I have unlimited free superintelligence running on my desk. GLM 5.2 launched a few days ago and it is taking the world by storm. Benchmarks and many people's own experiences are saying this is just about as good as Opus 4.8, definitely or right around 4.6 or 4.7. But something major just happened today. Unsloth released a version of GLM 5.2 that you can run locally on just 250 GB of memory. I downloaded and tested it on my Mac Studio and I'm going to be honest with you, I am completely blown away. It is just about as good as Opus 4.8. It is very very good. In this video I'll cover why this model is so good, show you some demonstrations, show you how to set it up so you can have unlimited AI as well. If this is your first time working with local models, I'll give you what local models are, how to set them up for your first time, what kind of computers you need. I'll tell you why I believe all of this is the future and everyone will be running their own local models very very soon. And I'll tell you how to start preparing for that future today. You are going to learn so much in this video. It will blow your brains out your wazoo. So let's lock in and get into it. So anyone who's watched my live streams before, by the way they're coming back soon, you know the 3D first-person shooter test. This what I'm about to show you is the 3D first-person shooter that GLM 5.2 ran completely locally on my Mac Studio. It is a good-looking game. You can see a great environment, good enemies. You can see a lot of video effects. The colors are nice. You can see hit counters and all of that. It is a very good test and is just as good for me as my Opus 4.8 test I did. It has waves, it has points, it has ammo, it has score, it has everything. It is really really good. This was all built by GLM 5.2 running locally, which is powering this Hermes agent I have right here. I told my Hermes agent to build the game. It built it out, said the game is fully working, and on top of that, it even tested itself, played the game itself, and then self-improved. So, it actually made its own skills for creating 3 JS games. This is a completely self-improving agent running on my Mac Studio. That is like mind-blowing to think the most powerful technology on planet Earth is just sitting on my desk right now. So, let's talk about GLM5.2 and running it locally and what makes it so special. It is open weights. That means you can run it on your computer. By the way, we're going to cover a ton in this video. Feel free to look down below at the different chapters and skip wherever you need to. I'm going to cover you some beginner stuff, some advanced stuff, like what local models are, how to run them. If that's not relevant to you, feel free to skip around. But, if you are brand new to local models, stick around for that in a second as well. But, it is a local model. It is open weight. So, that means you can right now download it, load it onto your computer. I'll also talk about what kind of computer you need to do this, and start using it completely for free locally. Based on my tests, it is comparable to Opus 4 8. There are weaknesses compared to it. I'll go into that as well very shortly. But, some of the things it's done, some of the tests I've given it, like that 3D first-person shooter, basically matched what Opus 4 8 was giving to me. It's running on my Mac Studio, one singular Mac Studio. I didn't have to link my different Mac I didn't have to make a cluster or anything like that. One singular Mac Studio. I'll talk about how much memory you need in a second. And what's amazing is it can power your Hermes agent or your Codex. So, right now, as I showed you, I have a Hermes agent running. Every prompt I give this Hermes agent stays local, is unlimited, doesn't limited work on my computer. It is powering this whole Hermes agent. I still have Hermes running on Opus 4 8 and another Hermes on GPT 5 5. I'll go over when you want to local models, when you want to use frontier models a little bit later as well. But now I have a third agent on my computer that's running completely locally. And as I said, Codex, which shout out to OpenAI, they allow you to use any model you want inside Codex. You can now do vibe coding in Codex with GLM 5.2, a model that is very, very good. Now for those newer to local models, they don't know much about it, let's talk about how they work and what type of hardware you need. You can run local models on any hardware you want. If you have a Mac mini with 16 GB of memory, there are local models out there that you can run on there. And I'll tell you how to do that a little bit later as well. But for this model specifically, GLM 5.2, it is a beefy model. It is a chunky boy. You need hardware for it. I am running the two-bit quant version of it. We'll talk about that a little bit as well. That version of the model is about 250 GB in size. That means you need 250 GB of memory, which means you can technically run this on a Mac Studio with 256 GB. You won't have much room left. It might crash. But if you were one of the people who listened to me early on back in January when I was spouting about how incredible Mac Studio 512 GB were, you can run this easily on a Mac Studio 512 GB. So you also have the DGX Station which Nvidia just started releasing across many different providers. That has 750 GB of unified memory, so you can run it pretty easily on there as well. That is just a very expensive computer. So you still do need a good computer to run this. But again, no matter what computer you have right now, there is a local model out there that you can run. And I will go over that very, very shortly. So let's talk about real quick the upsides and downsides here. Then we'll go into the more educational what are local models and how to set them up for the first time. The upsides of this model is it's free and unlimited if you're running it locally. It does cost if you run it through the cloud. I'll tell you about cloud versus local in a second as well. But if you run it locally, it's free, it's unlimited, it's private and secure. None of your messages go to the cloud. So if you want to have personal conversations with your AI, which I know some of you want to do, it is private, secure, no one else can read it. And it unlocks way more use cases. When you have unlimited private and secure AI, you can do a lot of things. Like for instance, I have my AI, my GLM 5.2 running on a loop right now. It is going through my code base of the new SaaS I'm building, Henry Intelligent Machines. It's making sure it's secure. It's fixing any bugs it finds. And it's doing this 24/7 365. These are the benefits of running local models is you unlock these incredible use cases. The downsides to local models are one, it's slow. I'll admit it. This is a very slow model. This is not going to be as fast as Chat GPT 5.5 or Opus 4.8 running on the cloud. It just won't be. That doesn't make it useless. It still has incredible uses. If it is passively working in the background doing things for you, you don't need snappy in the moment decisions to be made. It's doing work for me constantly around the clock in the background. So I don't need it to be lightning fast. I still use Opus and Chat GPT for the things I need done very fast. It does have a smaller context window. It just is what it is. And the more you shrink it, the smaller it gets, the dumber it gets. This is a two-bit quant version, which on most models would make it very, very dumb. But with this onslaught version, they actually found it has 82% accuracy, which is really nice. In a second, I'm going to go over how to set this up if you have the correct hardware. If you are newer to local models, I want to go through a few things first. I want to go through what local models are, and even if you don't have great hardware, how you can set them up. Again, if you're familiar with all this, feel free to skip down below to the different chapters. I'm throwing everything local models at you in this video. So, a lot of interesting information if you want to skip around. But, what are local models exactly? Just so we're on the same page. So, as I talk about how to set this up, it all makes sense. Local models are Local models are LLMs that run on your computer. When you talk to ChatGPT or when you talk to Claude right now, you write a prompt. Your prompt gets sent from your computer over the internet to the cloud or a data center like you see right here. This data center is filled with thousands, if not millions, of hyper-powerful GPUs. In a very, very, very simplistic explanation, basically what's happening on these GPUs is they get your prompt. It takes the prompt and turns it into numbers. It takes those numbers and runs a whole bunch of calculations, which gets you a response in numbers. The GPUs then take those numbers, turn it back into letters and words, and give it back to you on your computer. Basically, at the end of the day, all these GPUs are just doing a tremendous amount of math. The downside to all of this is you are paying for the GPU usage, right? You're paying for those tokens. And also, it's not very private at all. All your chat logs get sent to the cloud, get stored on servers, and anyone can read them at the companies for these frontier labs. Local models are different. Now, these LLMs are running on your computer. So, whether you're running on a Mac mini or you're running on a Mac Studio, now instead of prompts going to these servers, they're just staying on the computer, and these computers are doing the math of your prompts. That has many benefits. Now, your prompts are not leaving your computer. They're all being stored locally, so it's very, very private. And you're not paying a toll booth as your prompts go in and out of data centers. They're all local, so it's completely for free. It just costs the electricity going into your computer. The challenge with local models has been it's been hard for these AI companies that to make local models that are powerful on your hardware. The GPUs in these data centers are super, super powerful. But luckily, over the last year, these AI companies have done a great job of making the models more efficient, so they're still powerful on cheaper hardware, and figuring out ways to make the models smaller as well, so they're still smart even though the size is getting smaller. Those advancements have allowed things like today that have happened, which is GLM 5.2, the super Opus-level model, being just as good on your local device. Now again, downsides, it is pretty slow, so you're probably not going to be using this as your main daily drive. You're still going to use Frontier Cloud models to do things you need done quickly, right? Like if I was relying on this for live coding, I'd be sitting here forever. But because I'm using it passively to kind of review code in the background, it's not a big deal, and it's still super helpful. So let's talk about the computers you need to run local models, even if you're just running on a Mac mini, it's really dependent on the memory of your computer. When you load local models, they load into memory, right? So the more memory you have, the bigger the models you can run, the more intelligence you get. If you're on a Mac mini, if you're on a smaller Mac mini, you're probably going with Google's Gemma 4, which is a really, really small, but still pretty smart and efficient model, or NemoTron, which is a model from Nvidia, very happy to see Nvidia getting into local model game. If you're on better hardware, so you have like a good Nvidia chip like a 5090, or you have a DGX Spark, or a DGX Station, or a Mac Studio, you can run bigger models like GLM. If you're not on like the top-tier hardware like the 512 GB Mac Studio, I'd recommend for most people Qwen 3.6 27B. You're going to get excellent intelligence out of that. It's going to be pretty fast as well, and it can run on most kind of mid-tier hardware. So, let's get back to GLM 5.2 and how to set it up locally. I do basically all technical work through Hermes Agent. You can use OpenClaw for this as well. This is why I highly recommend everyone have a Hermes or an OpenClaw on their computer. And basically all I did was message my Hermes Agent, give it a link to the tweet from Unsloth, which I will put down below if you're running on good hardware, and say, "Can you get this exact model running on my second Mac Studio?" It went in, it built a plan, it researched it, and by the end of all of this, if we scroll down to the bottom, boom, brand new Hermes Agent set up with GLM 5.2 running. And so, now I can use the model, and I have a Hermes Agent powered by it as well. Basically all complex technical work is taken care of if you have a Hermes Agent or an OpenClaw running on your computer. Because it is pretty technical loading these local models up. You have to download them, you have to set up a server, you have to do a whole bunch of things. But if you just tell your Hermes Agent to do it, it just goes and does it and figures it out for you. One step, I hit enter, it was done and all set up. From there, now I have a Hermes Agent I can go to, ping anytime I want, and get it to do any work that I think would be appropriate for a local model to do. Which again, for those wondering at home, okay, what do I do with local models? What do I do with frontier models? Frontier, anything that requires the top-tier intelligence, right? Fable 5 is going to be better than all of this. Or if you need speed, right? If you're vibe coding, you're building something out accurately, you probably need speed. I'm using frontier for that. But local models, again, something where I want privacy, I'm having some sort of private conversation that I don't want Sam Altman reading in the servers, I will go and do that here, too, if it's something that can be done passively throughout the day. So, for instance, I have it checking every 2 hours my code base of my new SaaS looking for security issues, looking for bugs to fix, and it just fixes that passively 24 hours a day. It's just going and chugging through the code. It's doing it pretty slow, but because it's just a passive act, I don't care about the speed. If I were to do this with Opus or if I was to do this with ChatGPT, it cost me a lot of money. It would cost a tremendous amount of money to have Claude or ChatGPT running 24 hours in the background. So, it's perfect for local models. Now, if you were to use GLM 5.2 in the cloud, which you totally can do, so you use it like a regular model, the pricing is pretty good. It's much cheaper than ChatGPT and Claude. You're getting a lot of usage for a better price. It's pretty good. Now, there's questions that come up, okay, can I trust, you know, Chinese models? That's up to you to decide. I'm running it locally. When you run models locally, the data never leaves your computer, so you don't have to worry about going into other governments' hands to read. If you run locally, it's fine. If you run in the cloud, it's up to you. Although, I do know there are a lot of companies out there that are hosting GLM 5.2 on American servers if that's something you're concerned about. So, let's real quick talk about the future, why I think local models are the future, and how you can prepare for it. I think this is important for everyone to watch. By the way, if you learned anything so far, make sure to leave a like down subscribe, turn on notifications. I'm also going to do a full live boot camp on local models in the Vibe Coding Academy, the number one community for people in AI. Make sure to sign up for that down below. It's the best decision you'll ever make. Link for that down below. So, the future, everyone has their own super intelligence on their desk. This has all been converging in one way. Over the last couple years, local models have gotten smarter and faster and been able to run on cheaper and cheaper hardware. We are going to hit the point in the next year is my prediction where you can have amazing amazing intelligence running on the cheapest Mac mini out there. And at that point, I think that level of intelligence will be good enough for 90% of people. And so I believe in the near future everyone will have their own super intelligence sitting on their desk, none their data going to the cloud. It will be completely private and secure. It'll be your own personal intelligence. No Nobody working at OpenAI or Anthropic will be reading your chats and it will be doing work for you 24/7. So it'll be monitoring everything you do on your computer, helping you out where it can, building decks and documents and writing code all for you 24/7 passively in the background. I think this is a future that's coming within the next 12 months. So how do you prepare for that future? What do you need to do? Well, first you need to understand how local AI works. I just gave you a pretty good explanation, but make sure you understand how it works. If you watch my videos, you'll be in a good place. You'll understand how it works. Experiment with the hardware you have. So even if you have a crappy Mac mini right now, just install something that goes on it. What you can do is go to your Hermes or Open Claw agent say, "Hey, take a look at our computer. Figure out what local models we can run on it and what use cases would be good for that type of local model." Even if you're on a small Mac mini, you will still be able to run some version of Gemma 4 and do small tiny little tasks on it. So go to your Hermes or Open Claw now and do that and experiment with the hardware you have. The best way to learn about AI is just by taking action. Just by doing it. So install the model, even if it sucks, even if it can't take care all your vibe coding, still install it and use it and you will learn so much about AI and how it works. And then just keep up with what comes available. AI moves so freaking fast. New models dropping every single day. Make sure you keep up with AI, with local models, what's coming available for the hardware you're using, and stay on top of it to stay on the cutting edge. I really believe the only way to win right now in this new world is to stay up-to-date on the most trending latest technology and use it as quickly as you can. If you watch my channel, if you watch my videos the moment they come out, leave likes on them, you will be up-to-date on all the latest tech and using the latest tech and have a distinct advantage to your competition. So, make sure you subscribe down below as well. I'm going to be doing way more tests and showing you way more use cases with this GLM 5.2 running locally. I want to show you the coding loop I set up. So, if you want more information on coding loops, let me know down in the comment section below. I'll make that my next video if I get enough demand for it. I'm not sure if people are into like loops and coding loops. So, let me know down below about that. I hope this was helpful. I have the greatest job in the entire world. All I do is experiment and create videos on my experiments and teach you guys about it. It means the world you'd sit here and watch these videos and learn from me. So, thank you. Thank you. Thank you so much. I'm so appreciative you watch these videos. Hope that was helpful. I'll see you in the next video.
GLM 5.2. This is the model everybody's talking about. Now, it's not only about its impressive scores on benchmarks. It has a long context window of 1 million token, which is five times bigger than the previous generation. But, in this video, I really want to focus on the architectural details, which makes it one of the most impressive open model releases so far. It seems like this is the first open weight model that is actually close to the frontier. Okay, if you have been looking at Chinese labs, you are going to actually see some interesting patterns. And this is constant across DeepSeek, MiniMax, and now Z.ai as well. They're not winning by throwing more computer at the wall. They are winning on efficiency. GLM 5.2 is an open weight model under MIT license, which means you can just download it and run it yourself. But, even if you're using it through their API, it's pretty cheap. And the interesting question is how this model, which is so big, stays so cheap. So, let's start with how it is actually built. Now, this is a mixture of expert of 744 billion parameters. It's definitely much smaller compared to some of the other frontier models, but in terms of performance, people are actually reporting some really incredible results. Now, since it's an M- MOE, these parameters are split into 384 separate experts. And the beauty of this is that for any given token, a little router picks only a handful of those experts to actually do the work. So, even though the model holds 744 billion parameters, only about 40 billions of them fire for each token. Now, this is a trend that we are seeing across the industry. We're seeing more and more MOEs because they are a lot more efficient, and you need a lot less compute in order to run them. And this is kind of the trick that everybody is trying to adopt these days. Now, as I said, this context window is about five times larger than the previous generation, which also makes it harder to run. In classic attention, every token has to look at every other token to decide what matters. When the context is small, that's no problem at all, but the number of these connections grows with the square of the the context. So, it really explodes. At token, that will be a wall of connection exactly where the cost blows up. So, the whole game becomes on figuring out which connection you can safely skip. So, this is where the concept of sparse attention comes in, and GLM-5.2 build on Deep Seek's version of it. Now, here the idea is to add a small cheap component called indexer. Before doing the expensive attention, the indexer scans the context and picks out just a handful of tokens that actually matters. Then, the real attention only runs on those. So, most of the expensive connection just disappear. It's kind of like a librarian who, instead of making you read the whole library, hands you the three pages you actually need. Now, the catch is that the model has many layers. And naively, you would run that indexer all over again in every single one. And that's exactly the problem Index Share solves. So, instead of computing a fresh indexer in every layer, GLM-5.2 compute it once and reuses it across four layers in a row. So, three quarters of that indexing work is actually just vanishing. The result is that it's a 2.9 times fewer compute operation per token at the full million context window. And it's the same librarian now serving you four floors of the building instead of hiring a new for each one floor if you want to go back to that example. And the single trick here actually makes it possible to serve these models at 1 million context window. Okay, so that was the cost of reading a long context, but what about the cost of or the speed of writing the answer? So, GLM 5.2 uses multi-token prediction, which is kind of becoming a standard. Where it guesses several ahead and then verifies all of them in a single pass. When those guesses are good, it keeps them. So, it gets multiple tokens for the price of one step. They improved this enough to raise the acceptance rate by about 20%, which directly shows up in the speed up of the inference. So, between indexer and this, both reading and writing at the long context gets a lot cheaper at the same time. Now, one more practical touch here. GLM 5.2 gives you two thinking effort levels. So, there's a high mode which balances performance against how many tokens it burns thinking. And there's a max mode which basically opens it all the way up for the hardest problem. So, you get to pick the cost versus capability trade-off that task. So, you get to pick the cost versus capability trade-off per task instead of being stuck with one. And we are seeing this pattern that when it comes to reasoning models, models providers are enabling multiple different reasoning budgets or thinking tokens. So, you want to tweak these thinking efforts based on the complexity of the task. Now, the way they launched this was very interesting. Initially, there were no benchmarks at the launch, but later on with the open weight release, they actually also provided the benchmarks. And you probably have seen them by now, right? So, it's really impressive on agentic coding. On frontier suite, the long horizon test, it delivers 74.4%, which beats GPT-3.5 and basically is tying up with Opus 4.6. Kind of incredible. Now, the main story is not the benchmarks. I think there has been a lot of conversation regarding benchmarking. But the main thing is that when it comes to actual coding task, especially from front-end UI designs, it's a beast. It's really close to Opus 4.8. And if you haven't tried it, I'll highly recommend to test it out. Now, let's talk about pricing because I think this is where the Chinese nerds are actually winning. Although they're supposed to be compute constrained, but the pricing they offer compared to some of the US counterparts, it's simply incredible. For example, this is almost 10 times cheaper than Claude Max for similar amount of tokens. But even if you're not using the GLM coding plans, you have quite a few options when it comes to this model because there are US-based hosting solutions who are providing inference. Or if you have enough compute, you can simply host this model yourself. But you will need a few H100s for that. So, being open weight, this gives you a lot more flexibility and you are not tied up to a single provider. Okay, a couple of things before we wrap this up, right? So, the first thing is that if you're using their hosted API, you need to think about how comfortable you are in terms of sharing your data. There's a lot of concern about it, but as I said, there are options available. Because the open weight side is basically the escape hatch here since you can run it on your own hardware. Now, it's a beast when it comes to agentic coding. But there are a few other things to consider. First, it's text only, so it doesn't have vision capabilities. Second, you actually want to look for a harness that is well tuned to this specific model, rather than using it in something like cloud code, which is not probably specialized for GLM 5.2. So, you want to experiment with quite a few different harnesses if you want to try this model out. But, in general, we are actually seeing a very interesting pattern that I I kind of alluded to in the beginning of the video. The Chinese open-weight labs keep winning on efficiency, and the index share is exactly that kind of unglamorous trick that quietly moves the cost needle. So, I won't pay much attention to the benchmarks, but the main thing is these labs are actually competing with frontier labs now at a much reduced cost and being open-weight that gives them a lot more flexibility. So, what do you think? Is efficiency the real frontier now, or does raw scale still win in the end? Let me know. I hope you found this video useful. Thanks for watching, and as always, see you in the next one.
Loops are emerging as the single biggest unlock for people building software with artificial intelligence right now. But most people don't even know what loops are. And so today, I'm going to tell you what loops are. I'm going to show you why they're valuable. And then I'm actually going to give you many specific use cases that you can use loops for today. So what is a loop? A loop is a way to allow your AI coding agent to work autonomously towards a specified goal. The most important thing about loops is that it removes humans that allows the agent to work much more quickly towards this defined goal. And if it sounds very theoretical, I am going to break it down. So what is a loop more specifically? Well, you need two things. You need a trigger and you need a goal. With those two things, you can complete the loop. A trigger is what kicks off the loop. And there are three ways to kick off a loop. One, you can do so manually. You literally tell the agent, go do this loop. Two is schedule. You can schedule a loop to happen at a certain time of day or on a repeating schedule. And then three, you have actions. You can have the loop kick off based on some kind of action like opening a PR. Now to fully remove the human, we wouldn't want to kick everything off manually, but sometimes it is required. All right? And for the goal, the goal can be basically one of two things. It can be verifiable or we can use LLM as a judge. So if it's verifiable, it is something concrete, some specific number or some way to test it deterministically. If it is LLM as a judge, that means we're giving the model the ability to determine when it has reached the goal. Let me give you two examples. So for verifiable 100% test coverage in our codebase as an example, that is something that we know for sure and we have a nice way to test against when it is true. And for LLM as a judge, one example would be refactor until satisfied. And the satisfaction just means you as the LLM get to determine when we are satisfactorily refactored enough. All right, enough of the theoretical. Let me actually show you some examples. So, a lot of people talk about loops, but they don't actually give concrete use cases. And I wanted to fix this. That is why I am launching the loop library. It is a free library. I'm basically taking all the loops that I use and the ones that I see other people use and putting them in a single place so you can see them. You can be inspired by them to create your own loops or you can simply copy them straight from here. It's free. I'm going to drop the link down below. So, let's go over it. This is definitely my favorite loop and it's going to show you exactly how loops work. This is the sub50ms page load loop. Let me click into it. And here we are. So the objective of this loop is to get every single page load in my app under 50 milliseconds. And so that is the goal. It is a very concrete well-defined goal which really makes building a loop easier. So what I tell it is continue optimizing the code for speed. After each significant change, measure page load performance across every page under the same repeatable test conditions. continue until that's the loop continue until every page loads in under 50 milliseconds. So it is literally going to go through my entire application, every window, every page, every modal, load it. If it's above 50 milliseconds, it's going to continuously optimize it until it gets it under 50 milliseconds. Once it's done with one, it moves on to the next. That's the loop. That's the goal. But how do I actually do that? How do I actually kick it off? Well, the trigger in this case is me. I am the human and I'm going to manually kick off this loop. You can certainly set it on a schedule and you can even trigger it on, let's say, a PR open. So, every time you open a new PR, you also want to make sure that that new PR doesn't make the page load over 50 milliseconds. So, let's kick it off. So, we're going to click copy right here. All you have to do is paste it in. So I have the prompt right there. And then at the end or at the beginning, it doesn't matter. Type slashgoal. And this is a feature in codeex. Claude code also has a /goal feature. But as soon as you have this slashgoal, it's telling codeex to continue working until the condition is met. The condition of every page loads under 50 milliseconds. That's it. You just hit go. And it might run for 10 minutes. It might run for 10 hours. it will just continue to run until it meets the goal. And so you do have to keep a close eye on it if you're under a token budget constraint. So here it is in action. I sent this as a goal. Look for more optimizations to make sure every page loads in under 50 milliseconds on production. It worked for nearly 50 minutes. So I'm treating this as a production performance goal. I'll first measure the real team's page request path. And it basically, as you can see here, went through every single page and optimized it to load under 50 milliseconds. Loops are the frontier of AI workloads. And if you want to power them reliably and at production scale, use the sponsor of today's video, Digital Ocean. If you're running production inference, you're probably running into some of these problems. Your inference stack is too complex to operate. costs are unpredictable and I'm spending more time managing the infrastructure than actually building the things to be on the infrastructure. And most teams find out the hard way that the hard part of building AI applications is not using the model. It's actually everything around the model. The operational overhead, the fine-tuning inference complexity, the costs that become harder to predict as you scale. And that's why I want to tell you about Digital Ocean, the partner of this video. Digital Ocean is designed to minimize the total cost of ownership by giving teams a simpler path to production AI. They provide infrastructure that is optimized for inference and a vertically integrated core cloud that provides efficiency at scale. Vertically integrated is the key word. And with transparent usage based pricing that makes costs easy to predict. So, if you want to spend less time managing your infrastructure and actually building the thing you're excited about, Digital Ocean is the way to go. So, go check it out. They've been a fantastic partner. I've actually been using Digital Ocean for well over a decade at previous companies, so I can vouch for them. Go check them out. Link down below. Now, back to the video. Here's another loop that I really like. This is called the overnight docs sweep. Each night, review the codebase in full and make sure all documentation reflects the latest changes from the previous day. update the documentation as needed, then open a poll request with those changes. So, what I am doing is I'm making sure we have complete documentation based on any changes we may have made. This is an example of LLM as a judge. There's no verifiable way to know if we have complete documentation coverage. There may be some ways that we can say, okay, as long as a piece of documentation covers this section of the code, but ultimately what we're doing is saying, okay, LLM, you decide. So, how do we actually use this? Well, once again, just hit the copy button. We're going to come into codeex. We're going to click this automations tab. We're going to create via chat. We're going to delete this portion. I don't know why they put that in there, but I want to set up an automation. Then, we paste in what we just copied, and then each night review the codebase in full. hit go and let it run and hopefully it will set up an automation just like this. So there we go. I'll set this up as a recurring automation. So first I'm loading the automation tool rather than writing a one-off note. Perfect. So this is a way to keep your documentation always up to date. It is awesome. And by the way, I created this website with here.now. So shout out to here.now the partner on the loop library. I created it and I simply said deploy to here. Now and it was done. It's so easy. Next is the architecture satisfaction loop. This is one that Peter Steinberger himself says he uses often. Here we go. Refactor until you are happy with the architecture. Here is the trigger and the goal all in one sentence. Refactor, which is what the loop is going to do, until you are happy with the architecture. Happy with the architecture is the goal. This is another example of LLM as a judge. We can even give it more guidance on what happy with the architecture means. We can say be very strict about simplicity or make sure every single line of code is dry. Then after each significant step, live test the system, run auto review and commit. Track progress in and then we give it a markdown file to track the progress. This is fantastic. So it's tracking its loop as it's actually looping. Now you can kick this off manually or you can run it every night. So let's say during the day you're deploying a bunch of code and then every night you're just making sure that it's refactored, it's dry, and it looks really solid. So very good way to keep your codebase very clean. Next, another one of my favorites, the logging coverage loop. So let's click into it. Basically, what this loop is going to do is make sure that we have thorough logging throughout our app. And there's another loop that builds off of this that I'm going to show you in a minute, which these two loops together, you can start to see how loops can become so powerful. So, this says, "Review the systems logging and add missing coverage until every important path produces useful tested logs." And again, this just makes sure that we have logging for everything. And this is going to be manually kicked off. And this is going to be LLM as a judge because it says every important path and important is non-deterministic. It just means the LLM gets to decide what's important and what isn't. And by the way, if you want hands-on help with loops and other AI topics at your company, my team is offering free consulting sessions. I'm going to drop a link down below. We're only doing a few of these, so go apply if you're interested. Would love to talk to you. All right, so now imagine this. You have full logging coverage, but what do you actually do with those logs? Well, I have another loop for you. This is called the production error sweep. Every single night, we're going to review our production logs for errors. If you find an actionable issue, trace it to its root cause, fix it, verify the fix, and open a pull request. Then, ping me in Slack with the findings and PR link. If no actionable errors are present, ping me with that result instead. So we are kicking off a loop every night and the loop is looking for every error in the logs and we'll fix them one by one with the end goal being no more unressed errors in the logs. So that is a very concrete goal for this loop. All right, here's another loop. Something incredibly important to any website owner, any app owner is SEO. And not only SEO, now GEO. So, here's the SEO GEO visibility loop. Run an SEO GEO audit across crawlability, indexation, page intent, titles, internal links, structured data, source citations, and answer first content. Rank the gaps. I'm not going to read the whole thing. Fix the highest leverage issues. Rerun the same crawl. And here's the loop. Repeat until no critical technical issues remain. Again, you might have one issue. you might have 50 issues. The point is we've now kicked off a loop that fixes all of them until no more issues are present. So, this is a really cool one to run, let's say, once a week. All right, here's one of my favorite and one of the most handwavy loops that I have, but listen to this. This is called the full product evaluation loop. Create n realistic scenarios covering every major capability. Before testing, define clear success criteria and choose a consistent evaluation method such as past fail checks or a scoring rubric. Run every scenario under the same conditions and record evidence for each outcome. Fix the underlying cause of anything that that does not meet the criteria. Rerun the affected scenarios and then rerun the complete test. Continue until every scenario meets the original quality bar. Now, a lot of you might be thinking, "Wow, that just sounds like tests, right? It's just like a test suite. Well, kind of. But this is actually non-deterministic. This is allowing the model to go through every single use case in your application, in your product, figure out if it's good enough, determined by the LLM, and update it if necessary. This one really does work. It takes like 12 hours at times or more, but it really does come up with very good optimizations. Now, you can also customize this for your specific app. So, for example, I'm building something right now that requires me asking a question of an LLM and it providing a really accurate response with sources. So, I tell it, come up with 100 different use cases, wide ranging use cases for asking the LLM questions and judge whether the response is good enough. If it's not, iterate and improve it. So, I could keep going, but if you want to find all of the loops and any new ones that I discover, go check out the loop library. I'm going to drop a link down below. And once again, shout out to here. Now for hosting the loop library. Okay, so there are two major caveats with loops that I have to tell you about. Number one is it's not for every problem yet. Designing a loop isn't always easy. Specifically, coming up with the goal for the loop is not easy. If something can be verified like every page loads under 50 seconds, that is perfect for a loop. When we have to have the AI judge, LLM is a judge whether a goal is met or not. That's when it becomes a little more brittle because we are leaving taste and judgment up to the model. This becomes even more difficult when we're talking about building features. I have not really found a way to build features with loops. You cannot say loop until we build a full permissioning system. I mean, you technically can, but I'm not doing it because I don't know which direction the AI is going to go. I don't know what features it's going to build. I don't know when or how it's going to decide which features are worthwhile versus which are not. So, that makes it not great from day zero feature building. Now, one example of building a product from scratch using a loop is something I did where I told the model as a goal to clone Excel feature parody and it was running for days and days and days until I finally stopped it. It actually opened up Excel on my computer, used computer use, and literally clicked through and made sure that it had feature par. And yes, it was running for days before I finally stopped it. So, I do not recommend doing that. And that brings me to the second big caveat. Loops are very expensive. They are churning through tokens autonomously until they hit the goal. Some of these agents might run for 10 minutes. Some of them can run for days. So, for you token maxers out there, loops are fantastic. But for those of you who don't have an unlimited token budget, this might not work for you today. And by the way, if you like coding with loops, you might also like these four open- source projects that I reviewed that you can use right
So I've been getting the same question in the Cyberflow Discord for months now. "Should I learn C++?" And my answer has always been the same if you're serious about security, not learning C++ is like being a mechanic who's never looked under a hood. You can get by. But you don't really know what's happening. Tonight I want to show you why C++ is probably the most exciting language you can learn as someone in this field and how to actually make it feel like that instead of a university course that makes you want to cry. First let me be honest about something. C++ has a reputation. People treat it like this ancient terrifying thing that only grey bearded systems programmers touch. And yeah, it can be complex. But that complexity is also exactly why it's so powerful for security work. You are writing code that talks directly to memory, directly to hardware, with almost no abstraction layer between you and the machine. When you understand C++ you stop thinking about programs as things that just run and start thinking about them as sequences of memory operations. And that mental model is what makes you dangerous in this field. The way most people try to learn C++ is completely backwards. They open a textbook, spend three weeks on syntax, do some exercises about printing numbers to the screen, get bored, and quit. The reason it feels boring is because the projects are boring. The language itself is not boring. What you can build with it is genuinely insane. So here's how I'd actually do it. Start with learncpp.com and I mean actually start there it's free, it's comprehensive, it's written by people who understand the language deeply, and it doesn't waste your time. Get comfortable with the basics in about two weeks. Pointers, memory management, classes, the stack versus the heap. Don't rush past pointers because pointers are where everything gets interesting for security work. Understanding that a pointer is just a variable that holds a memory address, and that you can manipulate that address directly, is the moment C++ stops feeling like a language and starts feeling like a superpower. Once you have the basics the projects are where everything changes and this is the part nobody talks about enough. Your first real project should be a port scanner. Not because it's the most impressive thing you can build but because building one forces you to learn socket programming how your code actually opens network connections, sends data, reads responses and suddenly you understand what Nmap is doing under the hood at a level that reading documentation never gives you. You wrote it. You know exactly what's happening. That feeling is worth more than any certification. After that build a keylogger for your own machine. I know how that sounds but hear me out building one teaches you how Windows hooks work, how the operating system handles input events at a low level, and how software intercepts system calls. The same knowledge that goes into building one is the same knowledge that goes into detecting one. Offense and defense are the same subject viewed from different angles and C++ is where that becomes viscerally obvious. Then write a basic shellcode injector in a controlled lab environment. This is where everything you've learned about memory, pointers, and system calls comes together in one project and the understanding you come out with is something that Python programmers doing security work simply don't have access to. You are operating at the level where the actual interesting stuff in security research happens. This is also where I'll say something honest. Learning C++ for security is one of those things where the resources are everywhere but the structured path through them is not. You find a great tutorial on socket programming, then a blog post on memory exploitation, then a YouTube video on Windows internals, and you're jumping between things with no clear picture of how it all connects. The C++ course inside Cyberflow is built specifically around this not C++ as an abstract language but C++ as a tool for security work, with the projects that actually matter and the concepts explained in the context of how they're used in the field. Everything connected. Link is in the description, code Cyberflow50 for fifty percent off. Now back to it. For resources beyond learncpp.com The Cherno on YouTube is probably the best C++ content creator alive right now. His tone is engaging, he goes genuinely deep, and he doesn't talk to you like you're stupid. Watch his series on how C++ works, his videos on memory and pointers specifically, and his game engine series if you want to see what serious C++ architecture looks like in practice. Also read "The C++ Programming Language" by Bjarne Stroustrup eventually he invented C++ so he has some authority on the subject but don't start there, For the security specific side of C++ read shellcode. Actual shellcode. Go on exploit-db, find a C based exploit, read through it line by line and figure out what every part does. This is uncomfortable at first and then it becomes one of the most educational things you can do. You are reading the output of people who understand this language and this field at an extremely high level and that exposure changes how you think about both. The other thing worth saying is that C++ makes everything else you already know better. If you know Python, learning C++ will make your Python faster because you'll understand what's actually happening when Python runs your code. If you do web security, understanding memory corruption at the C++ level changes how you think about certain vulnerability classes entirely. It's not a replacement for anything. It's the thing underneath everything. The honest timeline if you put in consistent work two weeks on fundamentals with learncpp.com, one month building the projects I mentioned, and by month three you're reading real exploit code and actually understanding it. Not all of it. But enough to know what you're looking at and what to learn next. That progression from zero to reading real security research is genuinely one of the most satisfying things in this field. A lot of you have been asking about the course material and the roadmap I personally follow for this so I put a full C++ learning roadmap with all the resources linked in the first comment. Everything in order, nothing skipped. Grab it.
I've been using DeepSeek V4 as my primary AI tool for the past 3 weeks, and for about 80% of what I do, I haven't needed to open ChatGPT or Claude once. The reason isn't that V4 is smarter, it's that it has features I wasn't using that make the output significantly better once you know they're there. So, in this video, I'm going to walk you through every feature that matters so you can get the same results without paying for a subscription. DeepSeek V4 launched on April 24th, 2026, and it comes in two versions. V4 Pro is the stronger model built for complex reasoning, and it performs in the same range as Claude Opus 4.7 and GPT 5.5 on most benchmarks at a fraction of the cost. V4 Flash is the faster model built for everyday tasks. It's lower cost, quicker to respond, and still strong enough for the majority of what you'd use an AI tool for. Both support a 1 million token context window, which means you can feed them massive amounts of text in a single conversation. And the part that changes the equation is that the chat interface at chat.deepseek.com is completely free with no limits on the number of messages you can send. You sign up with an email, and you get access to both models, all four modes, file uploads, web search, and the full context window without paying anything. And because the model is fully open source under an MIT license, developers can also download the weights, self-host it, and fine-tune it for their own use cases. That alone makes it worth trying, but the features are what make it worth switching to. The biggest mistake people make with DeepSeek V4 is using it in one mode for everything. The chat interface gives you four distinct modes, and each one is built for a different type of task. Instant mode runs on V4 Flash, and it's designed for speed. Quick questions and simple summaries where you want a fast response without waiting for the model to think deeply. If you're asking, "What's the capital of France?" Instant mode is the right choice because the answer doesn't require complex reasoning, and you get it back in seconds. Expert mode runs on V4 Pro, and it's where you go for anything that requires deeper analysis. Complex coding problems and multi-step analysis where the quality of the reasoning matters more than the speed of the response. The difference between Instant and Expert on a coding problem is noticeable. Instant gives you to working answer. Expert gives you a working answer with better architecture, edge case handling, and cleaner logic. Deep Think mode is the one that makes Deep Seek V4 stand out from most free alternatives. It uses chain of thought reasoning, which means the model shows you its entire thinking process step-by-step before giving you the final answer. You can watch it break down a math problem, evaluate different approaches to a coding challenge, or reason through a business decision in real time. This is the mode you use when accuracy matters more than speed. If you're debugging a complex issue, solving problem where the first approach might not be the best one, or evaluating a decision with multiple tradeoffs, Deep Think lets you see exactly how the model arrived at its answer, which means you can catch flawed reasoning before it becomes a flawed output. Vision mode is the newest addition, and it's still in beta. It lets you upload screenshots, diagrams, photos of whiteboards, and even handwritten notes, and V4 processes both the visual and the text together. The rule of thumb is instant for quick tasks, expert for complex tasks, Deep Think for high-stakes tasks where you need to verify the reasoning, and Vision for anything visual. Switching between them takes one click, and using the wrong mode for the wrong task is the single biggest reason people think Deep Seek isn't as good as ChatGPT or Claude. When the issue is that they're running a complex analysis in Instant mode instead of Expert or Deep Think. What most users don't realize is that you can chain modes in the same conversation. I start a coding problem in Expert mode to get a working solution, then switch to Deep Think on the same conversation and ask it to audit what it just wrote. Deep Think breaks down the logic step-by-step, and about half the time it catches an edge case or a cleaner approach that Expert missed on the first pass. Two passes from two different reasoning depths on the same problem without leaving the conversation. That one habit of matching the mode to the task changed the quality of my output more than any prompting technique ever did. The modes handle how the model thinks, but the next feature handles where the model gets its information. Deep Seek V4 has a built-in web search toggle that lets the model pull current information from the internet while answering your question. You click the search icon before sending your message, and instead of answering from what it was trained on, V4 searches the web, finds relevant sources, and builds its response from what [music] it finds, with citations linked so you can verify every claim. This matters because any AI model's training data has a cutoff. Without web search, you're getting answers based on information that might be many months old. With it turned on, you're getting real-time data. For anything time-sensitive, like news, pricing, current events, market data, or product updates, web search should be on. For everything else, you can leave [music] it off and let the model work from its training data, which is fast and usually accurate for non-time-sensitive questions. File uploads are the feature that barely anyone touches. You can upload PDFs, code files, spreadsheets, [music] and documents directly into the conversation, and V4 will read and analyze them. Upload a 10-page contract and ask, "What are the three biggest risks in this agreement?" [music] Upload your code base and ask, "Where are the performance bottlenecks?" The model processes the file, references specific sections in its answer, and because the context window holds a million tokens, you can upload multiple large documents in the same conversation [music] and ask questions that span across all of them. File analysis combined with web search, combined with the right mode, is where DeepSeek V4 starts producing output that rivals paid tools, [music] and almost nobody combines them. Where this gets interesting is when you stack all three at once. Upload your document, turn on web search, and ask your question. V4 reads your file, pulls in current web data for context, and reasons through both together. I uploaded a competitor's product page as a PDF, turned on web search, and asked, "What are they doing better than us right now?" V4 analyzed the PDF, searched for their latest announcements and reviews, [music] and came back with a comparison I couldn't have built manually in under an hour. Web search and file uploads get better information into the model, but there's a reason the output still feels generic for most users, [music] and it comes down to how much context you're actually loading. The context window is the amount of information the model can hold in a single conversation, and DeepSeek V4's 1 million token window is one of the largest available right now. In practical terms, that's enough to hold an entire year of company documents, a full code base, or hundreds of pages of research in one conversation and have the model reason across all of it at once. The way to get the most out of this is to front-load your conversation with context before [music] asking your questions. Instead of asking, "How should I improve my marketing strategy?" with zero context, upload your existing strategy document, your last 3 months of campaign data, and your competitor's latest annual report, and then ask the same question. The answer you get from a model with 500 pages of your context is fundamentally different from the answer you get from a model working with nothing but your one-line question. The first answer is generic. The second answer references your specific data, your specific competitors, and your specific metrics. I tested this with a product strategy question. >> [music] >> Without context, V4 gave me a solid but generic framework that could apply to any company. With three uploaded documents, my product brief, our Q1 performance data, and a competitor's latest press release, >> [music] >> the same question produced a response that referenced our actual retention numbers, flagged a specific competitive threat from the press release, and suggested a positioning change based on where our metrics were weakest. Nothing about the question or the model changed. [music] The only thing I added was context, and the output went from generic to specific. There's a second layer to this that goes beyond just uploading documents. Before you ask your actual question, prime the conversation with a framing message. Something like, "I'm uploading three documents. The first is our product brief, the second is Q1 performance data, and the third is a competitor press release. Read all three and tell me what you noticed before I ask any questions." V4 reads everything, surfaces patterns you didn't ask about, and by the time you ask your real question, it already has a working model of your situation instead of treating each document in isolation. The average user never loads more than a few paragraphs into a conversation, which means they're using about 1% of the context window they have available, which is why the output always feels generic. Once you're using the right mode and loading context properly, the natural question becomes whether you still need to pay for ChatGPT or Claude. And the honest answer might surprise you. The honest assessment is that V4 Pro is not quite at the level of Claude Opus 4.7 or GPT 5.5 on the hardest reasoning benchmarks. But for the vast majority of everyday tasks, the difference is either invisible or too small to justify paying eight to nine times more. The reason the cost is that low is that Deep Seek had to build a more efficient architecture from the ground up because they didn't have the same hardware access as US labs. And that constraint produced a model that does more with less compute. On coding specifically, V4 Pro is slightly ahead with the best models available. And on competitive programming benchmarks, it holds the top score. On math, it's within a few percentage points, but noticeably behind on the hardest tasks. On everyday tasks like writing, summarizing, research, email drafting, data analysis, and document review, the output quality is comparable. The way I decide is simple. If the task has a clear right answer, like code that either runs or doesn't, V4 Pro handles it. If the task is open-ended and requires judgment, like evaluating whether a strategy makes sense or catching subtle tone issues in writing, that's where I still open on Claude. And V4 Flash is 90 to 100 times cheaper than Claude Opus 4.7 on the API, which means if you're a developer building an application, the cost savings are enormous without a meaningful drop in quality for most use cases. The combination of free unlimited chat access, four specialized modes, a million token context window, web search, and file analysis makes Deep Seek V4 the most capable free AI tool available right now. Where it doesn't compete is an ecosystem integration. ChatGPT has its plugin ecosystem and app store. Claude has projects and artifacts. Google has workspace integration. Deep Seek V4 is a standalone chat and API with strong capabilities, but without the surrounding ecosystem that the paid platforms have built. The capabilities make a strong case for switching, but there are three things worth knowing before you go all in. First, Deep Seek is a Chinese company, which means data privacy works differently than with US-based providers. Your conversations may be stored on servers subject to Chinese data regulations. For personal use and general work, this is likely fine. For anything involving sensitive business data, proprietary information, or regulated industries, evaluate whether your organization's compliance requirements allow it. Second, there are content restrictions on certain political and historical topics related to China. If you ask about specific sensitive subjects, the model may decline to answer or give a filtered response. For the vast majority of professional and creative use cases, you'll never encounter this, but it's worth knowing. And third, while V4 Pro is close to the frontier, it's not at the frontier on the hardest tasks. If you're doing advanced scientific research, solving the most complex competitive programming problems, or need the absolute best reasoning available at any cost, Claude Opus 4.7 or GPT 5.5 is still the stronger choice. For everything else, V4 is more than capable. Knowing these limitations is part of using DeepSeek V4 well, because it means you know exactly when to use it and when to reach for something else. But getting one tool down this well kind of gives the whole game away, because it was never really about the tool. Every result I showed you today came from one habit. Knowing how to actually use whatever's in front of you. >> [music] >> The people who have it don't just win with DeepSeek, they pull more out of every AI tool they touch, while everyone else sits there blaming the model. And that habit is the structure behind my school, AI Fluency. So if you want to get scarely good with all of it, and not just DeepSeek, go ahead and click the first link down in the description and join me in AI Fluency. See you on the inside.
First, Sonnet changed engineering. Then, Opus outclassed Sonnet. And now, Fable 5 and Mythos 5 are outperforming Opus. By now, you've seen the headlines, you understand that this model is an absolute beast. And you also know Anthropic is rugpooling Fable 5 from our subscription plans, ProMax, Team, and Sebast. This is completely unprecedented. Forget the June 23rd subscription. Rugpool everyone was mad about. On June 12th, the government pulled the entire model. Both Fable 5 and Miffals 5 are no longer available. A federal export control order made Anthropic suspend the best two models that anyone has ever seen after the government apparently found a jailbreak. I'm not sure the federal government has the skills, resources, and agentic engineers to confidently find and address these brand new jailbreaking techniques, but apparently they have. The very controversial piece here is that Enthropic says these same jailbreaking techniques work on models like GBD 5.5, and it's not specific to Fable 5 or Mythos 5. So, it's very weird that these same tricks work on GBT 5.5, but Fable is the model that got rugpulled. It's not clear what's true yet. What is absolutely clear is that this is temporary and Fable will be re-released soon. So instead of focusing on all the mania, all the hype, all the chaos of this situation, let's refocus ourselves on three actionable ideas you can use for all next generation mythos level models. In this video, I have three observations on Claude Fable 5 specifically for agentic engineers that can help you get the most out of this model. I'll list them on the screen here and at the end I want to share exactly how this model is changing my approach to agentic engineering. Now running a model in isolation is meaningless. I used Cloud Fable 5 to orchestrate itself along its two little brothers Opus and Sonnet. They all ran the exact same five specs each in their own agent sandbox. They looped until the work was complete and created a single public URL where we can see their results built end to end. These are full stack applications orchestrated by Fable 5. That's 15 sandboxes in total. This is enough for us to understand what this model really gives you because it's not price per token. Each one of these URLs is live built by the respective agent. Here we have Fable recreating Simon Willis's LLM price index. We have a hacker news clone. We have a scikitlearn model generator. We have a chat room where we're chatting directly with the pi coding agent docs. We had an agent build an agent inside of a full stack application. And then we have this brand new type of application and we can directly chat to any one of the agents inside this chat room. But you'll notice something interesting here. Sonnet, Opus, and Fable were able to do the job. So what is Fable for? In this video, we break down the model that will do what Sonnet did. Change engineering forever. Let's talk Claude Fable 5. This is one of the first models that is truly feeling like a payto-play model. But then there's the other side of this coin which is this very very important statement. Fable 5 is not an Opus replacement. It's a tier above Opus. It's a mythos class model with mythos class pricing. So that's the other side of this coin. Now the question is is this actually true? Right away we can address this. In our total of 15 agent sandbox applications, we can see the price differential between Sonnet, Opus, and Fable. Sonnet ran $55 worth of tokens. Opus ran $91 worth of tokens and Fable ran 200 per token. Fable loses this battle. This model is not giving you an improved price per token. And the raw numbers don't look good. Throughout all these applications, Fable shipped this with a million tokens. Opus did it with 700K and Sonnet did it with also around 700K. Where is the value does more with fear tokens is not true. This model is not doing more with few tokens. So what is it doing exactly? So this is the first observation I really want to nail home with this model. It's not about price per token. It's not about intelligence per token. What we're really looking for is this price per intelligent agent hour. Time is our most scarce resource. This is what we're getting out of the new fable model. There's a sweet spot here. This leads us to another really important observation that we'll talk about in a moment. This model is only really useful if your missions, if your tasks, if your specs are big enough, are complex enough. We're getting into this really interesting place where we're not spending on tokens anymore. We're spending on agent time. We're spending on useful agent time. The best, easiest measure is in hours. But here's the breakdown. This is where Fable completely wins, right? And it's not really close. You can see here the time taken to generate every single one of our applications and there is a gap here right a sizable gap. Fable is completing the work with more tokens more expensively about 20% faster. So this is the key observation when you're using Fable 5. What you're getting here you're not buying tokens you're buying intelligent agent hours. Get out of task top to bottom mindset and get into feature shipped work completed end toend work done. the spec is driving Fable to those next level results where we can truly delegate, get out the loop and just judge the results and iterate from there. This model kind of pushes us further away from prompting back and forth, babysitting the agent, and repeating that cycle. It's more and more becoming about specs, proper delegation, proper looping, proper closed loop structures, making sure your agent can validate all of its work, and then understanding the review process. So, that's the headliner here, right? The harder the mission, the harder the task, the harder the feature you're trying to ship, the more Fable makes sense. Looking at our applications that were generated, the price just keeps going up. This model is just going to use more and more compute. You can see the cost roughly equivalent. It's just using double opus, double sonnet. And this is the big takeaway, right? On about 80% of these tasks, the sibling models, the brothers of Fable, Opus and Sonnet, did the job at a fraction of the price. put on our harder task here, specifically our multi- aent chat room application here. This is a significantly harder task. We have an entire micro IDE in here and our agents are operating on this inside of the UI full stack application. You can see all the personas in the bottom left. We're of course spinning this up using the pi coding agent under the hood on the server that fable oneshotted through this entire application. We can prompt this directly. I'll run on all give me your perspective on the application. And so all is going to kick off every single agent. You can see they're now starting to think here in the top left. We're chatting to everyone in this chat room and everyone is going to give me the work received. So they're showing their red files and they're reporting. So we're in this really really interesting application where we can just chat back and forth with all of our agents in a chat room. And so we just address all of them with at all. We can of course address our agents specifically. But you can see we got a nice perspective from every one of the agents in the chat room. And again, this is one of the harder tasks. We also have our PI documentation support agent. We have our market direction scikitlearn model predictor. We have hacker news clone much simpler and an LLM price index just presenting information again much simpler. You know what we're looking at here is all the fable version. So of course these are going to be the top-notch best versions but we also paid a massive amount for these results. And the big kicker here is that we didn't need to. This is one of the big ideas the kind of uncomfortable ideas we're going to get to in a moment. But one of the key observations, the larger the task, the harder the task, the more fable makes sense. And here's my kind of extreme version of this. If you can cure cancer with a million Fable tokens, $10 per million tokens is nothing, right? It's absolutely nothing because doing this work, curing cancer or whatever hard valuable work you're doing is is worth a lot more than $10 per million tokens. $10 per million tokens is nothing. At the same time, if you're centering a div, you're making a donation to Enthropic. Okay, let's just say it like it really is. The price premium here scales directly with the mission. It will create that chat app for you. It will do that, you know, minor front-end backend work for you. It'll create the migration file for you, but you'll be mostly wasting money. So, first takeaway, what we get out of Fable is price per intelligent hour. That's the new thing to compare up from price per token. If you just look at price per token, 2x opus, nobody likes that. That sucks. They're rugpooling Fable from the subscription plan. What we get from Fable is price per intelligent agent hour. Time is the resource that if you prompt contacts harness engineer properly with this model, this is what you get back. This is worth all the money in the world, right? It's time. All right. And so we can see this directly in our receipts and this small microcasm benchmark. The results show up pretty quickly. Just 15 items to compare here. 15 full stack applications that Fable orchestrated. Okay. And speaking of orchestration, that brings us to our second big observation for your agentic engineering. Cloud Fable 5 is not an intern. It is an orchestrator. I've been using this model since release and this is one of the standout pieces in terms of raw performance. Cloud Fable 5 is not an intern. It's not a worker. It's pushing toward the most valuable thing an agent can do and that is orchestrate. Every engineer's progression looks like this. Now it is you start with a base agent. You then make it better. You learn how to prompt and context engineer. You then add more agents. You then customize them. You specialize them to outperform any agent without that unique information. And then at the last level, you orchestrate every previous level. Last week we talked about Cloudflare's review software factory. They are using multi-agent orchestration. They understand the value of it. And of course, Enthropic understands the value of multi-agent orchestration. If we go to the system card and we just search for multi- aent. Guess what pops up? An entire section on multi- aent. What did they find? They found what everyone finds when they start scaling their work with agents. What do they do here? They have a bunch of benchmarks where they're specifically battle testing multi- aent orchestration with the best model and then with the best model scaled up to three, five, 10 async and then I think they have some versions where they're just running unlimited agents. Spin up as many as you want, right? Async sub agents. And as you can see here, accuracy on the left, latency per task on the right. What you get is exactly what you would expect. If you scale your compute, you scale your impact. aka if you use more agents and you have a great model that can steer them. If you have an orchestrator model like Cloud Mythos, like Fable 5, you get better results and you get them faster. Now, it's not always faster. Token usage across agents also adds a time cost. It's not always going to get the job done faster. Sometimes it's slower, it costs more, but you still get the accuracy, right? You still get the raw performance when I'm prioritizing what I'm looking for out of my models, out of my agents. It's always this, right? There's a trifecta for every single agent. You're always trading these three things off. This is the trade-off triangle. It is performance, speed, and cost. As a northstar, I'm always sacrificing speed and cost. But then very quickly, depending on if you're running a product agent or if you have a subscription where you can blast through tons and tons of tokens. That equation is going to change very quickly. There are many things you cannot deploy a fable level model or even an opus level model because the economics do not make sense. But back to multi-agent orchestration, the idea is simple and enthropic knows it. Again, everyone using models at scale that has pushed past the progression of agents base, better, more custom orchestrator. You know that if you want to push your results, you add agents and you add specialized agents and then you let the orchestrator do whatever it needs to do to get the job done. And so that's what we're seeing here. Enthropic seeing it in the benchmarks. This is not new for them. They found this pattern with Opus and they've been building their models to be better orchestrators. Another way to say that is that they're building their models to be better prompt engineers. So this is the second big takeaway for Fable 5. Okay, this model is not a worker, it's a leader. If you have this model center or make a stupid small change, you're wasting it. You're legitimately wasting it. This model is not a worker, it is a leader. And so, you know, in my head, this model is performing more and more like a principal engineer. For sure, if you're just vibe coding, if you're firing off random ad hoc prompts, you'll never see this capability out of this model. But if you are agentic engineering, if you're writing great prompts, if you're setting up your context, this thing can perform like a principal level engineer. What do principles do? They can do the job. Of course, they know how to do the job, but what they do best is they delegate to others. Okay, so Fable 5 is the ultimate orchestration model. If you're thinking about how to get maximum results out of this model, it's in delegation. It's in orchestration. It is in building multi- aent orchestration systems, multi- aent orchestration agent harnesses, so on and so forth. So this is the second really, really important observation. This is the pattern I used to create this right to benchmark this model. I had Fable 5 spin up multiple sandboxes. Five on itself. Five have Fable, five Opus, five Sonnet. They each got their own sandbox. They each ran the exact same five specs. And the interesting part here is in the results. Just to mention this, I think that every model release there is some fear that um everything can just be done with a prompt. I firmly do not believe this. I think engineers will continue to have a place in the world. In fact, we're going to need even more engineers that actually know what's going on. The way I think about this is the floor in the ceiling. So, models like this raise the floor, but they catapult the ceiling up as well. If you have been, you know, following channels like mine, if you've been doing the work using agents and trying to build systems that build systems, going to that metaentic engineering level, the ceiling is much higher for you. You can do a lot more with a model like Fable. That's a big observation here. Uh, Fable 5 is an orchestrator. Treat it like an orchestrator. Treat it like not even just a co-worker anymore. Like this is I'm really thinking about Fable as a principal engineer that knows how to given a great spec. I should preface with that. garbage in, garbage out for every system, right? That's just a foundational truth. But if you are writing great long specs with high detail with validation testing and review steps that make the loop very very clear, this model can perform like a principal engineer. And the best principal engineers delegate to scale far beyond themselves. Okay, it's the same thing. Fable 5 only has a million context window. And you know, a great example, just scrolling back up, I had a single Fable 5 instance kick off all 15 agents. And you know, I don't need to like prove this in any capacity, but let's just pull down the session I was using here. And this is the entire session to spin up the agents, run it, execute it, and then I use it to also build the presentation that we're looking at right now. Check out my context window. 62% 600,000 tokens. If I was not delegating, this work would be impossible. Look at the token usage inputs. Well, more than a million, okay, total combined outputs 100,000 looks like 2 million total. Look at the costs, right? The estimated costs from running all these agents and they're on sandboxes. Just to re-emphasize the point, this is impossible without multi- aent orchestration. And many things are impossible without multi-agent orchestration. That's why we've been talking about it on the channel for months, probably over a year now. I have no idea. I've lost track of time. We've done so many of these videos all the way back to some of the original ideas before Cloud Code was released and then of course Cloud Code's original release of sub agents, right? Going way back there. You know, drop a like if you were with the channel for that long and drop a comment as well. Shout out to you. But um here we are proving it in the future. Multi-agent orchestration is how you get outsized results. This was true back when you were prompting just a few cloud 3.5 sonnet models and it was true with cloud 4.5 and it's even more true here with Cloud Fable 5. So multi-agent orchestration super powerful. That's the second big observation. If you want to push this model to the limits, treat it like a principal engineer. Give it a large plan. Give it serious work to work through like spinning up 15 full stack applications in their own sandbox and then let it rip. Let it absolutely rip. And we can open up the other examples here, right? We also have Opus. There's our LLM price index. This is a real full stack application. I I just want to like express that it's hard to work through all of this content of what these models are capable of now. but Fable and then we can pull up Opus, right? Opus 4.8 and let's get Sonnet. Let's get this model comparison. And so this application is letting us quickly compare. This is directly inspired by Simon Willis's uh llmmpprices.com. He's had this site up for a while. You know, I told the agent, look at the site, clone it. This is one of the simplest things these models can do, right? Clone it and improve it. You can see here we have probably over a 100 models to compare. This is the Opus version. We can go ahead and pull up Opus Hacker News clone. Looks great. We can pull up scikitlearn via the obus version. Will spy close up or down tomorrow? This is our prediction scikitlearn models for any engineer that hasn't touched the data science world. You know, we've got 5,000 rows that we trained on. We split it up into train and test split. Then we have our pi chat application, right? Let's go ahead and ask a question about the pi coding agent. How do I customize the pi coding agent? Fire that off. Again, full stack application. You need to give these models even opus hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard hard tasks to understand their limits so one of my benchmarks now when a new model is released I give the orchestrator in this case Fable 5 access to a bunch of compute and by compute I mean CPU so using exev to create agent sandboxes and uh I'm giving each agent access to control the entire instance they need to then plan build and then host the application that they're building via a public URL so we can access it and look at it so this is all the work our agents are We have one more Opus version, the agent chat room. And you can see here that it's a similar type of application. It's fully built. Onto our third observation here, and this is kind of a scarier one for some engineers that might make you feel like you're getting left behind. The third observation is this is probably the first model engineers don't need. Why is that? It's because just like Opus 4.8 is saturating benchmarks left and right. That means you know with every benchmark with every tool call with every long agentic coding benchmark that's getting blasted through. Here we have Toolathon. This is a really popular one. All these benchmarks are proxies for doing real engineering work. You can see that these numbers just keep going up. First run pass, three run pass average turns to solve the problem. You can see that these numbers keep going down and performance keeps going up. Even the first shot, one shot performance keeps going up. So with each one of those benchmarks are getting saturated but also engineering full stack work is getting saturated. And this is why I use this exact benchmark that I like to use. Spinning up agents in their own virtual environments having them host full stack applications very quickly at scale thanks to our orchestrator agent here. But this is why I like to do this because it shows something really really important. Let's open up sonnet. Sonnet lm prices. Okay look a little bug there. Not too bad. Let's open up our hackernews clone from sonnet. Okay not too bad. different styling, but same thing. We can go into the comments right here. Some of the comments from the Cloud 5 release. This is a real clone using a snapshot of data, but it looks like it did the job great. Let's go ahead and look at Sonnet Scikitlearn. So, can we spin up new scikitlearn models? Okay, not too bad. Again, you know, looks like it's getting us 80 90% of what Opus and Fable have. So, let me just kind of nail this point home. Right, here's our PI chat app. How do I add uh new tools to my extension? Fire that off. Then we have our Sonnet chat room. For sure, one of the hardest to build. Let's see if we've also done it here. We'll do that same prompt. All give me your quick summary on this app. Okay. So, let's see if Sonnet has also done the job. You can see our agents are thinking over here. And let's see if we get that pi coding agent response. There it is. We have two responses. It looks okay. You know, you can probably assume there's something off here. But as we look at the application and we look through these files, it looks okay. And this brings me to the point. All these agents are doing the job and the difference is in how close are they getting to the result and how quickly. It's not just about cost, it's about time. So our third observation here is this is the first model that you might not need at all. This is like a weird realization that I came to while testing this model. In a lot of these cases, you can do probably 85 maybe 90% of what Fable can do. Very very weird, right? It's kind of a weird thing to think that I don't need the state-of-the-art model for this problem for many problems. And even further, there's like a checklist that you can go through. You can now create, you know, a model stack, state-of-the-art workhorse and lightweight models. We talked about this again in our last week's video around tokconomics. Cloudflare had S tier tokconomics because they built something like this out for engineering work, for product work. You don't need state-of-the-art all the time. And in fact, I think we've finally hit the point where many of us engineers don't need Fable. And Opus is enough. You don't have to pay to play. you just don't have to. Before you upgrade, you have to ask yourself a few questions. If you can't write long detailed plans, if you don't know what questions to ask, if you can't envision the solution end to end, maybe you don't need Fable. Maybe Opus is enough. Kind of a weird realization, but I think this is now true. This is the first model that you do not need to upgrade to. We can see that in a lot of the benchmarks. Opus is a great model, probably one of the best there is. You can see it in the benchmarks. It's not like Opus is very far behind Mythos or Fable. It's actually quite close. In a lot of scenarios, you won't be able to tell the difference. And that's kind of the key idea I want to communicate to you here. This is the first model that probably 80% of engineers don't need. I can say that pretty confidently. If you're not writing long plans, greater than 100 lines, 200 lines, 500 lines, multiple pages, right? HTML specs, specs with images. If you're not doing that, you probably don't need this at all because your problems aren't complex enough. And it's not like a slight. I'm not saying that to like be mean or anything calling your problems small, but like it's just true. Like don't waste money on compute you don't need. Going back to the first point here, it's really about price per intelligent agent hour. This is what you want to hit. And if it's not worth it, just use Opus. Opus is a great model. It's going to do a lot of great work. I'm going to continue to use Opus, especially come June 22nd. But for my best, most complex, hardest work, I'm for sure absolutely not going to be shy when it comes to using Fable 5. So three observations there. How is this concretely changing the way I am doing my agentic engineering? There are two constraints of agentic engineering. We talk about them on the channel quite a bit. There's planning and reviewing. A sign of a new state-of-the-art model is you can sit down, ask for more, and expect it to be built exactly as asked. And once again, let me just echo this again. 600k tokens. I built 15 full stack applications for sure with some issues, especially on the sonnet level. But I did this because I wrote a great detailed plan and then the agent took it from there. It wrote all the specs in my specific spec format that I templated. It's a rich HTML plus images plus structured document. It did all that. Then it executed and orchestrated agents as requested. And if you're in the cloud code world, you're likely familiar with, you know, you can fire off the workflows, you can fire off the loop to make sure that it's hitting all the validation commands. It doesn't really matter. It doesn't matter how you do it. It just matters that you know what tools are available to accomplish the work. So when I'm thinking about how I'm changing my work around this model, these state-of-the-art models always allow me to do this. More planning, less reviewing. And it's not less reviewing because I'm getting lazy. Never get lazy with your engineering work, especially when it hits production. It's more planning, less reviewing because it's listening more closely to everything I put in my plan. And this is one of the big differences between vibe coding and agentic engineering. Are you saying exactly what you want the model to do? and can it do it? And with these models, Fable 5 is a great example. With the leap forward, you can do this more and more. You can sit down, type exactly what you want to see, and then expect it to be built specifically because you're writing great specs. Again, this comes right back to our third observation. If you're finding that Opus can do the work literally like in this case, right, we have several examples where I can't tell the difference between Opus and Fable. And that is because specific example right our LLM price index our hacker news front page and will spy go up or down tomorrow scikitlearn classifier models I cannot tell the difference between these three between fable and opus. Why is that? It's because it doesn't matter. You might like the UI a little more. You might like a small tweak here or there but that's not what really matters. What matters is did the agent do the job you asked it to do. And so where things do start to deviate is maybe our PI documentation chat app where we had the agent build an agent inside a full stack app. So this you know definitely getting more complex here. You can see that Fable did complete this properly right we're getting some great docs right side left side classic chat on right type of application and it also built out an entire file tree inside the application for the agents to work in. And now I can chat with any specific agent at tester run our tests. So now all the agents are thinking is this relevant to me? Nope. Everyone turned off except our tester here. And now tester is going to run all these tests. Okay, it's executed our test. There it is. There's the run command. There's the results. This is getting very meta when you start doing agent on agent on agent on environment on environment. Type agentic engineering. But that is for sure where everything is going. Anyway, to continue to communicate the idea here, here's opus LLM price index looks great. Little spy close up or down tomorrow. Looks decent. And then it looks like it did finally get that uh last result here. Right. If we say show more, how do I customize the pi coding agent? It gave me a nice breakdown here. The UI here a little off. Would be nice to have this to be extendable. But we didn't detail this. This is this was my fault. I didn't detail the exact UI in the prompt, but it looks like Opus did accomplish this work. And then our multi- aent chat app. Let's go ahead and see did we get a summary here. Uh let's see at all. Summarize your understanding of this codebase. So firing this off in Opus. Once again, you can see all our agents thinking here on the sidebar. This is a aentic UI application I've been envisioning and haven't really put into work. Starting to get our responses from every one of our agents. I built out a version of this in the PI coding agent in the past, but as a standalone UI. Uh this is definitely an idea I'm thinking about a lot more for multi-agent orchestration, being able to spin up multiple chat rooms where you have specific agents working on specific parts of your work. We'll probably cover on the channel at some point or someone will probably steal the idea. The point here I'm trying to make like let me just dial all this in. In the end, every model release is about understanding how far you can push your model. For a lot of engineering work, a lot of engineers, myself included, at certain times, we're going to be wasting tokens and wasting money on Fable. And you can push against this by pushing the model to its absolute limit. Sit down, ask for as much work as you possibly can, and then let it crank. Understand where it starts to fall apart. And then that's where you start making up for its downsides inside of your custom agent harness by scaling more compute, right? Add more verifiers, add more reviewers, so on and so forth, right? That's where your real agentic engineering prowess and skill comes in. Maybe controversially, I fully believe every model can be improved by adding another model that substitutes what that first model was missing. Often times, I'm talking about the exact same model. Just add another model. Scale your computer, scale your impact. I think anthropic based on all their work would probably agree with that. Um, I'm certainly betting huge on that. When I think about my road map with this new model, it's really about this. I'm going to be planning more work and I'm going to be doing less reviewing because the model's adhering to the plan more. And so, outside of the orchestration realm, which is where Fable definitely shines and should be used, this model will absolutely be used and should be used for doing the actual work. You know, principal engineers do ship the best work as well, right? They delegate and they do the best work. And there's a whole slew of review focused agents, prompts, skills, so on and so forth that can now be wired up and now be deployed thanks to Fable 5. So that's one big change I'm making. The whole point here is to hit the goal of 2026, right? My goal for this channel, if you're an engineer watching, and for myself, is to gain full trust in my agentic system so that I can ship into production from a single prompt. And Fable is absolutely going to be a key part of that because it can execute longer, more complex plans over longer periods of time. So anyone can ship a readme change into production with one prompt. Uh not a lot of engineers are probably should be experimenting with that more. But where the real skill of agentic engineer comes in is in fully trusting your agentic system so well that you can run a single prompt. You can have your agentic system run, plan, build, test, review, document, and then it's in production. You ship from a single prompt. This is the northstar of agentic engineering. This is ZTE, zero touch engineering, which I realize is a misnomer. It's technically one-touch engineering, but zero touch when you compare it to everything we used to do as engineers makes a lot more sense. The way I'm thinking about this, just as like a putting all of our observations together, the more ambitious the instruction, the better the result. And this is where really where you're going to get your return on investment for using Fable 5. With that being said, this is the first model where for most work, you probably don't need it. If you can, and if you're solving hard problems, I absolutely recommend Fable. I'm going to be using it. It's very clear this model is a beast, but of course, we're going to have to pay for it. Kind of a weird headline, kind of a weird perspective, but I think, you know, the truth is always more nuanced than we'd all like to admit. I think that Fable 5 is an absolute beast, is an incredible leap forward over Opus, the previous state-of-the-art, and most engineers don't need it. This is an opportunity for you to prove if you really need this model. Push it hard. obviously largest and useful specs that you possibly can, right? Really think through everything you want done and how you would test and verify that it works. Um, but I do think that this model has outgrown what most engineers can even ask for. I'm super curious. Comment down below if you made it to the end. First off, you know, thank you. Big shout out to you. But also comment down below what's the biggest task, the biggest feature you've handed off to Fable so far. I'm super curious what you've one-shotted in Fable or few shoted in Fable. This is how I'm thinking about how to best push and use this model. Ask for more. Think fully about what you want to see and then hand it off to the agent and let it fire off a powerful multi- aent orchestration system. This is how we push the frontier beyond. And we have clear and clean signal from Enthropic that that is exactly how to push these models. If you want to scale your compute to scale your impact, multi- aent orchestration is the answer. The right way to be thinking about this model and probably more models in the future, specifically the state-of-the-art models, is this is a new tier with a new price and it gives you new results. The thing you want to look for is not price per token. A lot of engineers are going to convince themselves out of the most important technology of our lifetime by looking at it this way. You and I are looking for price per intelligent agent hour. That is the true signal. These are three observations I've seen coming out of the Cloud Fable 5 release. Comment down below. Let me know if you agree or disagree with any of them. If you made it this far into the video, huge shout out to you. Thank you. You know where to find me every single Monday. Stay focused and keep building.
Who was more important to the national team, Landon or Tim? So what’s the verdict? [chuckles] [Landon Donovan speaking] All right, Tim, soccer is about to take over this summer. So let’s talk about Google’s top searched soccer debates. So, the clock starts now. First trending debate ... Why is it so hard to score in soccer? It’s like chess. There’s all these variables. [Howard speaking] Me and you, by ourselves. The goal is huge. You then put all these pieces in front of the goal scorer, who are strategically trying to take away space. It becomes much more difficult. [Donovan speaking] Speaking of scoring, this is like a massively asked question. What do you guys think? Would you rather win 5–4 or 1–nil? That’s an easy one, 1–nothing. If I'm a striker, I'm like, yeah, 5–4. Let’s go. Looking at pro soccer finals from the last 10 years, do more games end in 5 to 4 or 1 to 0? So Google says, “In the last 10 years ... not a single major professional final has ended 5–4.” — Okay. — [Donovan] But I love 5–4 as a striker, I love it. It’s the best. I have seen those games and they’re exciting, but ... people lose their jobs when games are 5–4. [laughs] That’s right. The next one is ... in a penalty shoot-out, who has the most psychological pressure, the striker or goalie? Google said, “The striker faces significantly more psychological pressure. Goalkeepers have zero expectations.” Exactly. — [Howard] It’s easy on the goalkeeper. — [Donovan] Yeah, because you’re not expected to save it. [Howard speaking] If you make a save, you’re heroic, but all the pressure’s on the striker. [Donovan speaking] Why do soccer players fake injuries? Ahhhh, my hamstring! Say why. So Google says, “One reason is for winning penalties.” Sometimes if you don’t go down, they won’t call the foul. It's high risk, high reward. Alright, last one. “Who was more important to the national team, Landon or Tim?” [Howard speaking] Tim Howard. 100%. Just for the record ... the first thing it says is “Landon Donovan.” Hmm, you have to triple check some of that stuff. [chuckles] [Donovan speaking] “Landon Donovan was the team’s offensive engine. Tim Howard was more important in specific high pressure tournaments.” [Howard speaking] Well ... I had to carry the team a lot. So, on your off days, I had to make up for — Landon has no off days. [laughs]
1 year ago, I showed you how I saved $7,000 a month on web hosting. Today, I'll make you forget about paying for any automation tool. I will show you how I run unlimited AI bot automations without using an 8 and make Zapier and without any limits. If you are ready, let's get started. This is Pylon. I built it and it's completely free and open source. And it runs every automation you're about to see. Let me show you how it works with three practical examples. Number one, brand tracking. Tools like Brand24 charge up to $199 per month with limits. These tools scan the web for every time your brand or your product gets mentioned online. It's very important, but really expensive. Watch this. Done and look. It emails me the report automatically. Now, here's the part I love. I set the schedule once, daily, weekly, whatever you want. It runs forever without limits, without subscriptions. Look at this one now, the keyword rank tracking. Anyone working online with a website, with a product wants to track their site's rankings on Google. And if you go for an SEO tracking tool, you will commit to like $30 or more a month. With Pylon, the same. Click, done. And I get the report into my inbox. And you see what's happening here? We are replacing subscription-based tools with simple automation using Pylon. And we are just getting started. Now, here's how Pylon actually run under the hood. It runs simple Python scripts. And I know what you're thinking now, I don't know how to code, so how I can use this? No, don't worry. You don't have to code anything. I'll show you this in a second. And what makes Pyloners super powerful is actually Python itself. Python has over 650,000 libraries. You can't imagine the flexibility you get. You can build literally anything you want. Let me show you what I mean with a real example I built. I want to automate with AI my YouTube channel comments management. So, I'm going to build it in front of you right now to show you how easy it is. What I did, I built a closed skill. You just install in closed, it's free. I describe what I want in plain English and in few seconds closed writes the entire Python script for me. Copy it, paste into Pyloner, schedule if you want, hit run, and here we are. It is that simple. The same automation would take you an entire afternoon connecting nodes in N8N. And honestly, I hate this. With Pyloner now, I set automations in minutes. But, honestly, now I want to share the most important point in this video. The skill you should be investing right now in 2026 and beyond is not learning N8N or Zapier. Those tools come and go. The skill that everyone should be learning is learning how to communicate the right way with AI. Now, in this video I made things easy for you with the closed skill. But, you should be learning how to create skills so you can build anything you want. That's the actual unlock. Start today and thank me later. Now, you're probably thinking, "Okay, how I actually run this?" Let me show you, it's super simple. Step one, get a VPS. I use Contabo and you can grab one for like $5 a month. Step two, install Coolify with one simple command. This is the same setup I used in my $700 saving video. Link below if you missed it. Step three, deploy Pyloner on Coolify. Enter the Docker URL, set the environmental variables, hit deploy, and that's it. You are running. And if you already self-host like me, this costs you nothing extra. It's completely free. If you are starting from scratch, I put together a full step-by-step guide in the description below. It's totally free. You can check it out and get started. And if you want to go in deep every service I use, every trick, you can check my full course on self- hosting linked below, too. Look, I'm not just saving $200 a month with Coolify Runner replacing different SaaS and tools. I built a system that no SaaS could sell me, and you can, too. If you learned something new in this video, smash the like button, and see you in the upcoming videos.
It was a wild week in the world of AI and I don't want to waste your time. So, let's just get right into it. Starting with the story that pretty much rocked the AI world this week. For the first time ever, the US government forced a publicly released commercial AI model off of the market. Enthropic had to shut down Mythos 5 and Fable 5 for every user on Earth. And in last week's news video, I spent a good chunk of the video talking about how much I loved this model. But unfortunately, we only got it for less than a week. On Friday evening, June 12th, it was gone. So, here's what happened. The US government basically said, "You need to suspend all access to Fable 5 and Mythos 5 to any foreign national, whether inside or outside the United States, including foreign national anthropic employees." Now, that's a nearly impossible task. They really had no way of shutting it down for everybody but American citizens. So the net effect, they must abruptly disable Fable 5 and Mythos 5 for all customers to ensure compliance. Now, Enthropic claimed it was over a very minor vulnerability and went on to talk about all of the ways they were sort of securing Anthropic so that it shouldn't be an issue. Now, in my opinion, Anthropic kind of brought this upon themselves back in April when they first announced Mythos. They pretty much restricted it for everybody because it was just too powerful to release openly. They literally claimed that it could wreak havoc in the wrong hands. The White House also claims they'd heard Amodai liken the dangers of anthropics technology to a nuclear bomb. So they literally spent months talking about how dangerous and how scary this model is. And then last week, just days before any of this happened, Dario published this essay, literally arguing that the government should be able to block or reverse a model's release, like FAA style. Frontier AI models, like airplanes, should be required to go through technical testing and auditing, and the release should be blocked or reversed as a threat to public safety if they do not meet high standards of safety. In the same essay, he said, "We now globally and collectively need to activate a slow and rickety policy apparatus to deal with risks and opportunities that are going to compound surprisingly quickly from here." And well, the government did exactly what he was claiming the government should do in his essay. And now the loudest advocate for AI regulation is frustrated that it's getting regulated. According to David Saxs here on X, a highly credible, trusted partner of both Anthropic and the US government who was testing Fable came forward with a jailbreak of the guardrails. The admin asked Daario to fix the jailbreak or deploy the model and Daario refused. Of course, in the blog post where Daario announced they were shutting down Fable, he did claim that the jailbreak isn't that serious. Now, according to David Saxs here, the admin issued the export control, but they did it reluctantly. He goes on to say, "It's frankly bewildering that Anthropic hasn't wanted to comply with safety requests that it previously said were its highest priority." And also, as a random side note that's like super fascinating about this whole thing is that the whistleblower, you know, the trusted partner that went to the US government, turns out that it was Amazon CEO Andy Jasse. Jasse was among the tech leaders who raised concerns to senior Trump administration officials this week about security risks and Anthropic's most advanced models. What makes this so interesting is that Amazon is one of Anthropic's biggest investors and vendors. And that's what makes it so bizarre is it doesn't seem to be in Amazon's benefit for this to actually get shut down because Amazon has so much invested in Anthropic. Now, my take on this is kind of that it's less about security and more about like the government anthropic butting heads. Like you get the real impression that the US government just does not like Anthropic. According to this article on political, the crux of the issue was the lack of seriousness that Enthropic was applying to it. Had Anthropic taken it seriously and rather than dismissing it as isolated, moved to fix or pause access, this would have never happened. Combine that with the drama of the US government deeming anthropic a supply chain risk. Pete Hedgeth sort of talking crap about Daario on Twitter. All of this kind of stuff makes it really evident to me that the US government just doesn't like Anthropic. Anthropic stands so heavily on their principles. And when Anthropic actually dismissed this as like a no big deal vulnerability, the US government went, "Fine, shut it all down. You're not taking us seriously." But again, the fact still stands that Daario did spend months saying that this is really, really scary. It could wreak havoc in the wrong hands and that this stuff absolutely needs to be regulated. Like he spent a lot of time saying that kind of stuff. Now, cyber security defenders are trying to come to the rescue and a bunch of them all signed this petition here to bring back these models. Basically saying, you're taking the best safeguard we have against cyber security issues away from us. going on to say to pull the best capabilities away from defenders without a good reason when our adversaries are rapidly advancing is dangerous. They also claim that these mythos class models are not uniquely good at these tasks. They do say they are quite good at finding flaws and weaponizing exploits, but they're not uniquely good at this. Basically, meaning the other models can do this stuff, too, and they're still out there. Now, there is a very good chance we'll be getting these models back pretty soon. I don't know how accurate this article is here, but according to Korea Junang Daily, Enthropic's confident of reenabling Mythos Fable 5 access in the coming days. Now, I think one of the bigger picture issues here is more about the precedent, right? Enthropic is worth almost a trillion dollars. The last valuation had it at 965 billion. If you're an investor in one of these AI companies, it makes it extremely scary that the government can basically say, "Shut this off." And you just have to do it overnight. Now, personally, I have very mixed feelings about this whole drama between Anthropic and the US government because again, on one hand, Anthropic was essentially begging the government to do this to them. On the other hand, I really liked the Fable model. It was really good at coding up apps quickly with just like one shot. If you watched my last week's news video, we made some really cool stuff with it. At the end of the day, we don't know exactly what went on behind the scenes with the government and anthropic. So almost everything we're hearing is just sort of speculation and you know sources told us. But it definitely does make, you know, investing in some of these foundation labs a little bit scarier of an idea. But again, there's a good chance we'll have it back pretty quickly. Maybe even by the time this video is live, Fable is starting to roll back out again. Another cool AI product I want to show you is from Box. If you're not familiar with them, Box is an intelligent content management platform that helps companies use AI to unlock information and insights hidden inside your content. And this is becoming a much bigger challenge than most people realize. Like I run both my YouTube business and the Future Tools site and newsletter. And even at that scale, I'm dealing with files spread across different folders and formats and tools and systems. Now imagine that problem across an entire business. Box's latest state of AI in the enterprise report found that 96% of organizations recognize that AI agents need access to company specific content, but only 36% have actually connected those agents to trusted content across their business. So the bottleneck isn't really that AI models anymore. It's getting the right information to them. That's where Box comes in. Box's AI features like Box Extract, Box Agent, and Box Automate help transform all that scattered content into structured, usable knowledge that can be searched, analyzed, summarized, compared, and put to work by both people and AI agents. And because companies don't want to bet everything on a single AI provider, Box is model agnostic, supporting models from providers like OpenAI, Anthropic, and Google. Their research found that 68% of organizations are concerned about vendor lockin. So this flexibility is huge for businesses and I personally like the option of jumping between models when I'm building with AI too. I can see this being especially valuable in industries like financial services, insurance, healthcare, government, and media where massive amounts of content need to be accessed securely and intelligently. Now, if you want to learn more about Box AI, click the link in the description below. And thanks so much to Box for supporting my channel and sponsoring this portion of today's video. We also got another brand new model this week that actually happens to be an open-source model. and people are saying it's really good. The company ZAI just dropped GLM 5.2. It's their new openweight flagship model that's built specifically for long horizon coding and agentic tasks. The model has a 1 million token context window, putting it pretty much on par with all of the Frontier Lab models, and it's got an MIT open source license, so theoretically you could download it and fine-tune it and do whatever you want with it. Now, according to Hugging Face here, it's a 753 billion parameter model. So, I mean, most computers aren't going to actually run this, but you can do whatever you want with it on the cloud. What makes this really interesting is that again, it's an open model, but on some of these benchmarks here, it's like up there with Opus 4.8. It actually beats out GPT 5.5. On Swebench Pro, it actually performs better than GPT 5.5, but not as good as Claude Opus 4.8. I don't really pay a ton of attention to this. I think the more important model is deep suite which you know it's not beating GPT 5.5 or Opus on that one specifically but still really really really good numbers for an openw weight model. If we look at the code arena here GLM 5.2 is actually coming in number two just behind Claude Fable 5. Now, this is a blind test where you give it a prompt, it gives you two responses back and users vote on the best response and that's how this score is decided. And based on the sort of user testing blind taste test here, the only model that's performing better on the webdev arena is Claude Fable. It's beating Opus 4.8. It's beating GPT 5.5 by a lot. When it comes to the agent arena here, it's right up here in the top 10 along with the anthropic and GPT models, beating out all of the Google models. So yeah, it's a good model and the fact that it's an open-source openweight model shows that these open models are catching up very very quickly. And the other thing to keep in mind here is if we come back to our code arena is look at the price per million tokens here. Claude Fable is $10 in $50 out. GLM 5.2 is $1.40 in $440 out. Opus 4.7 and 4.8 are $525. So, you're going to get like close to Claude Opus 4.8 level coding ability, but at like a quarter of the cost. So, according to all the benchmarks and according to the leaderboards here, this is the new state-of-the-art open model. So, of course, I had to put it to the test. I came over to Z.ai here and then selected the GLM 5.2 model and I wanted to test it with a similar prompt that I gave Fable. I wanted to see how good it can recreate that Megabon game. That's kind of like my current testing benchmark to compare to. So, I literally just gave it the prompt create a Mega Bonk clone. Now, I obviously gave it that prompt in English. And all of its response here, I believe, is in Chinese, but I can go ahead and select all this and translate section to English. Translate full page. And there we go. I've created a complete Mega style clone game for you. core gameplay, visual design, feedback system, and let's see what this looks like. This screen looks good here. Let's open it in a full screen here. And off to a good start. I mean, aesthetically, it looks good. Uh-oh. Nothing happens. I can't start bonking, apparently. All right. So, that was just one prompt. Let's go ahead and give it another prompt over here. And just say nothing happens when I click start bonking. I can only see the home screen. Now, while we're waiting for that, one of the other things that this GLM 5.2 model is good at is creating slides. So, I went and gave it the prompt, create a beautiful and accurate slide presentation for me that walks through what an agent loop is, how they work, and several useful examples of them. This time, it actually gave me the response in English. I don't know why it did it in Chinese on the last one, but it gave me this slide deck presentation. It's 10 slides and it opens it in another web page. A field guide to agent loops. How autonomous AI systems think, act, observe, and repeat until the work is done. A model that keeps going. An agent loop is the cyclic control flow that turns a oneshot language model into something that acts in the world, sees what happens, and tries again. So, I mean, it looks pretty clean. I like the little graph here. Nothing like ultra insanely impressive, but definitely clean. Four moves repeating, perceive, reason, act, observe, and then it loops through less than 20 lines of structure. The react loop, the original in 2022. So this says pattern one of four. So now I think it's giving us those use cases here. The tool use loop, the self-correcting loop, many loops, one goal. You've got the coordinator, agent with a researcher, analyst, writer, and critic. Loops are not free. Reach for a loop when the task is genuinely multi-step. Intermediate results change the plan. The agent has real tool calls. You want an auditable trail. skipped a loop when one forward pass will do. Latency is the product. Cost compounds per turn. You can't bound the iteration count. And then there's the final slide. Again, it's a clean presentation. Am I blown away? Not really, but it looks good. All right, jumping back to our meabon. We can see that it's writing code. It's actually responding to me in English this time. That's good. Okay, so it explains the problem to me and what it did to fix it. It appears to be done. Let's go ahead and try it. Still not working. All right. So, after the third prompt, it did work. So, here's what we got. I can finally start bonking. And this looks nothing like mega bonk. So, it had no idea, I think, what Mega Bong is. It just went off and made its own bouncy game. Also, spacebar is supposed to make me jump, but space bar is not doing anything at all. Okay, it's working sometimes. Space or click to jump. Space bar. Space bar. There it jumped that time. Now it jumped. It's like it it's janky. So, does it hold up to Fable level of coding? Definitely not. I do like the color scheme, though. If I'm looking for one sort of positive out of it, it's got a good color scheme. All right, let's ask our like gotcha question here. Instead of max, let's just go to high. My car needs to be washed. The car wash is 100 m from me. Should I walk or should I drive? You should definitely drive. If you walk to the car wash, your car will be left behind and won't get washed. Okay. Okay. So, it could answer that. At this point, though, I'm starting to wonder if like this is fine-tuned into the training data. I don't know. Anyway, it's a promising new model, and I don't think I've paid anything. Anyway, you can play with it over at chat.zai. And as far as I can tell, it is currently free to use cuz I haven't paid anything and I can't even find a way to pay if I wanted to. Meta rolled out some new AI features on Facebook. I haven't personally tried any of these yet cuz well I don't really use Facebook as much as I used to. I do have a Facebook account. I just don't think I've logged into it in like 6 months, 9 months. It it it's been a long time actually. Let me know in the comments if you use Facebook and if it's worth using. I just haven't in a long time and had no plans on going back. But let's check out some of these AI features because I don't know if this makes Facebook better or worse. So, Facebook has a new AI mode and it's a new way to get answers to your questions right on Facebook. So, AI mode uses meta to give you answers grounded in what people are saying publicly across our apps like groups and reels. So, this sounds to me very similar to what like Grock is doing with X, right? If you go and use Grock, it's grounded in all of the tweets or posts or whatever that are on X, which makes Grock really really good for like real time information and real time sentiment out of users because it is looking at what everybody's tweeting about. So, we can see in this example they typed summer escapes near me and then it says the Bay Area has tons of quick getaways, coastal towns, lakes, etc., etc. Think Half Moon Bay. It pulled up some images. guessing that people posted on Facebook and then it gave some additional prompts that they can use like wine country day trips from San Francisco or then it says popular day trips and apparently this is all being informed by what people are posting on Facebook. I guess they also introduced new editing capabilities that leverage AI. So inside your camera roll there's like collage cutout templates, there's new transition effects and apparently this stuff is optin only and can be turned off. There's also new photo presets, so you can actually change your clothing and hair. You can have like a picture where you're wearing the clothes you're wearing and then you press a button and it'll swap out your shirt to like be your favorite football team shirt. I mean, it's all kind of gimmicky. It's probably not going to make me start using Facebook. All right, there's one more quick rabbit hole I want to go down before we jump into a rapid fire, and that's the big announcement that came out of Midjourney. And if you didn't hear about this announcement yet, and you know what MidJourney is, you know, the AI image generation platform, you're never going to guess what they just released. Midjourney just started a new branch called Midjourney Medical. And the first thing that they showed off was essentially this like replacement for an MRI machine. It's this like dunk tank that like lowers you down into the water and then uses ultrasound to bounce sound waves around inside the water and then it uses those sound waves to get like an image of your body and your internals to detect issues and stuff. Basically, it's like designed to do what an MRI does, but at like 160th of the speed and onetenth of the cost, I think they said or maybe that's the other way around. Basically, it's like an MRI type machine that uh does it really quickly and really inexpensively. So, here's the video that they put out that sort of explains it. You've got all these little transducers here that are both speakers and microphones. And we can see as it zooms out here, there's like thousands and thousands of them, just shy of about 9,000 of these. And then these things are put in this like ring so that the audio waves that are coming out and being collected just bounce around inside the ring. And then of course your body goes down in these rings and then the sound waves are sort of bounced around everywhere to get that imaging of your body. You can see this is like the sound waves sort of bouncing around and creating that image. And they can see up to 25 different biological structures. So like you know spinal cord and your rectus abdominis muscles and your genome. And at the end of it, it creates like this whole scan of your body, similar to an MRI that can all be sort of sliced out and looked at in like layers. Now, you might be thinking, "That's cool. That seems like something that they'll put into hospitals and make it a lot less expensive to do scanning on people." That's not their plan. Their plan is to create what they call the midjourney spa. The first spa is going to be opening in San Francisco in 2027. And it's going to have hot tubs, saunas, cold plunges, cozy rooms, and of course, these body scanners. So, you're going to have like this whole spa-like experience where you can go get scanned and then go chill in the spa or the sauna or whatever. Like, they're trying to make this a cool, relaxing, cozy hangout where people want to go. And the idea being you go get scanned more frequently so that you can detect when problems happen because like you get scanned once and maybe it detects some stuff that could be completely benign but if you get scanned multiple times then you can see if things are like growing or if things are becoming more and more of an issue. One thing that makes this super interesting though is that Midjourney has no investors. They're completely bootstrapping this. they've made, you know, tons and tons of money from midjourney, and they're taking that money and they're using it to go and do this, like something actually useful in the health space, something that's actually going to cut the costs of something that was oftentimes prohibitively expensive for a lot of people. It sounds like they have this goal of like democratizing health and democratizing the ability to get these kinds of scans, which I think is a a cool and noble effort. Like, you want to see brain power going towards stuff like this. Now, if you're wondering why MidJourney would be the company to go and do something like this, well, I see a couple reasons. One, they're like an imaging company. So, if you if you don't know much about David Holtz, the CEO and founder of Midjourney, before Midjourney, he had a sensor company that made it so like you could do motion tracking with your hands. Like there was little sensors that can see what your hands were doing and so you can like drag things around on your screen and almost do like Minority Report sort of stuff with your hands and it manipulated what was going on on your screen. He has a background in sensors, right? And then he got into Midjourney, which is image generation. This almost feels like the evolution of those two things, just sort of like smashing those kinds of projects together, like sensors and image generation. And I think the bigger play here is probably more of like a data collection for AI play. like if he gets hundreds of thousands of people in these scanners, well, now he's getting tons and tons and tons of data on people's like bodies and can probably better help predict things based on more and more data that they're able to collect. So, I think there's like a bigger AI play here where they're going to likely train models on all of the scans that they collect. I don't know, though. I'm just speculating. But, there are some people that aren't convinced. Hank Green here posted this on X called let's talk about whole body ultrasounds. He personally doesn't like the fact that they're comparing it to MRIs and CT scans. Claiming that all of these scans are better at different things. So basically claiming this is a replacement for an MRI is not quite correct because an MRI is going to do certain things a lot better than what this is going to do. So he says here MRIs are bad at imaging parts of your body that move. That's why we don't use MRIs to screen for colon cancer or lung cancer. A chest CT is much more likely to catch an early lung cancer. While a colonoscopy can identify colon cancers, different scans do different things. Now, ultrasound is really good at soft tissue. So, fluid fil things and things close enough to the surface that sound can get in and useful information can get back out. So he's saying it should be great for looking at thyroid or a breast lump or your testicles or your lymph nodes or gallbladders, kidneys, fetus, blood flow, flute around organs. Lots of abdominal stuff. Ultrasounds will be good at that kind of stuff. But there's also stuff that ultrasounds will not be so good at. Air is a problem because sound waves do not pass smoothly from tissue into air. They mostly bounce. That's why the scanner needs to happen underwater because if you were just in air, the stuff's not going to bounce around properly. But there's also air inside your body, like in your lungs. So, the ultrasound will be able to tell you things that are useful about the surface of your lungs, but it's not going to look through your lungs the way a CT scan can. Basically, his argument is don't compare this to a CT scan or an MRI because those will do things that this just can't. Now, I think he's hoping this thing is a success and will do well. And hopefully AI will be able to fix some of the things that are the shortcomings of a product like this, but again, the point that he's making is that don't say it can do everything an MRI can do when it just can't. and don't say it can do everything a CT scan can. That is just not true. It won't be able to do all of the exact same things. So, I guess Hank's argument here is more the way they're marketing it and not the fact that it's not something that's really actually kind of cool. And for me, I just wanted to talk about it because even though at the moment it has nothing to do with AI. I mean, there's future AI implications of it, but right now he flat out said there's not even any AI in the system. I just think it's absolutely fascinating that this is the next thing MidJourney was working on. Anyway, I've spent a lot of time covering just a few things, and I have quite a bit more I want to tell you about, so I'll just kind of rattle through them quickly in a rapid fire. Starting with a cool new feature that OpenAI rolled out inside of their codeex platform called record and replay. So this appears to be something for if you have like a repetitive task, you can call upon this record and replay that will record a video of you doing something on your computer, learn how to do that thing, and then in the future you can just tell it to do that thing again. So the example they give here is watch me upload this YouTube video so you can handle these uploads for me in the future. We can see down here that it starts recording his screen. He goes through the process of uploading a YouTube video, grabbing some information from a spreadsheet, you know, grabbing the files on his computer, setting it as private, and then he types done. And then it watches the video, understands what he did, creates a new skill out of it, and now he can go back and he gives it a PNG file, I'm guessing, for the thumbnail, a video file, and a subtitle file. Drags and drops them into his codeex chat box and says, "Upload this YouTube video using the YouTube upload skill that we created." It then goes and does the exact same steps. It selects the file. It grabs it from his computer. It uploads it. Grabs the thumbnail. Grabs the subtitle file. Saves it as private. So, it followed his exact step-by-step instructions that he did in the original video. That's pretty handy. Claude Design rolled out a new feature. You can now edit directly on the canvas inside of Claude Design. So, you can have it generate a design and then just edit it straight from within there. You can now bring in designs from GitHub repos, design files, or raw uploads. and Claude builds with your components, checks his output against your design system and makes corrections before you even see it. They've also made it easier to move back and forth between Claude and some of your favorite tools. So stuff like Adobe Base 44, Canva, Gamma, Lovable, Miro, Replet, Verscell, Wix, and more destinations soon. So you can actually, you know, design in one and move between them fairly easily. And these new features are available on pretty much any of the paid plans. Perplexity rolled out a new feature called self-improving memory for agents. They say this new brain feature is a self-improving memory system. It builds a context graph of the work computer performs. At set intervals, such as overnight, brain reviews the context graph and teaches itself how to do the work better. The more work you do, the better and more efficient brain makes your computer. To me, this sounds a lot like what that Hermes agent is doing that everybody's raving about lately, where it watches what you're doing, learns what you're doing, learns where it's making mistakes, and then constantly improves itself. Except now you're doing it inside of Perplexity Computer. Brain gets better as you use computer. Agents become more effective at updating context as they learn the projects, connectors, artifacts, and other sources that lead to the best outputs. They also learn from their mistakes, remembering when a user has made a correction or when a source was a dead end. This results in fewer turns, fewer model calls, and better outputs. This feedback loop makes brain continuously self-improving. So, Perplexity Computer is essentially like Perplexity's answer to OpenClaw or Hermes. It's like their agent, but it runs in a cloud instead of on your own computer. They also have what they call personal computer, which does run on your own computer, but I think this one just runs on their cloud computers. I didn't see anything in their write up about personal computer. Now, in order to use this, you do have to be on a Mac or enterprise plan. So, one of their higherend subscriptions. I thought this looked interesting as well. I haven't tested it myself yet, but this company Palmir made a video editor where you use Claude to actually edit the videos. So, you can see here they've got their Palmir video editor open, and they've also got Claude open in another window. They say, "Organize my media in Palmir." And then over in their editor window, you can see all their files get sorted into just a few folders. And then they say, "Now edit it for me." And it says, "Good. Now applying trim, speeds, and transforms." And it looks like it's just like editing it in the timeline for them and adding sound effects and adding text. And then it uses the various like video models that are out there to actually generate video. So you can see like this i video is all generated. So, you use Claude in combination with this video editor, and it will do things like organize your media for you, chop up the video for you, actually generate media that can go into the video if you don't have the exact shot that you need. Something I definitely want to test out. Although, when I look at their website, it costs 29 bucks a month. Normally 49 bucks a month for 5,000 credits. Now, 5,000 credits gets you 3 to 7 minutes of generated video. I don't know. That seems like that could add up very, very, very quickly if you're using a lot of generated video in whatever you're editing. But still, I'll probably give it a try. Maybe test it out in a future video. Since we're talking about video editing, Adobe is adding its AI assistant to Premiere, Illustrator, and Inesign. So, you can do stuff inside of Premiere like make a new sequence with the drone footage or generate images through prompts like this. The idea being that you can use these tools here, Premiere, Illustrator, Inesign, but just sort of text prompt it with what you want instead of actually doing the editing the oldfashioned way. Here's one for my marketer friends. Google introduced what they call Ask Ad Manager. And to me, this looks like pretty much the same thing as ask Studio that we have in YouTube where in the back end of YouTube, you can click on a video and ask Studio, hey, what is this video doing well? What should be improved? and their little AI will look at all the stats on the video and tell you what you could do to improve that video. Well, you're getting that same functionality inside of the ad manager now. So, you could be running ads inside of Google Ads and get a little chatbot to figure out how to improve or optimize your ads or, you know, get additional data out of your ads, things like that. I used to run a lot of Google ads in a previous life when I did a lot of digital marketing and this would have been really really cool to have. Pew Research did a survey recently where they asked Americans about their thoughts on AI and there was some interesting information that came out of it. Here's the key takeaways. About half of US adults now report using AI chat bots. We can see back in 2024 about 33% said they used them. 66% said they didn't. 2026 it's now 4951. But Americans including younger adults are deeply skeptical of AI. more adults predict that AI will have a negative rather than positive impact on them and on society. So, both things are true at the same time. More and more and more people are actually actively using AI while also more and more people are actively saying they're skeptical that AI is actually going to benefit humanity as a whole. And I actually think that makes sense if you think about it because we're sort of lumping all AI together is just like this one thing. There's just AI, right? But AI is like a whole bunch of things, right? You had video generators like VO and C dance and gromagic. You have image generators like Nano Banana and Chat GPT image. You've got coding tools like Codeex and Claude Code. You've got music generators like Sunno and Udo. You've got text to speech generators like 11 Labs. And of course, you've got chat bots like Chatg GPT and Gemini and Claude. But then you do surveys like this and say, "How do you feel about AI?" Well, it's nuanced. like you're allowed to think, okay, I really don't like AI generated music because it's starting to pop up more and more on Spotify and I don't like hearing AI generated stuff on Spotify, but Catchy PT is kind of cool and helps me do my work a little bit faster, so I like that. And every time I get on Instagram or YouTube, I see shorts of like AI generated slop videos and I'm getting sick of seeing that. All of those things can be true. you can like this part of AI that AI is doing and you can hate this part of AI that is not making your life better. And so you go and do these surveys and yeah, more and more people are using AI cuz chat GPT genuinely does have helpful use cases for most people. It could make me do my work a little more productively. I could take a picture and say, "What's this weird rash?" and get an idea before going to a doctor, right? Like there's things that make it useful to like everybody. But then there's also parts of AI like AI generated slop music or AI generated videos that are popping up on YouTube shorts and getting millions of views that a lot of people are sick of and getting overwhelmed by and not wanting to see it. And there's, you know, fake images popping up that look like real images like, oh, this building just got bombed in this foreign country and you believe it's real for a minute and then everybody starts telling you, oh, that's AI and now you don't know what to believe. And there's things that positively impact people with from AI and there's things that negatively impact people with AI. And so I think going around and surveying and saying, "Do you like AI or do you not like AI?" Well, that's kind of a loaded question because it's very nuanced and these surveys don't really like nuance. Anyway, rant over. Oh, and if you're curious which chatbot gets the most use, apparently ChatGpt still dominates with 44% of Americans claiming they use Chat GPT, which is interesting because it's easy to get into a bubble and hear everybody talking about Claude and think that everybody uses Claude, but I think in the general public, ChatGpt is still like the Kleenex or the band-aid of AI. Like, they're synonyms to most people. And finally, I'd like to end on robots. And well, this robot is one that'll take your crap. It's a self-driving toilet. So, this company Shiaoban will safely navigate your home, clean up after users, and empty itself all on its own. That's right, folks. We're getting self-driving toilets. So, this toilet sits on a dock, and when you need it, it can be summoned over to your bedside and positioned exactly where you need it so that you can uh shift over to it and then do your business. It's got like a bedet feature to, you know, clean you up. It's self-cleaning. So, the toilet bowl scrubs itself and cleans itself when it's done. And then it actually drives itself over to your toilet and then has a mechanism to empty what's in the toilet, the driving toilet, into your regular toilet. So, it will dump that waste into your toilet. And then once it's done dumping the waste into your toilet, it then cleans itself and scrubs itself all out, searches the room again, and then when it's all done with that, it goes and docks back up to charge and refills its water tank. So yeah, I mean, this is obviously designed for people with mobility issues that need the toilet to come to them. But come on, you know, some lazy rich people are going to get this, too. But that's what I got for you today. A lot of interesting stuff happening right now in the world of AI. I do record these videos on Thursdays and publish them on Friday. So, if any new news came out on Friday that I missed in this video, well, it'll likely make next week's video. There is so much happening in the world of AI every single day. Like, we are just getting flooded with news announcements constantly, and it's my goal with this channel to personally keep up with it every single day. Drink from the fire hose. I'll be overwhelmed so you don't have to. And then every Friday, I'll make a video where I break down all the news that I think is most important for the most amount of people to know. I'll try to cut through the noise and the hype and just give you the signal so that well, you don't have to feel overwhelmed by this cuz trust me, it's really, really easy to get overwhelmed. That's my goal is to help you out and prevent that. If that's something that interests you, consider liking this video and subscribing to this channel and I'll make sure more videos like this show up in your YouTube feed. But again, that's what I got for you today. I really really appreciate you hanging out with me, nerding out with me. I'm having so much fun keeping up with all of this stuff and, you know, testing it myself and then turning around and figuring out what's worth talking about. And hopefully you learned something and it's helpful to you. Really, really appreciate you. Hopefully I'll see you in the next one. Bye-bye. What's up everyone? Glad you could make it.
If you have always wanted to make your footage cinematic with VFX, you know it usually means learning After Effects, hiring pros, or buying gear. But, I just found a workflow that completely changed how I do this, and it runs entirely on AI in a few clicks. In this video, I'm going to show you exactly how it works so that you can transform any footage you already have into something fully cinematic. The tool I'm using for all of this is called Open Art, which has a dedicated VFX section that allows you to edit existing footage without ever needing a green screen. If you want to follow along, I've left a link to Open Art in the description [music] below. What makes Open Art VFX different from every other AI video tool is that [music] it doesn't generate anything from scratch. You just give it something that already works and change specific elements while everything else stays exactly the same. So, I have this video of myself right here in a plain studio, and the first thing I want to change is my outfit. Normally, inpainting is done on images where you paint over one part of a photo and swap it for something else. But, Open Art lets you do the exact same thing across a whole video. So, when you first log in, you'll see the main navigation bar. [music] To access the VFX tools, click on video and then select VFX. I'll upload my video and wait for it to load. And then, under the areas to keep section, I'll click select mask and paint over my outfit. Make sure you don't paint over your face so the AI can keep it the same. And then, click invert this area. After that, you can simply hit preview to see what the final cutout will look like. And once you're satisfied, click confirm to save your selection. To perform the replacement, we can either select an existing image or generate one using AI. Since I haven't created any images before, I'll select create with AI. I want to swap out my outfit for a magician suit. And for that, I don't need to worry about writing very long complex prompts. Since Open Art gives you access to the best models like Nano Banana Pro and GPT Image 2, which have a high level of comprehension and deliver excellent results. So, I'll just type that, select Nano Banana 2, and generate. The video is exactly the same as the original, but now I'm wearing a tuxedo and it's tracking every bit of my movement right down to me adjusting my sleeve. Things like my face, [music] my movement, and the lighting remain perfectly consistent. The only thing that's different is the area I originally invert selected. And the best part is that this is repeatable. I can take that result and put it straight back in to change something else on the same clip. This time, I'll select this chair, invert the area, and ask to change its color from black to white. I'll type in my prompt and generate. I'm still wearing my tuxedo, and now the chair is a completely different color. Changes like this would normally require hours of editing in After Effects, but I just did it in a few minutes. With OpenArt VFX, you can change specific elements in a video, but now I want to change the whole environment. And OpenArt has a tool built for exactly that, called replace background. Back on the main navigation bar, hover over video and click VFX, but this time select replace background. I'll upload the final video we made with the tuxedo and the white chair, since every change I already made carries straight into this step. With this tool, you can keep the main subject intact and rebuild the whole scene around them using a reference background of your choice. So, there's no green screen and no cutting yourself out by hand. You just tell it what should be behind you, and it builds the entire scene like that. You can upload your own background or use one of OpenArt's presets. For this one, I'll go through a few from the existing library to show you the range. First, I'll select this one called Clean Creator Studio, and in the prompt box, I'll simply describe what else I want in the scene and generate. This one is similar to the original video, since they're both in a studio, but the whole thing feels way more professional. My face, movement, and in-painted features have all been carried over smoothly. And my favorite thing is how grounded I look in the new room, and it's the kind of professional set you'd normally have to rent or build just to record a few videos. Now, to show how different it can get, I'll swap the background reference for a cozy bookstore cafe. I'll type in my description and generate. What I like about this is the depth. The shelves behind me blur out of focus the way a real camera lens would do it. So, it looks like an actual video shoot. And my tuxedo hasn't changed at all. The only thing that got rebuilt is the entire room. This is the background you'd reach for if you wanted something warmer and more personal than a studio, like a vlog or a sit-down story. And then for the biggest jump, I'll change it one more time and select this cinematic desert highway at sunset. This is a much harder task than swapping one indoor room for another because now it's a full outdoor scene with real distance behind me. So, let's see how it handles that by typing this and hitting generate. Even with all that extra difficulty, OpenArt did a very solid job. It's still me with every element intact, but I'm now standing on a desert road with the sky going on for miles. And it even added cars passing behind me on the highway, which brought even more realism to the scene. And honestly, that's what impressed me the most. All three are just as good if not better with what you'd get from a green screen. And switching between them can be done in seconds. So, from a single pre-existing video that I filmed, I can change any element I want and swap out the background. This way, you can create content at scale without ever leaving your desk. One thing you may have noticed in all the results is [music] how well OpenArt adapts the character's lighting to the new environment, bringing even more immersion to the scene. This relighting process is something you might also want to do independently. Whether you want to salvage a bad recording, make a scene more realistic, or simply bring a new mood to your project. And we have a specific function inside the platform that does it for you in seconds. Still inside the VFX workspace, this time select relight. What this tool does is change the lighting across the whole shot at once, the light on you and the light on everything behind you, so the two finally match instead of looking like two shots stuck together. Your eye checks the light before anything else, which way it comes from, whether it's warm or cool, how soft or harsh it is. And that tells you what kind of place and what time of day you're looking at. So, if the lighting in your videos is wrong, the whole shot stops looking real, no matter how good the background behind you is. Now, instead of carrying on with the video we've been building, I'll start fresh with a different video of myself. That way, we can really focus on the lighting and how it alone can affect the entire scene. So, with that clip loaded, I'll open the lighting presets. And now you'll notice there are three different categories: lighting, mood, and atmosphere. Each one affects how your video will be lit, but in a unique way. Lighting focuses only on the actual light source and how it interacts with the environment. Mood is more about the overall aesthetic, and atmosphere is more broad with different scenery. First, I'll pick one called moonlit blue under the lighting category and describe my scene. We get back the same clip of me waving, but instead of plain daylight, it now looks like it was shot at night. There's a cool blue light across my face with soft shadows, and the whole thing is suddenly calm and quiet. Nothing else moved. I didn't touch the background or reshoot anything. The light by itself just changed the whole vibe of this scene. Now, let's try something different. I'll keep the exact same clip, but I'll switch the preset to neon rim light under the lighting category, and I'll paste in my description. This is my favorite so far. The whole scene is now lit by neon lights, and yet everything looks realistic. This is the lighting you'd normally see in a music studio. That's the whole point of this tool. The lighting decides the entire mood of a shot, and you can get a completely different feel out of one clip in a single generation. And just so it's clear, you can absolutely run this on top of everything else. Take that tuxedo and desert clip from earlier and relight it the same way. So, all three tools stack into one finished shot. The whole process is chainable and made to produce content at scale. I just wanted to show relight on its own first, so you could see for yourself how much lighting can actually change a scene. So, out of one video I already filmed, I've changed my outfit, my whole location, and the entire mood of the shot, and I never touched a green screen or After Effects to do it. And because everything starts from the same source footage, the final result still feels natural and consistent. What would normally take multiple shoots and hours of editing can now be done with just a few generations. And the best part is that all three of these tools exist under a single tool. And there's honestly way more looks and presets in there than the ones I showed you. So, if you want to make your own cinematic content with VFX, use the link in the description to sign up to OpenArt. Thanks for watching and I'll see you in the next one.
Most people pick a logo tool because it showed up in a recommendation or looked popular. That is genuinely the wrong way to choose. I tested five AI logo platforms this week using the exact same brief on every single one. The results were not remotely equal. The gap between the strongest output and the weakest was bigger than I expected going in. Here is what each platform actually produced and which one I would use on a [music] real project. Getting a professional logo made still costs real money in 2026. A freelance designer will typically quote anywhere from $300 to $1,500 for a basic brand identity. An agency starts at 5,000. For a founder or creator who needs a brand live this week, neither option fits the moment. AI logo platforms change the math significantly, but they did not all change it by the same amount. Today, I am walking through five of the most widely used logo platforms available right now. The test business is a fictional interior design studio called Amber Oak Studio. Every platform received the same business name and the same keywords, so the only variable is the tool. The first platform is design.com, and this section is sponsored by them. I want to be upfront about that before anything else. I tested the platform on my own before I agreed to work with them. What I am about to walk through is the actual workflow I ran for real. I did not put this together specifically for the video. design.com positions itself as a full AI design platform. The difference from the other tools on this list shows up immediately in how the generation space works. You enter a business name and a set of keywords. The platform uses AI to tailor designs to your specific input. Business name, category, and aesthetic direction all factor into what comes back. What you see in the grid is shaped around your brief, not a generic browse through all available options. The first generation for Amber Oak Studio came back noticeably strong. Several options in the grid were close to commercially ready without editing. The AI pulled a warm palette that suited the interior design direction well without any manual palette selection. I was not expecting a good result from the first batch. That matters if you need to move fast. The fewer rounds of editing you go through, the sooner you actually have a brand. The editing interface is where the AI first position becomes concrete. Instead of digging through a color picker or a font menu, you can type what you want and the AI interprets it as a design instruction and applies the change directly. The manual controls are still there if you prefer them, but the chat approach is genuinely faster for directional changes. That instruction came back applied correctly in about 5 seconds. The result did not need any cleanup afterward. One thing worth flagging here, the AI editing works best when your instructions are specific. I typed exactly what I wanted, a specific color direction and a lighter mark weight, and it came back correctly. Vague requests like make it more modern tend to give less predictable results. Here is where the platform goes beyond a logo tool. Once the logo is confirmed, design.com auto generates a brand kit from it. Every design template inside the platform automatically inherits the logo's colors and fonts. You do not re-enter your brand colors when you switch to a new design format. You lock in the logo once and the system carries the entire identity forward from that point. Every design is 100% commercially safe for real business use. One quick note, AI-assisted logos on a standard license are not exclusive by default. So, if trademark protection matters for your business, the extended license is the one to get. If you want to try it on your own project, head to design.com. The free tier lets you run the full AI generation workflow and preview the brand kit without paying anything. You only need to upgrade when you want to download your final files. Premium starts at $3 a month on annual billing. Month-to-month is available, too, if you prefer flexibility. The second platform on this list is BrandCrowd. This one has a strong following among small business owners looking for a specific visual style quickly. The search-based discovery flow is good at surfacing something on brief. The quality at the top of the results is genuinely solid. BrandCrowd starts with a business name input and generates a grid of logo options organized by style category. The library is large, and the search engine is good at surfacing relevant styles quickly. For Ember Oak Studio, the first batch came back with a strong selection. The botanical and organic directions were the best match for the interior design brief. Several options were close to ready without significant editing. The customization experience inside BrandCrowd is manual. You work with color swatches and font drop-downs. The layout controls let you reposition elements. The output improves noticeably when you invest time in the editing step. The tool is not doing that work for you with AI. The free tier lets you preview your designs and download a watermarked low-res PNG. High-resolution and vector files require a paid plan. You can either subscribe or buy individual logos as a one-time purchase. If you need a print-ready SVG for real business use, the paid tier is where you end up. BrandCrowd produces solid logo output, and the library depth shows up in the variety of directions you get back. The limitation is the editing model. Customization is manual rather than AI-guided. If you are comfortable working through a traditional editing interface, this one will serve you well. The third platform on the list is Canva. If you have made anything online in the past few years, you have almost certainly used it for something already. Canva is the most widely used design platform on this list by a large margin. That popularity is worth examining closely when you try to use it specifically for a logo. Canva approaches logo creation differently from a dedicated logo platform. You start by selecting a style category and browsing a grid of AI assisted designs. The AI helps you customize the direction you choose. It is not working from your business name or building something around your specific brief. For Ember Oak Studio, I searched the interior design logo category inside Canva. The results came back with a solid range. Clean word marks were the strongest direction, and the template quality at the top of the grid was genuinely good. The editing experience is manual and familiar. You move elements directly and swap fonts from a drop-down. Changing colors works through a color picker. Most people who have used Canva before will be comfortable within 30 seconds. But the AI is in a supporting role here. It is not making design decisions for you. Downloading a vector file requires a Canva Pro subscription. The free tier gives you a PNG. That works fine for digital use, but not for print-ready output. Canva is genuinely excellent at what it does. It just was not built specifically around logo first brand identity work. If you are a pro subscriber and you already live in the Canva ecosystem, the logo tools are more than capable. The trade-off is that you are using a broad design suite that happens to do logos well, not a platform built from the ground up around that problem. The fourth platform is Wix Logo Maker. If you are already on Wix for your website, this one has a natural integration that makes it worth knowing about. Outside of the Wix ecosystem, the story is very different. Wix Logo Maker starts with a short preference questionnaire before generating anything. The questions cover your industry type and general aesthetic direction. The AI uses those answers to filter its output before the first batch appears. For Ember Oak Studio, the first generation came back in warm earth tones across three distinct directions. The botanical illustration was the strongest option in the batch. The editing experience covers both manual controls and an AI chat interface. You can use color pickers and font drop-downs directly, or type instructions in plain English and let the AI apply them. The chat feature works. It is just not as refined as what design.com offers. Here is where the Wix integration earns its place. If you download the logo and build a site on Wix, the brand colors flow into your website automatically. That connection between logo and site is the platform's real differentiator. For Wix subscribers, the logo tool is effectively free, and the site integration is the best on this list. For everyone else, the case for it weakens considerably. The standalone logo output is solid, but it does not justify switching platforms if you are not already on Wix. The fifth platform on this list is Vistaprint. This one is primarily known as a print-on-demand service. You can order business cards and branded apparel directly through the platform. The logo maker is built into that ecosystem. Knowing that context is key to understanding who Vistaprint is actually for. Vistaprint's logo maker follows the standard entry flow. You enter a business name and select an industry. The style preferences come after. The generation quality is more basic than the other platforms on this list. The output serves as a functional starting point, but it does not match the quality level of design.com or BrandCrowd. For Ember Oak Studio, the first batch came back with simpler mark options. The pallet choices were standard industry defaults rather than something tailored to the brief. The results were usable but not exceptional. Vistaprint's logo maker lets you create and download your final files completely free. High-res PNG, SVG, and PDF with no watermark and no purchase required. Where the platform really earns its place is what comes next. Once you have a logo, you can apply it directly to physical products and order everything from the same platform. There is no exporting files or uploading to a separate print service. Vistaprint makes the most sense if your primary goal is getting a logo onto physical products quickly. The logo maker output is weaker than the other platforms but the print ordering workflow is the best on this list. If you are launching a local business and you need a logo on business cards this week, this is a reasonable path. Here is how the five platforms rank after testing the same brief across all of them. At number five, Vistaprint performs best as a print service rather than a design platform. If physical products are your priority, it earns its place. For pure logo quality, it sits at the bottom of this list. At number four, Wix logo maker is the right call if you are building a site on Wix. The integration is seamless and the cost is effectively zero for existing subscribers. Outside of Wix, the case for it does not hold up well. At number three, Canva is the strongest general-purpose option if you are already a pro subscriber. The template range is genuinely excellent. For dedicated brand identity work, it gives up ground to platforms built specifically for that purpose. At number two, BrandCrowd produces strong output and the library depth is real. If you know the aesthetic you want and you are comfortable doing manual edits, BrandCrowd is a solid choice. At number one, design.com is where I would start for most people building a brand from scratch. [music] The first batch quality stood out across the test, noticeably stronger than the other four tools on the same brief. The brand kit auto applies across everything inside the platform, and the free tier gives you the full experience before you decide whether to pay. It is the most complete package on this list for someone who needs a logo and a full brand identity in the same session. I tested five AI logo platforms with one brief and got five genuinely different results. My pick out of these five is design.com. I am not saying the other tools are bad, Brandcrowd and Canva both produce solid work depending on what you need, but design.com is the only platform here where your logo and your brand identity come out of the same process. You generate one and get the foundation for the other automatically. That is the part that saved me the most time. Not just on the logo itself, drop a screenshot in the comments if you try design.com on your own project.
I almost burned out working on my own business because I was doing every single task myself. My emails were all piling up and I had meetings to recap and a bunch of content to create and hundreds of small tasks that took up my day before I got to the important stuff. Doing it all myself cost me a full day of work almost every single week. So, here is the exact workflow that I automated with a tool called Syntra AI. Instead of just being a chatbot, Syntra gives you a team of AI helpers, each with its own job. For example, Soci creates and schedules a full week of social posts. Cassie, on the other my inbox. Busy takes my meeting notes. And Comment built me a full website just by chatting with it. And the best part about it is that you don't need any technical skills to set any of this up. Whether you're a founder or a freelancer or running a side hustle, you just pick what you want done and connect the tools that you're already using, like Gmail or your Google Docs or calendar or Instagram, and the helper runs with it and builds on top of that. So, if you're tired of juggling all these tools, doing everything yourself, this is definitely worth a look. Use the code build85 for 85% off and use the code and the link in the description for the discount. Follow for more AI tools that will actually save you time.
If you're a marketer, content creator, knowledge worker, and you're using Claude code, Codex, or some AI coding agent, you're in the top 1% easy, right? And I always tell people like especially in the marketing industry, I'm a marketer, I'm not a developer, or software engineer, anything like that. People just can't comprehend that because we live in such a bubble. It seems like everyone is doing way more than us, but that's just not reality. If you go down the street and start talking to these local businesses,
This is a great week for generative AI, and I have some interesting things to show you today. Amongst a plethora of ChatGPT updates, you can finally schedule tasks properly. Oh my god, can't wait to show you this. They made a few quality of life improvements, and the Gmail connector can now actually send emails. When I look into my sent emails, hello, I hope you're well. Yeah, it did it. Amongst that and the Fable 5 discussion and a Chinese open-source model that came out that is as good as GPT 5.5 or Opus 4.8. Yes, you heard that right. It's been a very interesting and practical week. So, let's get into it in this week's episode of AI News and Views where we look at all the releases in generative AI, we filter for the ones that actually matter to a non-technical consumer trying to use this stuff, and then I, Igor, have the honor of presenting it back to you. Let's begin. And our journey this week begins with the scheduled tasks, a big and long overdue update. They had a version of this, but in short, it was so bad it wasn't even usable. I'm not going to go any further into that, but just know that the new version actually works. You can access it in the sidebar. There's a new tab called scheduled, and within here you can see, well, if you're brand new, a bunch of suggestions on what you could get. A World Cup recap or a daily news brief. Or you could set up alerts for concerts. All of these are great ideas, and usually the sweet spot for the scheduled tasks, especially as people get into them, is a combination of doing some information research and googling, and then filtering that for your taste or needs. That's why all of these examples follow that formula. Hey, I'm interested in the World Cup, particularly maybe one team, and I want to get all the updates about what happens. Get me a summary every day. And then it gets your personalized summary. So, what I did is I used the daily briefing preset, which honestly is very, very basic. All of these prompts here are like one or two liners. It's going to be up to you to add more context to that, but that's not the point of this video. The point of this video is showing you the scheduled task. So, if you create a new one here under edit, you can see the layout. There's a prompt up top. This is kind of the basic prompt that it comes up with. There's a frequency, which basically says how often this is supposed to run, and a repeat. So, here's how it works. The repeat field dictates what shows up underneath. So, if you have it set to daily, it gives you an option. Do you want it in the morning, afternoon, evening, or night? Okay. Now, if I set it to hourly, which is the maximum frequency, and the frequency of this will depend on which plan you're on. So, you might not see all of these. But, on the pro plan, with hourly, I could say every 2 hours, for example. And if I go to custom, I actually get less choices than with the others. It's only turns into daily, weekly, monthly. So, really, you want to make your choice in the first one, and then the rest makes sense. Also, down here, you can set when this expires. They did make a comment in their release notes that they might turn off some of your tasks if you haven't used them in a while. I mean, I get it. They just don't want millions of people running tasks on a daily that they don't even use. But, basically, you set this to never, and now, every day in the morning, I will receive a daily news brief that looks something like this. I ran the first one here for you. Again, it's based on this prompt, and it, yeah, finds Fable 5 and Mythos 5 stories. There's all the other stuff. And that's a feature. I recommend you start with these daily news briefings. I recommend you add some of your own context. And over time, my goal would be to refine the prompt so that these results actually match my interest. If I'm looking at the results, and I'm like, "Okay, two out of four stories are completely relevant to me," then you have some work to do in the prompt to make sure that those two stories would be filtered out on the run tomorrow. That's how you improve it over time. It will take a little bit of work. Probably the easiest way to do that is just to copy this prompt into a new chat, and then give it context on what you're trying to achieve with the daily task, and what stories you didn't like, and how you could modify that prompt now. It will help you. All right. Next ChatGPT update. They added interactive visuals. And somebody at OpenAI basically took their mouse and keyboard, went over to Claude, saw the feature, and they're like, "Oh, this is great." Copy, ChatGPT. Paste. >> King in the castle, king in the castle. >> Because this is a feature that Anthropic released a few months ago. And yeah, here you can see some comparisons between the same prompts being run both in ChatGPT on the left and Anthropic's Claude on the right. Across like five test prompts, we found one difference where Claude actually failed on this one. Create a scatter plot showing the 2025 and 2026 NBA teams. And then while plotting the payrolls, ChatGPT actually nailed it while Claude made a mistake here. I mean, that's one case. The others looked really good. Then in another case, ChatGPT refused to make a pie chart that we asked for. But then on this comparison, there was a little bit of overlap of the text in ChatGPT while Claude nailed it. Potato, potato, they're different implementations. Pretty good. These visuals, as a reminder, are meant to be a quick visual aid while you're exploring a new topic. So, for that, it's amazing that they have them now and they work pretty well in both services. Okay, next up they made some model changes, too. I actually really like this change. Look, they just cleaned things up. When you pick your model, they removed some older ones, so that's good. But then even if you're on GPT 5.5, which is the newest one and probably the default that everybody should just be using all the time unless you have a really, really, really particular reason that's based on your experience. And when within the model, you now get these options: instant, medium, high, extra high, or pro. And then behind pro, there's a pro extended that actually hides in there. So, if you're on a pro account, this is kind of a good button to know. This is how you make ChatGPT work the hardest. It will also take up the most compute from them, so they hid it away here. The release notes show this beautifully. You can see the before on the left side and the after on the right side. So, basically what happened is all of these different levels of thinking were removed and now it's just something a bit more intuitive. And in short, if you're not familiar with what these do, it's just basically how often AI prompts itself before it answers to you. So, on instant, you're going to ask it, "Hey, give me 10 ideas for XYZ." and it's going to give you 10 ideas right away. If it's on pro extended, it's going to generate those 10 ideas, but then it's going to ask itself like 50 or 100 different follow-up questions to stress test those ideas or evaluate if it actually read your intent correctly and how it could make these better. That thinking process that goes in the background really changes based on what you give it. So, if I ask for 10 ideas for a dog food brand, um, duh, I suppose, you're going to see this is going to think forever. And if you click it, you can see the thinking process. And if you want me to save you some time, any question that is not complex, just stick to the lower ones. Added value is barely there with these. As soon as you're going to something that would take a human hours to work through, that's where you can go to these higher ones. That's a good rule of thumb. Just know that ChatGPT takes the freedom and a little bit of flexibility there, too. So, you can see even though with the biggest thinking model here, it only thought for 40 seconds because this is a simple problem. So, they're not going to waste their compute just because you tell them to. But, the higher you go, basically, the more possibility of compute being used on your question, you open up. Whereas, if you go to instant, it's always just going to be just that. And the rest is a ladder in between. Few more super minor updates. On iOS, when you were sending messages with attachments, you couldn't change anything. So, when you attached an image and you wanted to change the prompt next to it, not possible. That changed now. You can long press the message, as you can see in this example from a teammate, adding a chain to his dog, and you can update the prompt and keep working with attachments. Really nice to have. Another brief one, again, on mobile apps is if you long press the send button, you get to pick the intelligence level. Look something like this. And it's actually really good to know and not very obvious. Back to desktop, there's two really interesting changes that I want to show you. Remember Canvas, kind of the word editor built into ChatGPT? Well, they've been phasing that out, and there's some weird things about it. If I switch the model to like GPT 5.3, for example, you will see the Canvas is still available here under more. If I change it back to GPT 5.5, I will go on the more and there's no canvas. So, what they're doing is they're trying to interactively insert it when you write something like an essay or some other long-form text, see it appears here automatically. And then if the text is really long, there's also a dynamic table of contents that appears on the left side of the screen. It's a beautiful table of contents implementation and this word editor in here is also kind of nice. As a reminder, if you full screen, there's this button that says add to library and then you could add your essay into the sidebar here under library. You have the penguin essay and then I could easily keep working with that or just save it for a later point in time. Whenever I create something that's worth saving. And then finally, ChatGPT. And this is a really fun one. The Gmail connector now has the ability to actually send emails. So, this would be under plus, more, and Gmail. If I enable that, send an email to info@apple.com, ask them when the new Mac Studios are coming out. I sent an inquiry about the upcoming Mac Studio release. Wonderful. When I look into my sent emails, hello. I hope you're well. I wanted to ask whether Apple has any information it can share regarding upcoming Mac Studio models and their expected release timing. Yeah, it did it. So, this is not the most useful example in the world, I realize that, but what you could do is a one-two punch. Let me show you something fun with the scheduled task that we talked about earlier. Go to scheduled and then you could do something like this where you set up a email scan. It was actually one of the recommendations where you can scan your recent emails for anything that needs your attention. In this particular scan, I set it up to focus on events. The frequency of the scan is once every hour and when it surfaces on email like this one, this is just my burner email address that I kind of use to sign up for random stuff. I'm going to event in Cannes next week. We're looking forward to that and it found an email with my badge barcode in there. Now, this is not something I would want to reply to cuz it's just informational, but if you wanted, you can now combo it with the Gmail connector and just say reply XYZ. And this is a way that ChatGPT can keep re-scanning your email inbox and you can just reply right in there. It's not fully automatic, but it makes your life a whole lot easier if you have a lot of emails and you can filter. One note of warning, connectors are notoriously unreliable, meaning they'll get the job done eight out of 10 times, but then sometimes things will slip through the cracks. So, don't think of this as your complete email operating system, more like it's a really simple way to get a lot of them done, but then still go into the inbox and look at if everything actually got pulled in and handled. But it's pretty neat, we're getting there. I'm looking forward to the day where automating your entire email inbox is going to be as simple as this. Right now, there's custom solutions and particular apps that help, but still take oversight. And hey, if you're enjoying this video, make sure to subscribe, it really helps out the channel. Let's look at the next thing. Let's look at a few more stories that happened over the last week that I want to talk about. One of them is the Fable story. I mean, last Friday we made the video on Fable with all the different use cases. I think it's amazing video showing you what it actually can do, what it did for me, what it did across the internet. A few hours after the video came out, Fable was taken down, right? US government said, "Hey, this this model is too powerful for anybody who is not a US citizen." And then Anthropic just turned it off cuz it was just a simpler thing to do. I think the big learning from that is that with Fable, there's a whole new tier of model, Fable or Mythos tier, and it's just a question of uh probably days or weeks until a competitor comes out with a level that powerful or Anthropic re-enables it. So, for now we don't have access, but it is a new level. And talking about new levels, there's one open source model this week that I should bring to your attention, GLM 5.2. Have you heard of this? It's not often that I bring up open source models on here, they really have to be exceptional and this one is. 1 million token window like some of the best models out there, and agentic coding performance close to Opus 4.8, above Opus 4.7. Honestly, depending on the reasoning level, it's almost the same. And agentic coding is one of the benchmarks that matters the most because that's a lot of what agents do. It often also translates to agentic reasoning and how well the system works. This is unbelievable. Basically, what this means is we have a open source model that you can run on your machine. Um yeah, if you have a 15, 20,000 dollar machine that is as good as Opus 4.8, but it's fully private, fully local, doesn't cost you any API cost. Amazing stuff and it's just available for free out there under MIT license, which is a fully open source license. I guess one important nuance is that it's available for the API now and the full open source open weight release is coming next week. No more API bills. It's kind of crazy if you do a lot and pay a lot for APIs these days. And last story and this is a really fun one is Oasis free by Decart AI. It's a world model and you can drive around in it. How about this? Let's take this foggy coastal bridge, generate a new world, and look at that, with WASD, you can kind of just drive. And they made this for self-driving cars I think initially, but as you can see, I can kind of just go left and then I'm inside of the bus, I guess. It's really interesting how well this is stitched together and the use cases for this are only going to become apparent over time outside of self-driving. The point of this originally was creating unusual driving scenarios, um which I think I'm doing a good job at right now, actually. And we're in the fields. One last note is that this does render in real time, so all of the stuff you see, that tunnel just got produced for me. All right, that's all the new interesting stuff in AI this week. I think it's really interesting how these scheduled tasks are sort of popping up everywhere. It's not just agents using them, but also consumer products getting them and it's definitely the direction that this is heading in. If you solve a problem once, you shouldn't have to re-solve it over and over again. That's kind of the point of infinite intelligence at your fingertips. So yeah, I'm going to keep a close eye on all these scheduled tasks, scheduling features, cron jobs, whatever you want going them because I think it's a big opportunity for individuals to get a lot of time back with AI. That's what we're all about here. All right, my name is Igor and I hope you have a wonderful day. >> [music] [music]
AI is starting to recreate history in ways that have never been possible before. You can now fly through the Trojan War, walk through ancient Rome, or place yourself inside a historical battlefield as if you were there. But this goes much further than just making cool AI videos. We can take ancient statues and create realistic interpretations of what historical figures may have actually looked like. We can remaster footage from the early 1900s into crisp 4K detail, making the past feel almost uncomfortably present. We can even take old recordings and translate them into new languages while preserving the original speaker's tone, rhythm, and delivery. And in some cases, AI is even helping researchers decipher [music] ancient texts that have remained unreadable for thousands of years. So, in today's video, I'm going to show you how AI is reshaping the way we see history, how creators are using tools to build entire historical worlds and how you can start creating your own AI historical artifacts as well. And later, I'll also show you how some channels are already making thousands of dollars a month from AI generated historical content. This isn't just educational. This is entrepreneurial. If you're new here, I'm AI Samson and welcome to the channel. Take a look at this. This is a cinematic AI interpretation of the Trojan War. What I find interesting here is not just that it looks impressive, cinematic, and real, but that it gives us a totally different way to relate to a historical or mythological event. Normally, history is something we read about. We see it in textbooks, paintings, documentaries, or [music] maybe even a few museum displays. But with AI, history becomes something much more immersive. You can create the feeling of flying through a battlefield. what it was like to storm the beaches of Normandy or walking through a lost city, an Aztec empire, or witnessing an ancient ritual from the perspective of someone who might have actually been there. Now, using the latest AI techniques, we can take real sources and apply them to make the most realistic interpretations possible. And this is what it makes it so powerful because a lot of how we understand history has always involved imagination. paintings, films, reconstructions, museum displays, historical dramas, even maps and diagrams are all attempts to help us picture something that no longer exists. The difference now is that AI allows almost anyone to create these interpretations with a level of speed, scale, and visual richness that used to require thousands of dollars and an entire film studio. Take [music] a look at this. There are these videos that have been going viral on Tik Tok which depict accurate [music] interpretations of what ancient Rome would have looked like in 10 AD. And you can see we get this wonderful aesthetic that showcases some of the realistic natures of this time period. Now, one Tik Tok account has amassed more than 500,000 followers creating simple point of view historical videos. Here is one from the day that Pompei >> [music] >> erupted. And it's certainly a wonderful and immersive way to learn more about history. Now, here is another one about waking up during the plague in 1351. And it certainly gives you a real appreciation for what life is like now. Might be wondering how historically accurate are these depictions. Now, some historians have weighed in calling these amateur and dangerous, but all historians that have spoken to the BBC agree that there are merits [music] to these type of videos, that they can be a gateway into history and can inspire people to do their own research. And my view is is that the more care and interest that you apply to this type of content, the [music] more realistic and reliable you can make them. And that's where things get really interesting because I came across this fascinating [music] YouTube channel on AI history. And they have more than 200,000 subscribers. And some of their most popular videos have gotten millions of views in just a few months. And these are entirely AI generated. The content, the audio, and likely the script was all made with AI. but they're giving us enough information [music] to be entertaining and worthwhile. Now, what's particularly interesting is how they take existing media like this shot of London Street and then they use AI to interpret it from a different perspective. So, they'll take a realistic drawing and then they will overlay a rendering of it from AI and then animate it to give us this real sense of immersiveness. Now they interperse this type of taking a real source either a map, a drawing, a painting or even a written account and interlace [music] it with AI video generations. And this is not just an interesting creative experiment. AI historical content is already becoming a serious content category. There are channels using AI to recreate ancient battles, lost civilizations, historical figures, mythology, and alternative versions of famous events. And some of them are getting millions of views and creating tens of thousands of dollars a month. But the larger point is clear. There is real demand for historical content. People love the past. They love ancient worlds. I love this saying, "A man thinks about the Roman Empire every 15 minutes." And we love the allure of kings, queens, empires, mysteries, and great conquests of the past. And AI gives creators a completely new way to package those stories. Instead of just telling people about history, you can now show them a version of it. Now, it's incredibly easy to do this, and I enjoy using Google's Nano Banana for this type of work. All you have to do is take a historical image. Here is an engraving from 1616 of London. And we can simply drag and drop this into the window and I ask for the following prompt. Photoreistic cinematic aerial view of London in 1616. Exact composition preserved. I then use AI to elaborate on some of the details you would expect to have in this exact scenario in 1616. We're merging both what we know about London in the 1600s and we're using this real source material from this engraving and that can give us an incredibly realistic depiction. And here we have it. We have this beautiful drawing. And now what we can even do is go ahead and animate that. You can simply pop it in the prompt box, change from image to video, and go ahead and ask it to animate. But that's not all because there's a whole host of exciting things we can also do with another technique. There is another YouTube channel that I found that has also amassed millions of views and more than 500,000 subscribers and they simply take old historical photographs and animate them realistically with AI. So here we have Abraham Lincoln and you can see here it's been colorized [music] and animated to give us a real sense of what he may have looked like in motion. Now he's also done this with Edgar Alan Poe. There he is stroking his hair quite seductively. Now, one thing you would have to be a little bit skeptical about is perhaps some of the behaviors. I think what makes this even better is that if you can take your understanding, your knowledge of their personalities and create specific [music] acting maneuvers that will embody that character. So, who knows if Edgar Alan Poe was so fond of his rather thinning hair, but this is one of my favorites. And this is Zar Nicholas II of Russia. And you can see it's been excellently colorized. And we certainly get a bit more of this sort of stern intensity of the man, apart from there when he breaks out into a big smile. And here's a lovely one of Albert Einstein. And this one works very well, I think, because what we have here is a lot of emotion already transmuted via the photograph. It's very clear what emotional state he is in. And it allows the AI to more accurately [music] animate that out. Now, what's great fun is not just looking at these ourselves, but actually going ahead and creating our own. because you can do this for a whole host of things. Specifically, if you have your own photographs from your family from this area, you can take those, recolorize them, and animate them yourself in a very simple way. And to do that, I will be using a tool from today's sponsor. And it's truly one of my favorite places to create any type of AI content, and that is Hicksfield. And it's a place where you get access to the latest image, video, and audio AI models all in one place. That means you just have one single subscription and you [music] can perform any of the tasks that we've talked about already in one browser tab. So for this we would go to image and I'm going to be using an image of the author France CFKA. And this is particularly interesting because he was alive in the early 1900s which is where we're starting to get some quite highquality black and white photographs but [music] we don't have any decent video footage of him. Poor man also died at just age 40. Often the best are taken too soon. So, we can simply come into the image tab, drag and drop that in, and use the following prompt. And I'm also giving you every single prompt that I have used in this video for free in the link in the description below. And that way, you can go ahead and test out these exact historical techniques yourself for free. And if you are new here, I do invite you to subscribe to the channel. But let's get back to making friends CFKA come alive. So, the prompt basically tells the AI to recolor this into a modern highresolution photograph. Preserve the subre's face, expression, pose, and composition. exactly as they appear and it specifically outlines that we do not want any type of [music] painting or illustration. Now for this I recommend using GPT I2 or Nano Banana. These are probably the best models for performing [music] these types of actions. You can also select the aspect ratio and go ahead and generate. I like to get four out at a time for this type of work. Now what I love about Hicksfield is it has a whole host of other useful tools including a plugin that you can put into Adobe Premiere if you're already editing. You can create videos directly from inside Adobe Premiere. It also has the ability to connect directly with Claude. And you can use the intelligence of an LLM to create your own media, which is an extremely exciting way to apply this. And here you have the output. And you can see it really adds some realism to Kfka. What an intense man he was. Now, of course, the next step is to go ahead and animate this. So, we can go directly into the animate tab, [music] and you can pop in a prompt. Now, this is where we can get really interesting and try to leverage everything that we know about the map and try and create something much more realistic. What we can do is we can ask AI to help us leverage everything that we understand about CFKA to write an intelligent video prompt to give us more realistic acting and behaviors of the man. So, I've asked Claude to use everything we know about him to write a short video prompt that would describe his manner, his speech, his movements, and his demeanor. This will be for a mediumshot portrait about 5 seconds long. So I want to give you the full process here without editing anything out to make it look easier than it actually was. Now a couple of challenges that I came up with here is first the video prompts that Claude wrote for me did not take into account that we're using an image as the first frame. So it started to describe things that would not fit. For example, it was talking about his hands when actually his hands are out of shot. So we give it the context. We give it the first frame and then ask for more details. Specifically, we want to have a line of Kfka-esque dialogue in the way that you might imagine him speaking. And then it's gone through and given him his accent in Prague German. Precise consonants, slightly formal cadence. The vowels carrying a faint bohemian softness beneath the clicked central European diction. Beautiful. So we can take that entire prompt and use it to create our video. Pop it in the prompt box. I recommend using sea dance too fast for this. And generate and you get out this. >> I have been informed that I exist. I'm still reviewing the documentation. Now, that's not all we can do because we can also take other types of sources and use them as inspiration. And one fantastic source is statues. And here's an example of taking Marcus Aurelius and creating a visual representation of how he might have looked. This is from vshart.com. Now, you can have a look at Daniel Vosart at voschart.com for some of his interesting projects. Another one that I like is his Egyptian portrait renderings. [music] And you can see we get some beautiful interpretations from these portraits. Now, if you'd like to try Hicksfield yourself, you can check it out using the link in the description below. And a big thanks to Hicksfield for sponsoring this segment of the video. But I tell you, there's a lot more that we can do. And one of the most interesting areas where we're developing a huge amount of capability is audio and voice. And we've looked at this how we can perhaps imagine how CFKA might sound, but we can do something that I think is quite remarkable. and that is to take a recording of an individual from one language [music] and then use AI to not only translate that into another language but to preserve the exact delivery, [music] personality, intonation and cadence from the original. Now, this is incredibly interesting for historical understanding because it means that we can take significant speeches from the past and have them translated and transformed into English to give us a better understanding, a better feel for the personality of different people. Let me show you what I mean. Now, before I show you any of these examples, I want to preface this by saying that some of these come from dictators who performed some of the greatest atrocities in the history of humanity. And in no way am I promoting or condoning anything that is coming up in the next section. And that if you are sensitive to such pieces of history, I'm giving you fair warning that this may not be appropriate for you to watch. But I want to make the argument that it is important that we do not forget history so that it cannot repeat itself. And that is the reasoning as to why I want to demonstrate these capabilities. Now, I will leave links to some other ones in the description below in case you want to look at those. But the only one I'm going to show you is from Herithro that I have created myself. And this is the Hirohito surrender broadcast, also known as the jewel voice broadcast. Is one of the most significant recordings of the 20th century. On the 15th of August 1945, Emperor Hirohito addressed the Japanese people by radio to announce Japan's surrender in World War II. This was following the atomic bombings of Hiroshima and Nagasaki and the Soviet Union's entry into the war against Japan. What makes it especially powerful is that for many Japanese people, this was the first time they had ever heard the emperor's voice. Until then, Hirohito had been treated as a distant, almost sacred imperial figure. So, the moment was not just political, it was spiritual, cultural, and psychological. A divine-like ruler was speaking directly to the nation to announce defeat. And here it is. >> In my numerous journals, I have observed the superstous gazes. Truly, my original intent was never to seek wealth, let alone to kill part. >> Now, as you can see, it's quite compelling to be able to feel how he delivered this. There is a sense of being able to connect with the character, the emotion, and the significance of this moment that is only revealed with the use of this AI technology. Now, as you can see, the speech itself was not simple or direct. It was delivered in highly formal classical language. So many ordinary listeners reportedly struggled to fully understand it in the moment. He also did not plainly say Japan has surrendered. Instead, he used more indirect language saying Japan must endure the unendurable. Now, you can simply do this with a tool like 11 Labs. And you can even do it with your own voice as well, which is quite an interesting way. So, here's me speaking in Japanese. Now, here's another interesting application of this voice training technology because here we have the speech that JFK was supposed to give on the day he was assassinated. And what they've done is they've created an accurate voice clone using his existing recordings to create that speech that was never given. And we can listen to a little bit of that now. We in this country, in this generation of our destiny rather than choice the watchmen on the walls of world freedom. Now what's also interesting is how we can take some other footage and use some of the latest restoration techniques to give them a [music] high quality modern feel. Now one example of this is taking the 1966 World Cup which was won by England and recolorizing it. And you can see that now we take this old black and white footage and turn it into full color. But there's some other interesting ways where we can do this. And one of the most interesting I think is taking some beautiful classic films from the early [music] 20th century that was shot in 4x3 and using AI to extend them into 16x9. Now of course we can also incorporate other techniques like recolorization to this. But here you can see some of the most beautiful films like Charlie Chapen recreated in full 16x9. And this is really giving a new lifeblood to these great classical movies. Now where I think things get even more remarkable is the way that we can use AI to decipher ancient scrolls and scripts that have remained unreadable for thousands of years. And one of these is a Herculanium scroll that's almost 2,000 years old. And using AI, a team at Oxford University have been able to look inside these scrolls, pull out the text and translate it into modern English. And there are many cases of this happen. There is this ancient scroll that was owned by Julius Caesar's father-in-law. And this has been read using AI as well. And it translates to reveal some of the pleasures in life referencing music and food. In one passage, Fodamius questions whether things in lesser quantities bring more pleasure. As to in the case of food, we do not right away believe things that are scarce to be absolutely more pleasant than those which are abundant. But now I wish to move on to some more dramatic and concerning developments. And this is the ability to bring characters from the past back to life using AI avatars. Now what we're doing here is using all the techniques we've seen, but perhaps also including everything that they wrote themselves. So for some individuals perhaps like Shakespeare, there is a huge amount of written work attributed to them [music] and using that you can create an LLM that can inform how these characters may well have conversed and then you have the opportunity to even talk to them. Now there is an app that allows you to do exactly this. It's called Hello History and you can talk to all manner of different characters. This one looks a bit like Borat, but you have the Queen, you have Cleopatra, Winston Churchill, Gandandy. But it also raises questions about what happens to a person's likeness after they dies. There are a number of viral videos showing Martin Luther King performing outrageous acts like stealing from shops. And based on this, different platforms have banned the likeness of some individuals being used. Now, where things are starting to get quite concerning is with the invention of open world models. Now, this is where an AI can generate a simulation of a certain context. So, we have the possibility of generating realistic simulations of specific ancient periods. We could also then include characters based on what we know about them that you can meet and interact with. It's basically like creating your own environment that you can go in and explore using AI. And one of the best models is Google Genie 3, which you can actually access and use yourself today. Now, what I find fascinating about all of this is that AI is not just changing how we create images or videos. It's changing how we imagine the past. It allows us to build environments that give us a sense of what ancient cities, lost worlds, and historical moments may have felt like. It lets us restore old footage, reinterpret ancient faces, and translate voices across languages, and even help researchers recover texts that were once thought to be unreadable forever. And yes, there is a danger here because the more convincing these images and videos become, the easier it is to confuse interpretation with truth, the easier it is to manipulate [music] history. That has an important ramification. A cinematic AI reconstruction of history is not the same thing as evidence. But used well, I think this technology can become an incredible tool for curiosity, education, and it can make history feel alive again. It can help people engage with the past, not as something dead and distant, but as something human, vivid, strange, emotional, and connected to who we are now. Because ultimately, [music] history is not just about old buildings, ancient wars, or people who lived thousands of years ago. It's about us, about you and me. Is about where we came from, what we believed, what shaped who we are now, and what we destroyed. And that is why it matters. [music] AI is giving us new ways to learn, to imagine, to teach, and to explore. And we are still at the very beginning of understanding what creative AI can actually do. Now, if you enjoyed this video, you can watch this one next, where I look at how AI is changing design forever. Because many of the same capabilities we've talked about here, reconstruction, reinterpretation, visual storytelling, and creative control, are also transforming the way that we create brands, products, images, and entire visual worlds. As always, thanks for watching and have a delightful
Today we're launching Claude Fable 5, the most capable model we've ever released to the public. Fable 5 is a Mythos class model with safeguards that make it ready for general use. We didn't broadly release our previous model with this level of capability because when we finished training and testing it, [music] we saw that the model, Claude Mythos preview, was finding thousands of cybersecurity vulnerabilities. A model that can find flaws like that can also [music] be used to exploit them. So instead of releasing it, we handed it to the people who protect [music] the world's critical software and put it to work fixing the holes before someone could break through them. It was the right call for the moment, but it was never the goal. We believe powerful AI should be safe and [music] accessible. That's why we went to work on Claude Fable 5. >> Every Claude model has safeguards to keep it from doing harm. Fable needed more cautious ones than anything we'd built before. Our safety systems for Fable 5 automatically [music] review requests that touch on high-risk areas like cybersecurity or biology. Those requests are then redirected to Opus 4.8. [music] We do that intentionally so people can continue to benefit from the capabilities of a powerful model like Fable without the cyber and biology risks that come with it. The safeguards are broad today, but we'll keep refining them so that they're better at allowing safe requests. >> We built Claude Fable 5 for your most ambitious [music] work. It can stay with a problem far longer than any model before it. It's highly autonomous and can operate for days without intervention. And it's not just coding. It can take on projects in finance, research, economics, law, complicated tasks that used [music] to need constant supervision. So point it at something that matters. What's the problem we'll look back on and wonder why it took so long to solve. We know what Claude Fable 5 can do. The interesting part is what you'll do with it. >> [music]
Today, I built three real multiplayer games in a single afternoon with zero lines of code. I published them, went to sleep, and woke up to nearly 4,000 players and 120 remixes. A week ago, this was impossible. Today, I'm going to show you the exact AI combo used to do it, and how you can ship and monetize your own games this weekend. Hi, I'm Adil. If you've seen the games built by Clot so far, you know they genuinely work, but they also look like this. capsule characters, gray boxes, one texture for the whole world. The code is brilliant, but the visuals are very basic. Today, I'm showing you the fix. Three complete games, every character with a real skin, every object with real textures, like an actual design team built them. All of it with Claw Fable 5 plus the Hixel MCP, the combo that does what nothing else could do before. One more thing before we dive in. I showed all three of these games to one of the biggest game studios on the planet, the people behind a shooter with over 670 million players, and filmed their honest reaction. It's later in this video. I had no clue how it would go. And stick around to the end. I'll show you exactly what all three games cost me to build and how the same workflow turns one sentence prompt into something you can actually make money from. Let's get into it. Let's start by preparing our setup, which takes only 30 seconds and it's just two steps. First, the Kixold MCP. Paste the connector into Claude. Sign in. And now Claude can generate real game assets instead of gray boxes. Second, the skill we built ourselves. You'll see what it does in a second. The MCP link, the skill, and every prompt I'll be using are all in the description. We start with a big one, a first person pirate game. You crew a gallion, steal it, man the cannons, and when you pull alongside an enemy, you board it and fight on the deck. So, I'm going to type build a first person pirate game where I sail a gallion, fire cannons at enemy ships, and board them for a sword fight on the deck. And just look at this. Every texture in this game, every sound was generated. The model made all of it from scratch. I mean, real wood on the whole, detailed cannons, characters that look like real pirates and Navy. This looks like a team of artists worked on it. And that's the power of the MCP, generating custom characters, detailed environments, and smooth animations directly inside the chat. Now, before this video, I ran an experiment. I took this exact one-s sentence prompt and give it to Clude alone. No MCP, no skill, and here's what came out. Same sentence, same model, same working game underneath, but what's missing makes a huge difference. There are no skins, no textures, no real ocean, just gray shapes. That's the line between a tech demo and something you'd actually want to play. Fable writes the game, and Hicksfield makes it look real. And the look isn't even the biggest gap. Because here's the truth about every game demo built by Claude you've seen this week. You can vibe code a game, but getting it online so your friend can actually join you is a completely different problem. The old way that means hiring a back-end developer at $50 an hour, weeks of work, syncing players, and rent service every month forever. That is the exact bottleneck where most people just give up. But here, my friend clicked a link. Higsfield hosted the match and synced both of us automatically. One sentence with literally zero effort. Now, you just watched a friend join a game that didn't exist this morning. You could be in a lobby with yours an hour from now. Links below. Now, let's check out the next game. A block world shooter. You can build and break like that mining and crafting game, but with guns. for the pirate game. You just saw the final result, but this time I want to pull back the curtain and show you exactly what our custom skill is doing under the hood because it's wild. Watch this. I'm typing build a block world shooter with two teams against each other where I can place and destroy blocks and fight the opposite team. And look, it doesn't just blindly start coding. The skill actually starts interviewing me the same way a real game studio interviews a client. Now it took my answers and wrote a full design document. Mechanics, art direction, level structure, sound. This is the prompt that built our code. Not my one sentence, but a studio grade brief. That's what the skill is. An experienced game dev team that works before a single line of code gets written. Then they put it online by itself. a live link, zero setup. Now, you can technically deploy a game with a plain AI chatbot, but you usually have to act as the middleman. The Kixel MCP just makes the infrastructure invisible. The same prompt that writes the game also hosts it. The simple prompt ends up with a playable link you can send anyone within minutes with no deployment headaches. Now, let's play the actual game. Look at that. The interface is clean. I mean, tells you everything you need. And right away you can feel that each gun has its own shape and personality. Let's take the sniper. So it fires slowly, but in return we get zoom in and it hits way harder than a regular gun. The bazooka though, now that goes another direction, launching rockets that also leave a real smoke trail behind them. And since it's a bazooka, it can blow up blocks, too, which makes sense because something that heavy should tear through brick and stone in a way regular bullets couldn't. All right, now let's check out the next game. Game three is the one that genuinely impressed me, and it's the one I want you to remember. It's that fruit slicing game, but there's no controller, no mouse, no screen to touch. Your webcam watches your hand and your fingertip is the blade. I'm not touching anything. The game is tracking my fingers and turning them into blades. Look at this. Isn't this so satisfying? Uh, we got three fingers. I can move each one separately. Lost one. And it has the full game logic. So, if I hit the bombs, I'm losing a finger. I'm losing hearts. All right. One hand is out. As a computer science graduate, this one genuinely blows my mind. Fable and Kicksfield took handtracking, the velocity prediction, the physics, and a fully playable game and stitched all of it together into a browser from a single prompt. Getting all those complicated layers to work together without breaking. That's what gets me every time. There's no installs, no extra hardware, just your hand. This is the one people won't believe until they try it themselves. It's fully live right now. if you want to test the camera tracking for yourself. So far, these games live at links I send people. The last step is a different thing. When they were done, the skill asked if I wanted to publish them to the marketplace, and I said yes. Now, they're not just links I share. They're listed where strangers find them on their own, play them, and can remix them. That's the difference. A deployed game is one you send. A published game is one people discover. Here's why the timing matters, though. The marketplace launched this week and it's almost empty. Publish now and you get discovered first, played first, remixed first. And every one of those plays is free data telling you what's actually fun. Early YouTube, early app store, same story. Back in 2008, solo developers made fortunes building simple flashlight apps just because they were first on the platform and there was zero competition. The key marketplace is in that exact phase right now. The advantage isn't just in making games. It's in being the only game in town when the players show up. The night I built these games, I published all three and went to sleep. I figured I'd come back, check the numbers, and show you whatever they were, good or bad. This is what I woke up to, though. Blockfield, the block shooter. Almost 4,000 people have played it overnight. I didn't post it anywhere. I didn't run a single ad. Didn't really tell a soul. They found it on the marketplace by themselves and started playing up to 22 of them in the same arena building, fighting, and talking in a game I described in one sentence just yesterday. And then there's this number. 121 people didn't just play it. They remixed it. They opened my game, changed it, and shipped their own version. Think about what that means. So 121 times, someone looked at what I made and thought, "I could do that." And one click later, they did. Okay, so here's the moment everyone's been waiting for. What this actually cost me, I asked Claw to pull the full credit breakdown. Everything we spend, including the fail generations, because I'm not going to hide those, it cost me $68 to build all three games from start to finish. And here's how the money actually works. The marketplace is free top of the funnel. You host for nothing. Real people play it. And if they actually keep coming back, that tells you you found something special. Then, because you own the code that Claude just wrote, you take the proven winner straight to digital distribution platforms, and that's where it scales. The marketplace shows you what to bet on. The big platforms are where the bet pays off. Look, I built these games, so obviously I think they're great. My opinion is not really worth much here. So, I sent all three to a studio that ships games for a living, Smilegate. You might not know the names in the West. The rest of the world does. Their shooter, Crossfire, has over 670 million players and has been running since 2007. Their CEO watched the exact same thing you just watched. No script, no notes from me. Here's the verdict. Prototype questioning. Minecraft. Okay. Okay. A studio at that scale. looking at three games I made yesterday from three sentences. Whatever you thought about AI games walking to this video, that's worth sitting with. Six months from now, everyone is doing this. The marketplace is crowded and games like these are everywhere. The window where this is an advantage is right now. The people who win on every new platform are the ones who show up early today. That can be you. All three games are published and the remix button is right there. Everything is in the description. the games to play, the MCP, the skill, and every prompt to build your own. And as always, if you found this video helpful, hit that like button, subscribe, and I'll see you guys in the next one.
In this video, I'm going to show you how to create this design step by step, but more importantly, I show you the exact details that make your work look professional. So, we're going to build this fantasy scene step by step with each other, how to correct the colors, how to correct the lighting, how to paint the fog, the atmosphere, and how to paint light precisely. And finally, how to put the final touches and color grade. So, without any further ado, let's dig into business. Hello everyone, welcome to my channel. My name is Nur and this is Nor Arts channel. All right guys, before we start, let's have a look at the final output. The main idea behind this piece was simple. I wanted to create like a dark cinematic fantasy mood design. Something that looks mysterious, epic, and realistic at the same time. So the focal point is the castle. I wanted to focus on the storytelling of the visual by guiding your eye using the rider with the road and the moonlight to drive the entire lighting direction. Because whenever you were building a scene like this, you are not just placing images together. You are actually directing light. Now it's time to build the design step by step with each other. You will find all the images that I'm going to use in the link in the description so that you can download them and follow along with me. I'll create a new project with these dimensions 4,000 widths and the height should be 2250. The coloring is 16 bit to give me flexibility in uh coloring and press create. I like to start with not white canvas because this irritates my eye and it's a good idea actually to start with a gray color in the background if you are making a dark scene. So, we're going to start with the castle. Let's bring it to Photoshop here. And I really like this perspective of the castle, you know, because this makes it like it looks huge and gigantic. Can you see the vertical lines are not straight because of the perspective is threepoint perspective. So, let's start by choosing the object selection tool. And from here, I'm going to choose select subject. Start with the cloud option and then let's press select subject. This gives me a really detailed selection depending on Adobe AI for selecting the object as you can see right now. Then create a mask. Very nice. And then we can simply use the hard rounded brush and paint over these areas because we're going to need them later. Very nice. I'm not caring about the lighting right now. I'm just caring about putting images together and making the main focal points and then we will care about the lighting and the color matching. So, let's bring the road and put it right here, I guess. Just Okay, I guess we need to make the castle smaller just right here. Very nice. The road should be right here. And then we can create a mask. And using the mask, I'm going to just paint over these areas. So, let's just mask this part. And we will use the grass brush to mask it properly later. But I'm just going now, right now, I'm just putting everything in position, trying to get the perspective and the composition right. So, I'm just press Alt and duplicate this layer like this. And let's use the grass brush. I'll leave all the brushes as well as a gift for you in the description so that you can follow along with me as well. The brush options. I'm going to remove the transfer and the color dynamics cuz I just want the brush only. I just want to mask this part using the brush. And here as well, maybe we can make it smaller here. Press X to toggle between the white and black to paint with the grass. right here. Let me show you the end result here. Can you see the this effect? This you can never get this effect without using the grass brushes or something similar. All right. So, let's mask this part as well like this which will make it look like it's part of the design or the ground, you know. So very gentle cuz this is the number one thing that differentiate between the professional artist and the beginner one. The selections should be perfectly done without any franges, without any white edges or something like this. Not bad. Let's select all the layers. Press Ctrl T. And we can now press Ctrl 2 to make it bigger like this. I'm not sure about the castle, but I guess this not bad. Maybe the castle needs to be shifted a little bit to the left. And maybe should be smaller. So, draw something like this. Maybe should be rotated a bit. Is good. And let's get to the road. Let's make it bigger. Now, I'm doing my best to get the composition right because this is the most tricky part in this visual. That's very nice. Uh I intentionally wanted to create this design using the least number of images possible because when you are just starting photo manipulation, you are having a lot of ambitious ideas but you get stuck because you are not skilled enough to execute these ideas. That's why I always recommend you to start with simple ideas with low number of images to blend together. This will make your progress gradual and you will be improved step by step because the last thing we want is you lose your passion. Uh the next step we're going to get the horse, the man with a horse and let's put him here. Press enter and we are going to select it using as well the select subject tool with the cloud option. This makes a really nice selection and actually it saves us a lot of time. So let's create a mask and right now press CtrlT. Let's make it smaller. So let's put him right here. That's very nice. Now I guess everything looks nice. If we used the composition rule, the rule of third, which is dividing your canvas into three uh by3 grid. So something like this. And here the main focal points should be in these four points. And this rule of third is really good in terms of showing your subject and at the same time showing its relation with the environment. It shows you the whole context. Let's bring the sky which will be the main source of light. Let's put it right here. Right where it belongs. Maybe we can make it smaller. Very nice. Now the composition is solid. The next step is going to be correcting the lightness and the color. So, let's start by being organized and we'll put put everything into a group. And let's start by the sky. I really like how the sky looks right now. It's just small tweak in the coloring. I'm going to use hue saturation. Create a climing mask and just going to tweak its color more towards cyananish color. That's it. Here's before and here is after. Maybe give it some saturation. Yeah, just a touch. Very nice. Let's correct the colors and the lightness of the castle. I'll start by creating curves adjustment layer. Create a clipping mask. And now we want the castle to be dark, but at the same time, we need to see its details. So the last thing I want you to make is to burn it like this. Because this is a mistake I see a lot of beginners doing. you know, we cannot see any details into the shadows. And that is not good. You need to make it dark, but at the same time, don't just burn it. So, let's make it dark like this. Maybe open up the shadows a little bit. And that's enough, I guess. And we can tweak it later. Next, uh, when creating any kind of night scenes, you should know that the elements will not have so much high saturation. So because we are having a lot of yellows right here in the castle. So we need to get rid of this. We'll do this using hue saturation. So let's create a hue saturation adjustment layer. And let's decrease the saturation a little bit. Not to the point where it it turns into complete gray, but just a touch of yellow will be enough. So something like this. Here's before and here's after. Very nice. Next, let's correct the coloring and do I'll just do this using color balance. You can use selective color, you can use even curves, whatever you want. But color balance the easiest one to use. So, I'm just going to add touch of cyan and a touch of blue. That's it. I guess we don't need more than this. Let's see. Maybe a touch of green. Go to the shadows. Add some blues, some cyans. Not so much. Maybe some green. Let's see. No, this is enough. And let's go to the highlights. Touch of sand and touch. Now let's see. Now we are talking. Look at this. Before, after. Before and after. Now it's blending uh in a better way. Blended seamlessly. So let's see. Here is before everything. Here is after. Quick pause. If you are interested in creating amazing visuals using photo manipulations like these, I have good news for you because my course, the ultimate guide to photo manipulation is now released and you can find it in the link in the description. In this course, you will learn how to create advertising visuals and photo manipulations like this using all the techniques that explaining all the time into a practical way like 13 projects fully narrated in English fully explained in details. Not only that, we will talk about an introduction to marketing because creating visuals is a little piece of the big puzzle which is the marketing game. So you need to understand the basics of marketing to create visuals that sells and then and how to apply the branding guideline into your visuals. And we talked about visual communication meaning and understanding the brief and the elements of the brief, the graphic design principles, hierarchy, contrast, proximity, balance, etc. And how to create a design from scratch. Uh brainstorming and making the ideas from scratch. gathering the images, the references and all these stuff. And finally, the practical part which is 13 fully explained applications into photo manipulation. And we have a section dedicated for using the basics of AI to help you get better results in faster times. And finally, how to create a presentation to show the effort that you put into your work. A link of the course is in the description. And let's get back to the video. Very nice. We're trying to do the same thing into the road, but I guess the ground and the castle are having the same base coloring. You know, they are kind of similar colors. So, we can just take the curves, the same curves, the same color balance layer, and press alt and drag them. Take a copy and put them here. And we can take another copy and put it into the other part. Let's see. That looks good. Maybe the curves needs some tweaks. So, let's make this darker. Maybe increase the blacks. This is a tricky part, by the way. Getting the lights right. It's not easy. I guess this is not bad. Maybe into the colors. We should increase the blues. See, it's too much. Maybe into the shadows. This is very tricky, guys. Need to be careful because we don't want to ruin the whole visual. Let's make it darker. Something like this will be good. And then we can simply duplicate the same two layers. Press alt and duplicate them and put them right here. Now the next step, we're going to correct also the coloring and the lightness of the horse. So let's do this using curves as well. Make it dark and same time open up the shadows a little bit. And let's create another color balance and do the same thing. Don't forget to create a clipping mask. Something like this. Go to the shadows. Add also a touch of blue and touch of cyan. Let's zoom out to see. Always zooming in and zooming out to look if everything is going well. Okay. I feel like here we have a lot of greens. We need to get rid of this green. So, we can do this using another hue saturation adjustment layer. Create a clipping mask and then decrease the saturation a little bit. So let's increase the saturation of the whole image. Yeah, this is better. Now we are talking. Okay, we can duplicate this layer and put it to the other one. And there we go. Very nice. Now the horse, the man with the horse needs a shadow, obviously. So let's create a shadow for this man. But before that, let's decrease the lightness of the man a little bit. Okay. To make a shadow, I will just use the selection of the man layer by pressing control and pressing into the mask from here. And then create a solid color adjustment layer which will take the same shape of the man with the horse. And then I'll take any color from the ground. So something maybe like this. Of course, a dark one. Something around this area. Yeah. And then press okay. and press Ctrl T. Try to put it into the ground. But at first, we need to get this layer behind the horse layer. Press Ctrl T and try to match its shape with the shape of the shadow that it should have. So maybe something like this, I guess. Press okay. The color was affected with the color balance adjustment layer. So it needs to be edited. So something I guess like this. And we can change the blending mode of the layer to multiply. Yeah. And then make it kind of brighter. This gives the shadow some color, you know. Very nice. So something like this looks cool. But we need to soften the edges a little bit. So let's go to the mask. Press into the mask from here. And from the properties panel, we will increase the feathering. So something like this looks better. And we can as well create a contact shadow which is the shadow that is resulting from two surfaces connected to each other. So I'm going to use the soft rounded brush. Make it smaller and using the same solid color adjustment layer. I'm just going to paint over the areas where the horse is touching the ground. Just subtle touches will be good. Maybe increase our brush and just give it some, you know, ambient shadow. Something like this. Very nice. Here is before and here is after. I'm feeling it looks good, but you know, we should remove it from some parts right here. See? Yeah, it's not bad. Now everything looks good in terms of color matching, saturation matching, and lightness matching. But the image looks very flat, not interesting at all. So the next step, we're going to paint the lights of the moon into each element. So let's start by the castle. And the main advantage we're having here is that the castle itself was having almost the same light distribution that we want. So to show you here, the castle was having the same lightness, the same shadows that we needed. We only just want to emphasize or increase the intensity of the lightness and remove this yellow tint from it. Okay. So we will use this for our sake by creating a new solid color adjustment layer and choose the light color will be something like this. Create a climming mask and I guess this is very high. So let's choose the color something like this. I'm just going to choose it from the clouds in the sky cuz you know it's really tricky to have the light intensity, right? Because the last thing we want to have is to have a very high lights like these which is lower than the main source of light which is the moon. Okay. So we need something like the nearest clouds to the moon. So something around this area will be okay. And then change its blending mode into screen. And next I'm going to use blend F to remove the light effect from the shadows. So double click into the layer and remove the effect from here. I'm just moving this slider to remove the light effect from the shadow areas. And we can smoon this transition by pressing Alt and pressing into this cursor and trying to make this really smooth. Can you see that? This looks really really nice. We can decrease it like this. Here we go. Having light effect the same as we wanted. Look at this. Here is before. Here is after. We can pump this even more. But for now, this looks okay for me. We can of course remove this light effect from certain areas. So, let's bring the uh soft rounded brush and remove the effect from this area. For example, using the mask and increase the brush. I'm just using uh going to remove it from here. Maybe from the edges from these edges here as well. Maybe soften it in these areas. So yeah, this is very nice. We will do the same thing into the ground part. So let's take the same solid color adjustment layer and duplicate it and put it into this area. And it immediately give us great result. But we will tweak the blend if values even more cuz we want different distribution. So something like this. Press all tweak it to something like this. We want kind of harsh highlights. So the moon is giving us harsh highlights. So we don't want it to be very soft. No. Something like this. That's okay. And let's use the mask to paint the lights in the areas that we want. So in these grass areas of course here looks really nice. Maybe we can soften it like this. Maybe remove it from these areas. This paper here. Yeah, immediately looks really nice. Let's see. Here is before. Here is after. Before. After. Really nice. We will leave this part because uh this should be even darker. So we can get to the curves. As I'm always saying guys, this is a neverending process. You will tweak everything until you finish the whole visual. You know, you will always uh change things here. For example, this area should be darker. So, let's create another curves adjustment layer. You will figure out things in the go, you know, create a clipping mask and Ctrl I. And let's just uh paint over this area, but we will change its blending mode into luminosity because we don't want to affect the coloring. We just want to affect on the brightness before. Really nice. Next, let's paint the highlights into the man with the horse. And we'll take the same layer, adjustment layer, and duplicate it by pressing alt and drag it. Drop it into the man. And we will do the same blend if thing we have done before. So like this. And then press alt. Try to distribute the light. That way it should nice. So press okay. and press Ctrl I. Using the soft rounded brush, I'm going to paint in these areas here. This should not have some highlights. This is it. Maybe some other highlights. Yeah, this is very high. Where is it? That looks really decent. Now, it's time to paint the absolute highlights. And for this, I'm going to use the graphic tablet and special brushes. So for this one, I'm going to use this brush I'm using all the time of Numeir. I like this brush because it gives me like some texture into the highlights. So let's create a clipping mask and let's choose its color to be something like this. Press okay. Change it to blending mode into screen. And we can see its color. This is very saturated and bright. So let's just take this color. Yeah. Press okay. Press Ctrl I to invert the mask. And let me show you what I mean by the texture. Can you see this texture? This makes the highlights realistic when painting them using this brush. Okay. So, just going to make the brush smaller. And let's paint the absolute highlights. Follow the form of the object and try to make it realistic. This can be improved by practice, nothing else. And if you don't know how to paint light and shadow, I have a full video explaining this. You'll find it into my channel. Just search for painting light nor art and you'll find it for the horse. Let's make some here and here as well. I guess this is enough. Maybe some subtle touches here as well. Just no. Okay, that looks nice. We can also paint some highlights into the grass layer by duplicating the same layer and put it right here. create a clipping mask and remove everything from this. Press Ctrl Y and using maybe the same brush. Let's see. No, don't look right. Okay, let's remove everything. And using this grass brush, I'm going to decrease the flow. And let's paint some grass highlights and change the angle of the brush using the arrow keys. Looks really nice. Let's paint some highlights here as well. Now we're talking. Let's see. Here is before. Here's after. Here's before. Here is after. It adds a lot to the scene. Now, let's paint some highlights into the castle. So, I guess you know the process. Let's bring the same layer and put it into the castle. And let's paint the highlights. Press Ctrl A and into the mask. I'm just going to press alt backspace to remove everything from it. And using the same texture, the brush we've used before into the horse. This one. So, let's double click into the layer and remove the highlights from the absolute darks. Yeah, something like this is really nice. And now we can paint freely without affecting on the absolute blacks. So, we can just give it brush stroke like this here as well. Here maybe here some touches. See this edge of course and here the areas that are facing the moonlight directly. These are the areas that should be affected the most area and here as well. Now we are talking. Let's see before after. I guess this is very intense. Let's decrease it opacity. That is nice. Now the image looks good but at the same time uh kind of flat. It needs another color to be cinematic to be artist. So we will create some fire lights here and here. And at the same time I guess we can have like another source of light should be out of outside of the canvas that is reflecting some orange subtle lights into this area. Okay. So let's make this creating another solid color adjustment layer. Let's give it some you know orangey color like this. Create a clipping mask of course. Change its blending mode into maybe linear dodge. or color dodge. This is good. But we need to change of course the light to something like this maybe. And then double click into the layer. Use the same blend if technique. Press alt to split the cursor. Now we're talking and that's I'm just focusing on to this area. Never mind about this part. And press Ctrl I. And using the polygonal lasso tool, I'm going to select this part only. Then paint on it using the mask. Of course, press into the mask. Press control backspace. See this is very high. We should also remove it from here. So let's remove it from here. I guess we need to paint this part manually. So, let's just put it over edges here and remove it this area. Just want the areas that are facing this. Okay. So, let's remove it from here. Maybe here as well. Increase the opacity of the brush. And let's decrease it from these areas. Just want a subtle orange touch. We can as well decrease the fill. No, I guess the opacity. So, this is nice. Maybe we can paint here as well. Yeah, this is not bad. Next, let's put the fire right here. And for this, I'm going to take this fire and open it into a new Photoshop project and use the same select subject. I'm always using this option recently because it gives me really really good selection and it saves a lot of time for me. It looks really nice. Create a selection. Create a mask. And boom, we are done. So, let's take this, I guess. Right click, convert it into a smart object first. And let's put it right here. Press Ctrl T. And let's make it very very small. Something like this. So, we will correct the color and the saturation. And for this, I'm going to use curves only. Create curves adjustment layer. Create a clipping mask. Let's make it dark like this. And open its shadows like this. We're going to we're going to use the curves to correct its color. So, let's give it some cyan. And at the same time, let's go to blues and give it some blue. And here we go. Do the shadows as well. And let's get back to the greens. Yeah, now we're done. So, let's see. Here's before and here is after. First looks blended, but the fire lost its brightness. So, I'm going to use the the same brush to mask the curves adjustment layer to reveal back the the fire brightness. So, like this. This is the first step. A second step, I'm going to paint some foggy lights around it. So, I'm going to use solid color adjustment layer with almost very desaturated bright yellowish color without any clipping mask and change its blending mode into screen. And for this, I'm going to use uh fog brush. Let's select this one. Yeah, let's paint some fog around this area. Increase the opacity again. And we can paint like, you know, because this we're going to paint some fog right here. So, this should be a foggy scene. A foggy scene will have always a foggy lights because what is fog? Fog is actually the particles into the air reflecting the lights, right? So, we should have these particles into the whole sea. I guess this color is not good. Maybe let's make it more orangey reddish one. So, yeah, this is better. We can duplicate this part. And let's just duplicate it and put it right here. Simple as this. Very nice. Now we should paint some highlights reflecting this lights into the color into the castle itself. So we'll do this by duplicating this solid color layer and remove everything from it. And then use any texture brush which will be this brush as well. And we should paint some reflected lights into the castle. So at first should paint some reflected lights here. Oops. This is very high. Let's decrease the flow of the brush here. Here. This is also very high. One touch here. Press shift and press into this area. And then let's erase it. It should be the highest intensity at the closest point to the source. Okay, now we're talking. And we should also paint some touches right here. It should be um you know very linear. Uh, I guess this is maybe we can paint also here and here. Yeah, I should do the same thing into this part, but it has um moonlight. So, we will try here as well. I guess this is not going to look good. Yeah, cuz the moonlight is bright. So, we should paint it over this edge. And that's it. Guess we need to get back to this because I really don't like these very sharp edges. So, we can get rid of them using the mask from the castle mask. So, from here, press select and mask. Now let's try to shift the edges. Yeah, like this. This is way better. Maybe f the edges a little bit. It's okay. Before that, this is better. We can also have like a lantern here. The man should be having it hanged with from the horse. So let's take it and open it into a new project. Do the same selection again using the cloud select subject and create a mask. Then right click convert it into a smart object. Let's take it and put it into our composite. Let's Ctrl T. Make it smaller. Oops. It is right here. Let's put it into the horse group because we are very organized, right? Ctrl T and let's make it really small here. Perfect. Now we can mask this area. Oops. Should be hanged. So something like this. We can now mask this area. It's really nice. Now, let's correct its color and light quickly using curves. So, I'm just going to create curves adjustment layer. Make it really dark. Okay. It's shadows a little bit. Go to the reds and add some cyans to the blues. Add some blues. Yeah, very nice. Then we will remove the effect of the curves from the fire like this. And then we will do the same thing we have done with the castle lights using reddish color. change its bling mode into screen. Press control I and then using maybe this fog brush again to paint some, you know, reflected lights like this. Paint some highlights into the horse using this a duplicate from the same layer. Create a clipping mask. Just going to remove everything. And let's use the same texture brush we have used in every part into the visual. I'm going to paint some highlights. freeze the flow of the brush and let's paint it like this. Very subtle. Now we are talking. Everything looks cool. And it's time to add some fog into the design. So let's create a solid color adjustment layer and let's hide it for a moment. And let's choose a dark desaturated blue. So something like this which we we will use into creating the fog effect. So this color I guess yeah it looks good. So how we're going to do the fog effect is simple. Just going to paint using I'm just going to paint it using the fog brush like this. But at the same time try to make it using some flow or rhythm. Shouldn't be like very uh random. No. So, the next part should be erasing. Be careful because we want it to have a good look. Let's see. Nice and mysterious. And if this looks um hard to you, you can use some fog overlays. Something maybe like this. Let's put here. Yeah, this is going to be easier. Change its blending mode into screen. And let's put it here. Press CtrlU to change its color. Press colorize. Give it some blueish color. Decrease its saturation. Decrease its brightness. Press okay. Let's hide this for a moment. Yeah, this should be really good as well. So, we can create a mask and remove it from some parts. So, just erase this areas and you're good to go. Now we are talking. I guess uh we're good right now. Let's just decrease the effect of the fog and create a new layer. And then press alt control shift E to bring everything into one layer. and right click convert it into a smart object. Now it's time to apply camera row filter. So let's go to filter and camera row filter. Here we're going to put the final coloring uh effects. So let's start by increasing the contrast which give us interest to the design and maybe increase the highlights as well. What about the shadows? Let's open up the shadows. and darken the absolute blacks. And for the temperature, I'm not sure. Maybe a touch of yellow. And a touch of magenta. The vibrance. Yeah, here's before. Here's after. I guess we should also give it some final dreamy effect using a new layer. Change its blending mode into screen. And using a really special brush called magic magic to I'm going to put some magic touches. So I'm just going to pick this color and give it some touches like this. And then fix this coloring with some touches. Just magic touches you know here as well. the moon. Give it one big touch and into the castle itself. But this time I'm going to change its shape to be something like this. Let's make it big. touches. Yeah. Let's decrease the opacity of this effect because it's really high. Maybe give this one also a touch. Increase it opacity. Yeah, here is the final result. Here is before and here is after. Yeah, it was a long one, guys, but I really enjoyed doing this one because it's kind of different from my style. My style was always very vibrant colors, very colorful visuals, but this one is kind of moody, cinematic, desaturated colors. You may have noticed that I'm kind of trying to limit myself to not use saturated colors. And yeah, guys, if you like this video and you want more tutorials about photo manipulation, you will find it as mentioned into my course in the link in the description. Yeah, see you soon in next tutorials. Peace.
Have you ever had one of those moments? You wake up with a perfect anime idea. [music] You can already see the character. You open your laptop, start creating, and the first image looks incredible, exactly how you imagined it. Then you generate the next scene, different face, different hairstyle, different outfit, and suddenly the character you imagined doesn't even look like the same person anymore. That's one of the biggest reasons why AI anime projects fall apart. Not because the images look bad, but because the character never stays consistent long enough to tell a story. Now, watch this. >> In a small village, a young man dreamed of becoming a knight, not through privilege or birthright, [music] but through courage and determination. He trained day after day, learning sword craft, enduring trials, pushing his body and spirit to their limits. Other knights tested him. Each battle made him [music] stronger. But the final trial awaited in the arena. A seasoned knight, bigger and more powerful than any opponent before. In that moment, as they faced each other, the young man understood this wasn't about defeating a man. And when the dust settled, the young man stood victorious. He had become a knight, not through the strength of his [music] body alone, but through the strength of his heart. >> Same character every scene, start to finish. In this tutorial, I will show you exactly how this was created. Now, there are plenty of tutorials that teach this well, but most of them use multiple tools, stitched together across multiple subscriptions, and that gets expensive fast. The question I I getting, is there a cheaper way to do this? An all-in-one solution? There is. It's called Abacus AI Studio. I've used Abacus multiple times on this channel because for $20 a month, you get access to multiple AI models, video tools, and workflows all in one place. So, instead of paying for several subscriptions, you can handle the entire workflow on a single platform. Same result, a fraction of the cost, and to make this easy to follow, comment the word "anime" and I will send over the skills and prompts I will be using in this video so you can follow along. All right, [music] let's get into it. Use the link in the description to sign up for Abacus AI. Once you sign in, it will bring you to this page. Here, you'll need to click on either image or video. When you choose one of them, you'll be taken to the Abacus AI Studio. This lets you use the auto feature where the large language model picks the best models for what you want to generate. For images, if you click on it and then the drop-down, you can pick whichever model you'd like to use. All the main models are there. If you switch over to video, you'll also see all the available video models. I should mention the Abacus Studio or platform also has text-to-speech, which we'll use as we create what we want. Since we want the large language model to use the skill I've built to guide us in creating a consistent anime, I'm going to set this to auto and then drag in the skill I prepared. To get this skill, just comment the word "anime" below and I'll send it to you. I'll drop that skill into the box here and tell it that I want to use the attached skill to create a story. Then, I'll click and send it off. Now, we wait. The skill will follow the step-by-step process I made. If you click on the skill, you'll see the details of what's inside it. You can also download and look at the breakdown of the skill. As you can see, it's listed all the steps to help us complete this anime. Confirming the plan, writing the script, generating the character, creating the voiceover, breaking down the segments, generating the video, putting everything together, and giving you the finished product. The first question it asks is, "Do you have any assets?" At this stage, there are two ways to use the skill I gave you. If you already have a script, image style, or voiceover audio, you can upload it here. Just click attach to upload it from your computer. This will save you a lot of steps. But, for many people watching, you might be new to this process. If you're new to this, just let it know you don't have any assets, and it'll help you generate them. In this tutorial, I'll follow plan B and say we don't have any assets, so it will create them for us. I'll say, "I don't have any assets. Please generate them for me." Then I submit. I'll answer all the questions, so it can make a character sheet for us, and I'll come back to show you what I've done so far. The second question it asked me is what story I want. I told it, "I want to create an adventure about a young boy dreaming of being a knight, practicing hard, facing challenges, and going through a final battle to become a knight." It then gave me some options and asked if I have my own voiceover audio, or if I want it to make one for me. I'll keep answering the questions. It's important here. It's asking about the voiceover. I chose for it to generate the voiceover itself. After that, it gives you a list and asks what kind of anime you want. >> [music] >> Anime, 3D, claymation, 2D, and so on. I said I want anime and will use the Ghibli style. It's important to pick the kind of animation you want. At this point, it has asked me a few questions. It asked about the character and the vibe I want. I've given a detailed description. One good thing about using Abacus is you can use the mic here just to share your thoughts about what you need. That gave me a detailed result. It also asked clarifying questions, which I answered. Now, it's asking what the target length for the video is. For this, I'll say about 50 seconds of video. Now, it asked me for the duration in minutes, so I picked about 1 minute. Then, it asked about the aspect ratio and I chose 16 by 9. Now, you can see it has a project plan ready. With these steps finished, it's asking if we should go to the next step or if we want to make changes. I'll just say, go, and it'll start generating the character sheet for us. While that's happening, on the side here, you can see where everything that's created will be saved. Also, here is the computer that's handling all this work automatically. I'll close this and then we wait. Now, it's given us a script. As you can see, it wrote a detailed script and showed the word count. One thing I added to the skill is that you can approve each step along the way. That's important because you shouldn't let AI handle everything start to finish. It's good to add human checks and input and make sure things follow YouTube's rules. For this tutorial, I'll say I'm happy with the narration and go ahead. It says the spoken words are 45 seconds. Leave 15 seconds for environmental sounds and so on. So, I'll approve and let it move on to the next stage. Now, it's ready to generate the character sheet and it'll use GPT image 2 for that. If you'd prefer to use a different model, like we've said, other models are available. Just a reminder, if you go to images and use the drop-down, you'll see all the model options. But, I want to use GPT image 2, too. So, I've asked it to do that. Now, it's going ahead to create the character sheet for us. You can see it generated the character sheet. It's impressive. This shows the boy before and after he became a knight with detailed descriptions, hand movement, facial expressions, and the prompt it used, which looks great. When you scroll down, here's the knight the boy will fight at the end to become a knight. Everything is detailed. Now, it's asking if you want to approve this. If not, you can say you want to change the color, hair, or clothes. You can do that, but I'll just say yes. One thing about Abacus is, as you saw, what we just did used 358 credits. Remember, all this fits within your $20 plan. I'll confirm so it will start building our storyboard. All right. The next thing it did was to start generating the voice-over. We can see that it has written out the full narration and asked for approval. I approved it and it immediately began generating the voice-over. I'll quickly play it so we can hear how it sounds before we continue. Let me play a short portion of this. Boy dreamed of becoming a knight, not through privilege or birthright, but through courage and determination. That sounds good. The next thing it has done is start breaking everything down into segments or scenes. We are doing this specifically to keep the character consistent across all the scenes in the video. That is extremely important. Otherwise, elements may start changing as the video progresses. All right, let's go ahead and approve this and I'll come back. Now, it's important to explain what the system is doing at this point. It says it will generate the images for all five segments for me to approve along with the video clips. For the sound, which is crucial to how the final result turns out, it's asking whether I want only the voiceover or voiceover and the generic sound. Things like wind, sword, and footsteps that give the video more life. Or I can choose voiceover, the sound, and the background music to bring everything together. For the purpose of this video, we're going to use voiceover and the enigmatic sound. You could add background music yourself, but because of copyright concerns, I'm not entirely sure. So, for now, we'll go with the voiceover and the enigmatic sound and let it continue. I decided to double-check what kind of music it would generate. That matters because you want to avoid using copyrighted music and getting flagged. It says the music is AI generated and has no copyright. So, this time, I'll ask it to use the voice, the generic sound, and the background music as well. It also asks which model I want to use. If you want to use C Dance 2.0, keep in mind that it will use a lot of credit from your $20. However, you can also use 13.0. 13.0 is also very good, but I'm going to use C Dance 2.0 for this example. So, I'll select it and choose C Dance 2.0. >> Now, all the images are being generated, but to show what we've done so far, it has locked in the music it's going to use to build the story. It also gives a clear progress update, so you can keep track, which is really helpful. At this point, it has started generating all the prompts. So, it's not just creating random images behind the scenes. It's not a black box. You can actually see all the prompts, and if you want, you can take them and use them for manual generation yourself. Based on this process though, you don't have to. It has generated all the image prompts for every segment, all five segments as you can see, and it will also go ahead and generate Yes, these are all the images for the five segments, and it starts generating them automatically. Sometimes it will ask you to generate them, but just click generate and let it handle everything. As you can see, this is the first dream image. This is the second, and it's now working on the third. It's also important to mention that, as you can see, it generated the storyboard, but the storyboard used Nano Banana 2. We don't want that. What we want is GPT Image 2. We can either wait for this to finish or stop it, so it doesn't waste your credit. We could use the same image, and Nano Banana 2 is not bad, but we want the same quality as the reference images used here. That is why you see the reason I put gates in these. So, it responded and said, "Yeah, you're absolutely right." It's going to use GPT Image 2 as a reference. It generated all the images again, and I'll show them once this is done. All right, it has now generated all the character sheets. If you look at this first set, it appears very consistent. As you can see, the character stays consistent through throughout each panel. Then, this is the next set. This one looks consistent as well. Here is the next character sheet. This is the fighting scene. And this is the final battle. I like this. I've gone ahead and asked it to generate the videos and put everything together for us so we can see the final result. We'll wait for that to finish. While the video is generating, you can see that all the images used for this 1-minute video took about 1,039 credits, just so you're aware. And like I said, whatever you generate in Abacus AI using the Abacus AI Studio, it will show you exactly how much credit you are spending. >> [music] >> It has now generated the first video so we can play it and compare it with the storyboard here. That way you can see how consistently everything flows through. >> In a quiet village a prince was born, a dream of a distant castle and a knight's true path and practiced with all his might. The road was long, but his heart was ready. It was his path. >> This came out really well. However, it's using a different voice, so I'm going to ask it to make sure because obviously Ciders 2.0 can generate voice-overs. So, I asked it to generate this with no voice because we want to use our own voice-over for it. I'll let it go ahead and generate everything for us, and when I come back, I'll show the final video. All right, everything has finished generating, including all the video scenes as well as the It has stitched together the voice-over and the background music, and as you can see, we now have an MP4 ready to download. We can download it, play it, and see the final result. And that's it for this tutorial. The biggest takeaway here is that the problem was never generating one good image. AI can already do that. The real challenge is generating multiple scenes, 10, 20, or even an entire story without the character turning into someone else or looking completely different halfway through. That is exactly what this workflow solves. You create your reference image once or a character sheet, then build everything around it. Suddenly, your character feels like the same person from the first scene to the last. So, here's what I want you to do. Open Abacus AI, grab the skill that I'll be providing for you, and make one short scene today. Not the full series, not the full movie, just one scene. The moment you see your own character stay consistent across multiple shots, this entire workflow start starts us to make a lot more sense. And if you found this video helpful, let me know in the comments. And if you want the skills and the prompts I used in this video, comment the word anime and I'll send them over to you. If you're serious about building a faceless YouTube channel the right way with AI, this is the video you should watch next and I'll see you there.
You will not believe I made this with AI. >> I checked every screw twice. I painted the sail myself. Hund, blue, like the sky. Wait, Dad, my toolbox. And your toy sailboat has a spare rudder. The other boats are so big. Can we even keep up? We're just spinning in circles. Oh, no, Rap. The rudder snapped. We're spinning in circles. We're so close. Keep going. It's working. Paddle with me, Ann. Push. We're so close, Meg. Keep going. >> [cheering] >> We did it. Not the biggest boat training, but the bravest crew. >> The way every character stayed perfectly consistent, the mouth movements matching every single word, the cinematic lighting that shifts like a real production. No team, no studio, no animation skills, just me and AI. But here's what most people get wrong. Everyone can make AI animated clips now. A few seconds, a cool shot, and then it falls apart. What almost nobody can do is build a full cinematic animation where the characters look exactly the same in every scene, and the lip sync actually matches the dialogue. That's the wall. That's where 99% of creators stop. Today, I'm going to break that wall for you, step by step, from the very first image to the final exported film. No fluff, no skipped steps. Let's get into it. Here's the truth most beginners miss. Making one good clip is easy. Anyone can do it. But building a whole film where your character's face never changes from scene to scene, and their mouth moves perfectly with every line of dialogue. That takes a real process. So, let me walk you through mine slowly and completely, nothing skipped. We start with the script because every great animation begins with a plan. Open Claude and paste in your master prompt. Hit enter and in moments, you have your full animation blueprint. This one prompt generates the entire package, your locked character descriptions, every scene with its own image prompt, and every scene's video prompt with the dialogue already written inside it. That dialogue line is the secret to lip sync later, so never skip it. Now, open Google Flow because this is where everything happens, both images and video in one place. Before we touch any scene, we lock our characters, and this is the single most important step in the whole workflow. On the left sidebar, click characters, then click new character. Now, go to Claude and copy the first character prompt. Copy the whole expanded prompt, the full detailed one with face, hair, clothing, colors, and features, and paste it into the Flow prompt box. Then generate. And look at this. Just look at how good this character came out. The detail in the face, the texture of the hair, the colors, the lighting. It looks like it came straight out of a professional animation studio, and it came from a single prompt. This is your hero, locked and ready. So, save it. Now, do the exact same for your second character. Go back to Claude, copy the whole expanded prompt for that character, paste it into Flow, and generate. And again, look at that. A completely different character, just as detailed, just as polished, and perfectly matching the style of the first. These two now belong in the same world, so save it. Do this for every character in your story, and now your characters are stored inside Flow as reusable references. And Flow pulls them into every scene so they never change. This is how you defeat the number one problem in AI animation, character drift. Now, let's generate our scene images one at a time and I'll explain exactly what's happening in each one. Scene one is your establishing shot, the opening of your story. And before you type anything, attach your locked characters as references so Flow knows exactly who belongs in the frame. Copy your scene one image prompt from Claude and paste it in. This first scene sets the world, so it's usually a wide shot showing the environment. Generate it and watch your characters appear in a full detailed setting that introduces your story. This is your foundation, so take your time and regenerate until it feels right. Scene two brings us closer to the action. Copy the scene two prompt from Claude, keep your character references attached, and paste it in. This is usually a medium shot. Generate it and notice the most important thing. Your characters look exactly the same as scene one. Same faces, same clothes, same style. That consistency is the whole point and it's happening automatically because your characters are locked. Scene three is where you introduce a moment of emotion or discovery. Copy the scene three prompt. Paste. Generate. This might be a closer shot focusing on your character's expression as something important happens. Look at how the lighting and framing shift to match the mood while your characters stay perfectly consistent. Each scene is a new moment, but the same world. Copy the scene four prompt and paste it into the Flow prompt box, then generate. And look at scene four. This is where your story really starts to feel like a film. The framing, the lighting, the energy in the shot, and your characters still look exactly like they did in scene one. Not a single detail drifted. That is the power of locking your characters. Every new scene just works. Now, go to Claude again. Copy the scene five text-to-image prompt, paste it into the Flow prompt box, and hit generate. And look at scene five. This is your climax, your big emotional moment. And it shows. The most dramatic framing yet, the richest lighting, the most powerful shot in your whole story, and the characters are still perfectly consistent. The same faces, the same outfits, the same style. Five completely different scenes and zero drift. This is what separates a real animation from a pile of random clips. Now, I'd repeat this exact pattern for every remaining scene, but the process never changes. Copy the prompt, paste, generate, and move on. Once all your scene images are done, look through them side by side and regenerate anything that feels off, because fixing it now saves you hours later. Now, the exciting part, bringing every image to life with motion and lip sync. And here's the best thing, you don't need to upload anything because the animate option is built right into the images we already created. Let me walk you through five animations one by one. For scene one, hover over the image you already made, click the three dots in the corner, and choose animate. Now, go to Claude, copy the scene one image to video prompt, the one with the camera movement and the dialogue line inside it, and paste it straight into the flow prompt box. Click generate. This is Google VO 3.1 working right inside Flow, and in about 30 seconds, your still image comes alive. The environment moves, the camera glides, and your character speaks the line with the lip sync matching automatically. No uploading, no separate voice tool, it all happens right here. For scene two, click the three dots on your scene two image, choose animate, go back to Claude, copy the scene two video prompt with its own dialogue line, paste it into Flow, and generate. Watch closely, the character's mouth moves perfectly with the words, and it's the same character from scene one, now alive and speaking. VO 3.1 handles the lip sync for you inside the generation. >> I checked every screw twice. She's not the biggest, but she's the bravest. >> For scene three, three dots, animate, then copy the scene three video prompt from Claude and paste it in. Generate it. And watch scene three come to life. This is your emotional beat, and the gentle push in makes it land. Your character's expression carries real feeling now, and as they speak, the lip sync [music] matches every word perfectly. This is where your audience starts to actually care about your character. Not because of fancy motion, but because of a quiet, honest moment that feels real. For scene four, three dots on the image, animate, copy the scene four video prompt from Claude, paste, generate, and look at scene four come alive. The camera moves, the energy lifts, your character is speaking, and the lip sync lands right on every word. This is no longer a still image. It's a living, breathing moment in your film. The same character you locked at the very start, now moving, now talking, exactly as you imagined. For scene five, your climax, three dots, animate, copy the scene five video prompt from Claude, paste it into Flow, and generate. And look at scene five come alive. >> The other boats are so big. Can we even keep up? >> Don't worry. We got this. >> The motion, the camera move, your character speaking with the lip sync landing right on every word. Another scene fully animated and still perfectly consistent with all the ones before it. Five scenes in and your story keeps getting stronger, every clip alive. If the lip sync isn't perfect on the first try, just regenerate. It usually locks in within a try or two. Now, just do this for all your remaining scenes, and the process never changes. Select the image you already made, click animate, go to Claude, and copy that scene's video prompt. Paste it into Flow, and hit generate. And now your entire story is animated. Every character consistent, every line lip synced, all inside Flow. Once every clip is generated, download them one by one, number them clearly, scene one, scene two, and so on, and save them in one folder. Now we make all these separate clips into one complete film, and for that, we move to CapCut. Open CapCut and start a new project. Import every clip you downloaded, then drag them onto the timeline in story order. And because you numbered your files, this takes only seconds. This timeline is your film coming together for the first time. Next, add transitions so the story flows naturally between scenes. Go to the transitions menu, pick a clean cinematic style, and keep them short, around half a second, so the pacing stays tight and engaging. Long transitions kill momentum. Short ones keep viewers watching. Now we upgrade the sound because audio is what makes an animation feel alive. Open the YouTube audio library and browse the copyright-safe tracks. Filter by mood to match your story, something cinematic, emotional, or adventurous. Preview a few, download the one that fits. Then import it into CapCut and drop it under your clips. Lower its volume so it supports your dialogue instead of covering it. Then add a few quick whoosh or swipe sound effects on your transitions to make the cuts feel dynamic and professional. Now add captions because they boost retention dramatically, especially on mobile where many people watch without sound. In CapCut, go to text, choose auto captions, and let it transcribe your dialogue automatically. Pick a clean, readable font, give it a bold style so it pops on screen, and add a subtle pop-in animation to keep the energy moving. Finally, do one last pass through your whole timeline. Trim any awkward pauses, tighten the pacing where it drags, check that every transition lands smoothly, and make sure the music and dialogue are balanced. When everything feels right, export your video as an MP4 in the best aspect ratio for your platform. And that's it. You didn't just make a clip, you broke the wall, you made a complete animated film with consistent characters, perfect lip sync, music, captions, and clean editing using a workflow you can repeat for every story you ever want to tell. This is the difference between someone who makes random AI clips and someone who makes real animation. And now you're on the right side of that line. If this tutorial helped you, hit subscribe because I post new AI animation tutorials every single week, and I'll see you in the next one.
Hey, my name is Nolan Michaels. I didn't
know anything about coding, but I was
able to build my dream software just
from talking to Claude Code. Look at
this. It's a color grading app with a
cassette [music] futurism aesthetic. I
was able to build a feature where I can
take one picture and ask for my other
images to look more like it. This is
what they look like by default, and with
my feature, I'm able to anchor these
images to this one up here. That's
crazy. Off.
On. I can't express just how happy this
makes me. Are you seeing why Claude Code
is such a big deal? This stuff is really
intelligent now. It can do a lot. So,
grab a coffee, get cozy, and I'll
explain everything. This is going to be
your ultimate beginner's guide to Claude
Code. Well, what is it? It's a spin-off
of the regular Claude chatbot. Claude
Code is an AI that works directly on
your computer. It's not a chatbot that
hands you code to copy and paste. This
thing reads your files, it writes new
ones, and it edits them. Like, you're
going to point it at a folder on your
computer, you're going to talk to it
like a person, and then it's going to do
work inside of the folder that you
chose. That's sort of a leap that people
have to make. It's not going to give you
instructions, it does the job. And it
asks permission before touching
anything. I think that's probably one of
the most important parts. I know this
was super weird to me when I first tried
it. Like, am I supposed to be giving
this thing permission to work on my
computer? And I think the consensus is,
like, yes, it is probably okay. Sure,
there are risks, but there are quite a
few safeguards that you can build in.
I'll show you everything that I do to
keep my work safe. One thing I really
like about Claude and the way it works,
as opposed to something like Chat GPT,
is that with Claude, there's not any
monthly limit, per se. What you get with
Claude is a 5-hour limit and a weekly
limit. You can only do a certain amount
of work within 5 hours. Once that 5
hours is up, it refreshes. And then,
theoretically, if every 5 hours you did
a bunch of work, you might hit the
weekly cap. So, instead of taking down
the whole month, you get this 5-hour
window and the weekly window. And I
think it helps my workflow so much more
this way. And quickly, which plan is
right for you? I honestly think you
should start on the Pro plan, okay?
Nothing wrong with that. You'll get used
to working with the AI. You might run
into some limits, but those are only the
5-hour limits at first. So, after 5
hours, you can get back to work. It's
super affordable. It's a really great
way to get started. Once you start
hitting the limits and you start getting
frustrated by those limits, that's when
you'll want to upgrade. And let me tell
you, this $140 Canadian is the best
money I have ever spent. It will tell
you that you get five times the usage,
but you get way more than that. I
literally never came close to hitting
the limit when I upgraded. But, since
this is for beginners, don't worry about
Max for now. Stick with Pro. So, now
that you understand maybe what you're
getting into, how do you get it? Where
can you download Claude Code? It's
funny, I actually do most of my work on
my desktop computer, not my laptop. So,
I'm going to go through this fresh with
you now. You can follow along. We want
to download the desktop app. I think
that's the best way to do it. You can go
to claude.ai. That's where your regular
Claude chatbot is located. And from
here, you can even click on the code
menu. This will bring you to a screen
with quite a few options. All we want is
the Claude Code app. So, I'm going to
download for Mac OS. Again, it's going
to give you a few different options. I
would actually classify these as more of
the advanced section. I don't think you
really need to worry about these for
now. If you're a beginner and you're
just getting involved with Claude Code,
don't worry too much about these. These
are just different interfaces that you
can interact with Claude in. Until you
get used to the basic app, I don't think
you're going to need these. On my Mac,
I'm going to double-click. I'm going to
drag the Claude app into my
applications. I'm going to go into my
applications and find the Claude app.
It's going to ask me to sign in, so
let's just continue with Google. This is
what it looks like and we're already
seeing a pop-up. You're going to see a
lot of pop-ups when you start working
with Claude code. So, what does this one
say? The Git command requires the
command line developer tools. Would you
like to install the tools now? I'm going
to say most of the time, maybe 99% of
the time, it's okay to install what pops
up on your computer. Now, there are some
caveats to that and maybe you should
take some precautions, but there is a
bit of a leap of faith that you're going
to have to do. If something says it's
required,
maybe you'll have to do it. We're going
to get to how you can double-check these
things in the future. For now, this is
one of the first steps, so I'm going to
install. I have to agree to this
license. Somehow it says it's going to
take 10 hours. Okay, no, down to 4
hours. This should only take a couple of
minutes. All right, this is actually
taking a few minutes here, so let's let
that download. If you're following
along, do that download as well, and
then come back when you're done. How
about that? All right, looks like we're
done. So, this is the Claude code app.
At the top, you'll notice we're in the
code tab. If you click on the left side,
you have access to all of your chats.
This is connected to your Claude
account, and you can talk to the bot
normally here. You might want to do that
sometimes, for sure. But, let's go back
to Claude code. And the most important
thing that you'll see here are these
options at the bottom, your local
computer and which folder Claude is
going to be working in. We have to
create a home for the bot when it comes
to our computer. So, let's create a
folder, and I've been told to call this
starting playground. We'll go back to
Claude, and then in select folder, I'm
going to find that on my computer. So,
do we trust this workspace? Claude code
may read, write, or execute files in
this directory. Only proceed if you
trust this workspace. I think more
importantly, you don't want Claude
accessing files that it shouldn't have
access to, like personal information, or
your favorite pictures, or whatever you
want saved on your computer. Don't
direct Claude at those folders. Make a
new home for the bot. I made the
starting playground folder, and I'm
going to trust this workspace. Now, the
funny part is we can just start talking
to the bot. And
that's the end of the ultimate
beginner's guide. You're on your own
now. No, I'm kidding. However, kind of
not. Like, that's the main part of it. I
can't tell you what to start creating,
but let's go over maybe the first 10
minutes, okay? You don't need to
memorize a command, a prompt. You just
talk to it. You open the folder where
you want the project to live, and you
type something like, "Build me a webpage
with a button that changes color when I
click it." That's all you have to do.
Maybe that example is a little boring.
What can we do instead? I have an idea.
Watch this. Inside of my starting
playground, I'm going to create a new
folder. I'm going to call this images.
I'm going to fill this folder with a few
pictures of mine, and then inside of
Claude code, I'm going to say something
like this. First, I want to know that
you can click on this little dictation
button down here, and just talk to
Claude. I don't really like doing that,
so I'm just going to type out my prompt.
"Hey pal, can you look inside the
starting playground folder for another
folder called images?" Then the
challenge is for you to build me a quick
app that lets me scroll through the
images like they're album art for a
record collection. As I scroll through,
if I want to stop and click on one of
the images, that image should be
amplified in a spotlight. And then, I'm
going to show you a little advanced
trick that I use all the time. I end all
my prompts with these words.
Do you know what I mean? {question mark}
Before hitting enter, I can scroll down
here to the bottom, and I can see which
model I'm working with. Right now, it's
Opus 4.8. I would love to be showing you
Fable 5. Right now, 4.8 is the newest
model available, so we're going to make
sure that's clicked. And then, in the
bottom left, we have our permissions.
And there's like a scale here. So, at
the very top, we have ask for
permissions. Every time Claude wants to
do something in that folder, it's going
to ask if that's okay. I honestly
suggest you start with this for now. On
the other end of that spectrum is auto
mode. Now, if you click on that,
Claude's not going to ask you any
questions, it's just going to start
working for you. I do not recommend you
clicking on that right now. If you're a
beginner, get used to Claude asking you
for permission first. Now, I can hit
enter. Claude would like to access files
in your document folder. This is the
permission I was referring to. I'm going
to say, yes. Then, it's going to say,
hey, yeah, I know exactly what you mean,
like flipping through records in a
crate, and when one catches your eye,
you pull it out and it gets lit up in a
spotlight. I think that's the power of
this prompt right here, do you know what
I mean? It gets Claude to acknowledge
what you said first. And then, we have
more permissions down here. Allow Claude
to run list playground folder and images
contents. So, you're going to start
seeing a lot of code, and at this point,
you're not really going to know what
you're signing up for. This is that leap
of faith. If you're really not
comfortable for whatever reason based on
what you see here, you can deny, as long
as ask permissions is on. I'm going to
allow once right here. Look at that, it
says it found eight images. Let me build
you the record crate browser. I'll make
it a single self-contained HTML file
that you can just open in your browser.
And look at this, it's just starting to
work. It's not giving me code that I
then have to put in a program somewhere
to create this idea. I just had an idea,
I gave it a folder to work in, I gave it
some assets to use, and it's starting to
do its work. It's asking me if I want to
allow it to write an index. You could
always allow it. You could allow once if
you're still trying to get comfortable.
Look at that, built it. Let me open it
so you can see it live. And a crazy
update recently is that instead of me
having to go through the folders and
click on what it created, it can just
serve me the app itself. This opened in
my browser without me having to do
anything. It called it the create scroll
the collection click a cover to
spotlight it. I can drag, scroll, or use
the arrows. And look at that. Maybe it
got a spelling mistake. Maybe it got
that from this spelling mistake on image
or maybe it got it from the file name,
which would have been the prompt I used.
I don't think I spelled retro wrong in
the prompt, but either way this is a
glimpse at the intelligence that you're
working with. It got this spelling
mistake from somewhere, okay? The king
of the dolphins. Okay, so it is probably
getting it from the file name. Maybe I
did spell retro wrong.
But isn't this cool?
A cat in the hat.
Amazing. Up here it's asking me about if
I want Claude notifications. I don't
know. I'm probably going to say don't
allow because I don't want to be
annoyed, but you might want to allow it
so that when Claude is working and when
it is asking for permissions or when
it's done its work, you might want to
get notified if you're not paying
attention to that screen at all times.
Fine, you know what? Best practice, I'm
just going to allow the notification. If
it gets annoying, I'll figure out how to
turn it off after. Back in Claude code,
it even opened it in the app form. Now,
the images aren't showing up. It could
be a little bit of a bug, but as we saw
in my browser, it was working just fine.
So, I'm just going to close this at the
top here. It has a couple of notes for
me. It says titles are auto clean from
the file name, and it turns out I did
spell retro wrong. So,
how about that? You live and you learn.
Then it ends with, "Want me to tweak
anything?" Some easy ones like adding a
vinyl record that slides out behind the
spotlighted cover. That's a pretty cool
idea. We could play a soft crate flip
sound on scroll. We could show the real
full file name or make the crate auto
advance like a slideshow. Those are just
some prompts from the AI to me about how
we can progress with this app that we
just built. Pretty insane for our first
10 minutes, wouldn't you say? So, we
just built something. Let's take a step
back and let's cover some more basics,
ultra basics of what Claude code
actually works with. Like, what is an MD
file? What you need to know about MD
files, you're going to see a lot of
them. MD stands for markdown. That's
what it's called, a markdown file. But,
really it's just a text document. It's
going to have words, headings, bullet
lists, absolutely nothing technical.
It's literally a word document. But, an
MD file is how you and Claude leave
notes for each other. This is where you
can put your plans, your decisions, your
speculative ideas. Your project will
slowly grow a library of MD files. And
these files are the reason that chat
number 50 is way smarter than chat
number one. Like yes, we just started
talking to it and it took a pretty good
first guess at what I wanted. So, you
remember these things that it asked me
about right at the end? I'm going to
save this in my prompt. I don't know. I
like your ideas, but can you create an
MD file documenting all of the ideas so
that we don't lose them for later?
Then we can talk about what to do next.
Allow Claude to write an ideas.md.
Where's it going to write it? Inside of
my starting playground folder, ideas.md.
And it's going to say done, saved
everything right next to the app. And
look what it did, already built V1, so
we have a baseline, ideas to consider
next, and then open questions. Now, the
reason this is so powerful, why this
step is so important, is so that if you
open a new chat with Claude, it can see
that ideas.md and pick up right where
you left off. All of your information
gets documented, that's why chat number
50 is way better than chat number one.
We are building a bunch of markdown
files for the new chats to build off of.
You know what I just noticed? These
files are a little loose in my folder.
So, I'm going to go to Claude, and I'm
going to say, "Hey, can you clean up
that project we just started? Can you
make its own folder inside of the
starting playground folder? And can you
put whatever asset we just used or
created inside of that new folder?"
Boom. Says it's on it, all cleaned up,
everything now lives in one
self-contained folder. Starting
playground, the crate, and then
everything we just used. Let's go and
make sure that's the case. If I click on
the starting playground folder, boom, we
have the crate right here, and look at
that. And let me just show you what an
MD file actually is. First, we need a
way to open it. So, let's open it with
TextEdit. It's just got all your ideas
right here. So, every new chat is going
to read this and know exactly what
you're talking about. You never have to
explain yourself again, sort of. So, we
know what an MD file is now. But there's
this whole other concept to Claude code
that is really special, and they are
called skills. A skill is a recipe that
you teach once, then run with one
command. And this all runs on that idea
of an MD file, a text file. A skill is a
little folder of instructions that
Claude is going to follow every time. I
can show you that I have skills for
Midjourney prompting, lyric writing, a
bunch of stuff that I find myself doing
often. I take all of my knowledge and I
build a skill around it. The skill
creator will help you build those
skills.
>> Girls only want boyfriends who have
great skills.
>> Now, here's an example of a prompt that
you can use to create a skill, but
there's one thing you need before doing
this, and that is the skill creator
skill. Bear with me for a second, okay?
What you'll need to do to get access to
these skills. In the top left, you're
going to see a customize option. Click
on this. We're going to see connectors
and skills. I already have the skill
creator listed here, and it is just an
MD file explaining to Claude how to
create a skill for a user. It's a lot of
text, a lot of prompts written in that
file. We can also browse plugins right
here. I don't really think you need to
worry too much about these for now, as
long as you have the skill creator,
which you can find right here, that will
allow you to create any skills that you
might need just by talking to the bot.
Now, I can't really tell you what skills
you're going to want to be creating.
That's what you have to figure out as
you explore with the AI. Back to
philosophy, you need to remember that
you are the director, not the coder.
This mindset is going to make all the
difference. Your job isn't to write the
code. It's not even to really understand
what's happening. It's to see the final
results and to direct your workers from
there. You have to know what you want,
and you have to be able to judge what
you see. You're the director. Claude is
your crew. You look at what it builds
and you say, "No, I want it more like
this instead." You keep steering the
crew until it gets it right. Describe
what you see, tell it the good parts,
and tell it the bad parts. So, if you've
gotten to this point in the video,
congratulations, you're well on your way
to becoming a master in Claude code.
But, here is the biggest thing that you
need to understand. Everything that
you're ever going to want will happen
through iteration. Okay?
Do you know what that means? Let me show
you. So, this beautiful color grading
app took me 40 days to build. This did
not happen on day one, not even close.
It took me 40 days, but it is possible.
That's the crazy part. And look at this,
I got Claude to build me an iteration
timeline. This app was built through 163
iterations. 79 features were shipped,
four features got rolled back, and there
were 16 major redesigns. It starts with
research, then we get to the cork board
in phase two. Look how much back and
forth went into this app. It was not
built overnight. 40 days of iterations.
And I can actually show you here. This
is the first version of my app. It
didn't have a name, it didn't have that
many options, but it worked. This thing
proved that it could do my idea. I could
have an interface where I could change
the color on my images. And that was the
proof of concept. At around 25% of the
way, we get to something like this,
where I now have a cork board that I can
move my image around. If I click on one
of the sliders, we can make adjustments
here, but it's all very laggy. I can't
even increase the size of my image. Like
this is still very basic. And then at
around 50% of the way through, we're
starting to get somewhere. Hey, I can
resize my image. I don't know if you
noticed that, but I made it so that if
you adjust one of your settings while
you're on the cork board, it takes you
right to the full screen edit mode. I
thought that was really important. We
have this color story feature over here
that didn't really work quite yet, but
it's supposed to give you the colors
that it sees in the image. I thought
that was awesome. We have a color pilot
name, but this is still not a start
menu. We have some presets that we could
add. It's really starting to get
somewhere. And then again, like 160
iterations later, we have something that
looks like this. It's pretty crazy. I
can even change the themes, like maybe
you want a lighter theme. Maybe we want
the black and chrome future tech plus
theme. Take a look at the color story,
it's pretty accurate now. We have the
ability to save snapshots. So, if you
make some changes, but then want to try
something else, you don't have to lose
what you built. Like, let's remove all
the color. I can save the grade, and
then I can flip back and forth between
them easily. Iteration, okay? Are you
understanding? Everything you'll ever
want is on the other side of iterations.
So, yes, this crate idea is nothing
special right now, but after 160 back
and forth, I bet we could get this
looking pretty cool. And I know that
because I just went through the same
thing with my other app. It is possible.
So, I've stressed the importance of
iteration. Now, how do we iterate
properly and successfully? Let's start
here. I want you to build a memory file.
Let's start a new chat, and then I'm
going to direct the bot to open my new
folder, and then I'm going to see if
it's up-to-date based on the MD file
located in that folder. I'm going to
say, "Hey, can you tell me where we left
off in the crate project?" It's reading
the memory file. It says, "No memory
file exists for this project yet, so let
me read what's actually in the project
to reconstruct where things stand." It
has an ideas folder. At the end, it
asked me something very important. Do I
want to create a memory file? Yes,
absolutely. We have our ideas file, we
want a memory MD. Now, that idea file is
all about the app specifically, but the
memory file is more about how you
interact with the AI. So, if you don't
want sycophancy, if you want it to be
straight and forward with you, that's
the type of instruction you would put in
the memory file. This is all you'd have
to say. "Set up a memory file for me.
Whenever I correct you, or state a
preference, or we settle on a decision,
save it there. Document everything and
apply it in every future session without
me asking again. That's all you need to
do to keep your ideas and your memories
rolling forward. Make sure every new
chat is building off of the last one. I
already told you about my favorite
prompt, the six words. I showed it to
you earlier, but I want to stress it
again. I end most of all my prompts with
"Do you know what I mean?" It's that
simple. I just want the bot, before it
does anything, to acknowledge what I
said. And when it does that, you can
instantly tell if you're on the same
page. You'll know right away if it
understands what you were trying to say.
And using those words, "Do you know what
I mean?" can loosen you up to have fun.
Rather than trying to craft a perfect
prompt, I'm telling you that stuff isn't
that important. Just ramble with
enthusiasm, have fun, and then at the
end ask, "Do you know what I mean?" This
will get the bot to synthesize
everything you just said and
re-summarize it for you. And then you
can point out what it got right and if
it got anything wrong. This will save
you a bunch of headaches. You don't want
the bot going and creating something
that you didn't actually want. It's just
a waste of time and money. Another way
to make iteration more fun, this is my
favorite way to do it, I want you to ask
for the picker. That's what it's called.
All you need to say is something like
this. Every time you have a question for
me or you're proposing a solution, ask
me with the picker UI. This is the
clickable questionnaire. Finishing the
prompt with "Remember this" will make
the bot put that inside of your memory
file. And say, "Hey, I want to keep
building the app, but check this. Do you
know what I mean?" It says, "Yes, of
course. Let me save that as a standing
preference." That's what we want. Look
at that. Let me put it to work right
away. Here's what to tackle next on the
crate. So, instead of going back and
forth in a chat, it's going to load up
this questionnaire for me. This is my
favorite way to iterate. Because you're
going to be going through a lot of
changes, you might give it a full page
of notes, and jumping from point to
point can get kind of confusing. This is
where the questionnaire comes in handy.
Simple enough to start, what do I want
to build next? We can do the spotlight,
we can do a 3D tilt, auto load the image
folders, shuffle and search, or other.
You can type your answer here.
We can minimize it, bring it back up.
So, let's go with this one, arrow
through images without leaving spotlight
mode. I think that would come in handy.
And then it's going to start working on
it, and look, it's going to ask for
permissions again. Now, at the bottom
here, instead of ask permissions, I'm
going to go into auto mode. I do not
think you should do this, okay? This is
an advanced feature. Because now that
I'm in auto mode, it's no longer going
to ask for my permissions. Please only
select this when you're sure about the
goal that you've given it. Because now
it's just going to start working. And
it's going to do whatever it needs to to
get the job done. And then look, it
brought up another picker question.
Stops at the end, wraps around. I want
it wrapping around for sure. Now it's
updating the backlog in project memory,
editing the ideas file. I'm going to
actually click this option, I'm done for
now, or I could just close the picker to
deny it. I'm just going to ask it, can
it open it for me in my browser, please?
And then boom, here we go. So, we click
on it, we get the spotlight, and then
using the arrow keys, I can go through
my images. So cool. Iteration, folks,
that's how you do it. Now there's one
more thing that will really level up
your iteration development. Screenshots,
okay? This is so important. All you need
to do is screenshot your computer,
circle what you want changed, and send
that back to the bot. This workflow is
exactly how I built my app. Literally
all I do. If you're on a Mac, I
recommend getting this app. It's called
Shutter and it's an easy way to create
screenshots. You can find it at
shutter.cc.
We download this for the Mac. On
Windows, it's really easy to take
screenshots and edit them. It's command
shift two on a Mac. Once I have Shutter
installed, I'm going to select the home
screen of this Create app. It's going to
be brought up here. I'm going to select
a freehand drawing and I'm literally
just going to draw a circle right here.
I'm going to copy this. Back in Claude,
I'm going to hit control V or command V
to paste that screenshot. You can click
on it here to see what you're sending.
Because I don't really have any other
ideas right now, I just want to show you
that this works. I'm going to say, "Can
you put a signature of mine right
there?" by Nolan Michael's Future Tech
Pilot. Remember, I put this on auto
mode, so it's just going to do it for
me. I doubt we're going to see the
picker. Again, make sure you have this
on ask permissions. I really just think
that's going to be better for you off
the start. And then take a look at that,
by Nolan Michael's Future Tech Pilot. It
made a decision. It didn't put it in the
empty space here, it put it below the
title. That's fine. I could screenshot
it again, maybe put a little red line
through this and say, "No, I want it
over here instead." Either way, through
the use of screenshots is exactly how I
created this app, all right? This next
tip is more of a personal suggestion.
I'm not sure it will work for everyone.
It just sort of helps my mind be at
ease. And that is to tell the bot to
document everything. Absolutely
everything. I don't want to lose any
idea that I send to the bot. The literal
prompt that I use is, "Here are my
notes. Document them and talk to me."
And then I paste my notes. I want the
bot to go through. I want the bot to put
them all where they need to be so that
we can find them in the future and then
without doing anything I want it to come
back and talk to me first. Building off
of that another personal preference, I
want to make sure that the bot speaks
with me before coding. That's something
that would go in your memory file. It
goes in my memory file at least. I want
my questions to get answered before any
code gets written and I think this
becomes more important as your app and
your project gets bigger and bigger.
Because you'll find that like one wrong
move in the code can really hurt things.
You got to be careful the bigger you
build, the more careful you have to be.
And this is probably the key part of the
prompt that goes into your memory. If my
message contains a question, even one
that sounds rhetorical, answer it in the
chat first. Never start coding while a
question of mine is unanswered. Remember
this. But then I'm going to say in the
chat, "Hey, I need you to add something
to the memory file for me." And then I'm
going to post that idea about questions.
So it's creating a special reference
inside of my memory file. Theoretically,
the more of these we build, the stronger
and more advanced and more intelligent
these AIs are going to feel. They're
going to get to know you better.
This is how you do it. The next step in
iteration is pointing the AI at specific
tools. I didn't give it any special
instructions about how to create this
right here, but there are some better
tools more suited for different jobs. It
will take some effort on your part to
understand when you should be asking for
these things, but I have two specific
examples, okay? I want you to mention
this acronym GSAP. This is going to help
when you're trying to build something
with motion involved. The other name
that I want you to remember is three.js.
Those two tools in particular really
helped build out my color grading app.
You can run a prompt like this if you
want, but in all honesty, just direct
the AI to reach into its tool bag to use
the most appropriate tool for the
occasion. Otherwise, it's going to
default to the easiest way of getting
your idea to life, and maybe you don't
always want to go down that easiest
path. Another aspect of iteration that
is extremely important, I want you to
tell the bot to build non-destructively.
Always iterate and never destroy on what
was built. This might be a personal
preference again, but I like knowing
that I can always roll back to previous
versions extremely easily. And this is
best done through something called
GitHub, okay? A GitHub is a way to save
your code and always have snapshots that
you can go back to. That's exactly how I
was able to show you old versions of my
color grading app. The code is saved on
GitHub, and it's being broadcasted
through a website called Vercel. This is
one you absolutely have to know about.
Vercel is the coolest service I've ever
seen. You can literally create a website
for free. It's not going to have a
custom domain or anything. It will say
like the crate.app.vercel.
But the point is you can then give that
link to anyone, and they can go on and
see what you built. And to be completely
honest, if you get to this point where
you want to share your work online,
that's when you're going to ask the bot
about it. Ask the bot, how do we use
Vercel and GitHub to make this work
public? It will know what to do, okay?
All you need to know as a beginner is to
ask the questions. Another part of
iteration that I think you'll enjoy is
asking for a lab. I want you to ask the
AI to build a safe sandbox for testing
UI redesigns and other wild ideas
without touching the real app. And I can
show you an example. So, this is how my
app looked like by default. Pretty
basic. I mean, it's fine, you know? But
then I asked for a lab to test new
interfaces, and it came up with this.
So, now I can use the sliders, I can
click these toggles, I can rotate the
knobs, and I can see that this is
something that I wanted so that we could
then import this into my app without it
trying to build all of this in the app
from the start, which could get messy
for sure. Building on that whole idea, I
think you should be creating mock-ups in
other AIs like GPT, maybe even Gemini
from Google. Create the mock-up
somewhere else and then bring those to
Claude. Funny enough, that's exactly how
I created my cassette futurism theme. I
gave GPT my original canvas and I said I
want it to look a little more like blank
and I described the aesthetic. It made
me this mock-up, I brought this mock-up
to Claude, and I said, "Can we build a
lab so that we can build out each panel
individually?" And while you're building
UI, you might want to know about this
specific word. Signal. Like when you're
building an interface, ask the bot, "How
can we signal our intentions to the
user?" That word is going to unlock a
whole new way of thinking about design.
For instance, users keep missing this
thing. How can we signal it to them?
Maybe something subtle. And I can show
you an actual example. So, in Color
Pilot, when you make changes to an
image, we can click the show grade
button to turn those changes off or put
them back on. Now, for some reason you
forgot about this button and it was
accidentally left off and then you start
making some changes and you don't know
why those changes aren't appearing on
your image, you'll see a now subtle glow
around show grade to remember that, "Oh,
I need to click this and then I'll see
the changes." I asked the AI, "Hey, we
need a signal to remind people about
that button." And it built that little
glow for me. There are three more things
that I think will help you a a on your
journey as a beginner. Number one, while
you're building the app, have the AI
build in some dev tools. A menu that
will let you access certain parts of the
design that normal users wouldn't have
access to. And you can say this, build
me a dev menu for this for this app. A
way to test the unlocks, a C view of
what the code is doing, and something I
can record while I reproduce a bug for
you. You don't have to copy that exact
prompt, you can be more specific about
what your app needs. If there's
something in your app that you need to
test the limits of, ask for some dev
tools to do it. The next thing that you
might want to ask for is an HTML
checklist file. This sort of goes in
with the dev tool idea, sort of.
Basically, you want to ask the AI to
help you keep track of things. This is
what I prompted for. I needed to make
sure the app had good first impressions.
I needed to make sure the corkboard had
everything working, everything about
grading an image. I needed to go through
and make sure was all fit to spec. This
checklist helped me do that. Super
powerful HTML checklist file. And then
one last thing that really comes in
handy for me and sort of makes me feel
safer about making these big
adjustments, I think you should ask the
AI to make it confirm the obvious.
You're going to give it a bunch of
notes. Some of those notes might have
obvious answers.
I don't want you to rely on that
assumption, okay? And I say that for one
specific reason, because even though it
might seem obvious, you may have used
the wrong word in your notes. Again, let
me show you an example. So up here,
color palette. I refer to this as the
start menu, okay? Because that's where
all of our options live. But nowhere
here does it say start menu. So, if I'm
calling this the start menu every time,
and then all of a sudden I call it the
color palette menu, that AI better know
what I'm talking about. And instead of
assuming, I want it to confirm the
obvious. Like yes, color pilot menu is
probably the start menu, but maybe I
used a different word. Maybe I called it
something else. Even something that
might seem obvious, I want to make sure
the AI is going in the right spot. And
I'm bringing this up as an example
because as you're building an app,
you're going to have a lot of different
parts that don't have an associated
name. Like if we go to the crate, what
are these called? Maybe we call them
images, but are we calling this the
shelf where all the images lie? Like
there's a bunch of different names that
you could use for different parts of
your app. I just want you to remember to
make the AI confirm the obvious, okay?
This will save you a headache or two.
And yes, maybe you would have guessed by
now, but this whole presentation was
made by Claude. Not necessarily the
information inside, but like the slides
themselves. I created a skill to take my
images and my notes and turn them into a
presentation. Isn't this amazing? I'm
able to give it my raw notes, a goal.
It's able to harvest the notes with an
evaluation of how likely it is I wanted
that note included in the goal of the
video. It's able to create different jot
notes for me on the slides. These menus
are collapsible. The slides themselves
are rearrangeable. The slides themselves
are rerollable because I have different
layout options, I can just sort of
randomize them. You'll see my layouts up
here. And my favorite thing that I asked
for when I hover over one of the points,
it gets amplified to direct the
attention to the important part. And
when there's really something important,
I'm able to dive into that slide and get
a new slide with that hidden
information. Isn't that amazing? Now, I
still have to fine-tune it. This was a
test run. Maybe you noticed it looking a
little weird. Maybe you didn't notice at
all. I'm going to keep making it better
and at some point I will share the skill
with everybody for free. You'll be able
to make your own presentation slides
super easy. I'll be honest and say I
don't know when that will be, so make
sure you subscribe and follow the
channel if you don't want to miss out on
that. This was your ultimate beginners
guide to Claude code. I think I covered
everything that you'll need to get
started. I truly can't wait to see what
you create in this new future. I hope
you're doing well. Take care.
And I'll see you next time.
Peace.
Could not fetch transcript (it might be disabled or unavailable). Error: Could not retrieve a transcript for the video https://www.youtube.com/watch?v=9Ky5QTpVjAw! This is most likely caused by: Subtitles are disabled for this video If you are sure that the described cause is not responsible for this error and that a transcript should be retrievable, please create an issue at https://github.com/jdepoix/youtube-transcript-api/issues. Please add which version of youtube_transcript_api you are using and provide the information needed to replicate the error. Also make sure that there are no open issues which already describe your problem!
AI music is being tracked, regulated, and even banned on some platforms, while AI music generators keep growing in popularity. Suno AI just crossed 100 million users. Google launched its own AI music model. Spotify and Universal Music are working on a feature to allow users to make AI covers of real artist songs. And lawsuits against Suno keep piling up. And this summer, a major court ruling could help decide the future of AI music forever. So yeah, grab your coffee, because I'm about to break down everything happening in AI music right now, and what all of this actually means for you. Let me start with a big picture. AI music is not a toy anymore. Suno AI just crossed 100 million users. 100 million people using one AI music platform. To put that in perspective, that's more users than Spotify had in its first decade. Suno is also making serious money. They did $150 million in revenue in 2025. $25 million just in February of 2026 alone. They have 2 million paying subscribers. Great numbers. But here's the problem. Every new technology creates two groups of people, creators and exploiters. Creators are using AI music to experiment new genres, make songs, and express ideas they never could before. While exploiters just want one thing, easy money. The same people already generating fake streams before AI, now suddenly have access to a machine that could mass produce songs at ridiculous scale. Hundreds. Thousands. Endless uploads creating AI slop. Low effort tracks designed to game the algorithm instead of making actual music. So guess what happened? It obviously drew a lot of attention towards AI music. And someone had to put a stop to this, or at least regulate it, so that it's not flooding all the streaming platforms. Bandcamp, one of the platforms taking the hardest stance, basically said, "Nope, not here." And outright banned music that is fully or substantially AI generated, along with AI impersonation of artists. Meanwhile, Deezer, one of the big streaming platforms, says that 44% of all new music being uploaded to them is AI generated. That's nearly half of all new uploads. Deezer, instead of banning AI music, they started detecting it. They built systems to identify AI-generated tracks and limit spammy uploads. And then there's Spotify, arguably taking the most controversial approach. They're not banning AI music. Instead, they're exploring licensed AI agreements with major record labels. Because here's the undeniable truth. AI music isn't going away. The genie is already out of the bottle. Everyone wants a piece of AI music and all the money that it generates. On May 21st, 2026, Spotify and Universal Music announced a deal that lets Spotify Premium subscribers create AI-generated covers and remixes of participating artist songs. Meaning artists opt in, fans pay extra, and musicians earn revenue when people remix their songs. Will Taylor Swift, Billie Eilish, Kendrick Lamar, and other artists opt in for this? We don't know that yet. But here's where things get interesting. Because Spotify is trying to position this as the good version of AI music. An alternative to what they call AI slop. But in my opinion, this could also open a giant Pandora's box. Because Spotify reportedly described this system as something that could make one song become 10,000. Think about what that means for a second. Imagine one Taylor Swift song suddenly spawning thousands of fan-made remixes. Metal versions, Bollywood versions, sad piano versions, all generated instantly. This would create remix slop. Because the problem is this. What happens if Spotify starts getting flooded with AI remixes? And this leaves no room for upcoming artists who have no intentions of remixing existing songs. Or don't have deals with major record labels. AI songs are already so much better. What if AI-generated versions begin competing with the original songs? This could create a vicious cycle where artists feel forced to participate. Not because they want to, but because if they don't, AI content starts dominating the platform. Welcome to the weirdest timeline in music history. The same industry that spent years fighting piracy is now trying to monetize AI-generated remixes before somebody else does. And AI covers have been insanely popular on social media. With videos getting millions of views on YouTube and TikTok. So Spotify and Universal Music are definitely hopping on the trend. Suno AI has a covers feature that allows you to upload any song and convert it into any genre while keeping the same melody. Well, guess what? The exploiters started misusing this Suno feature to upload copyrighted songs to Suno to create those famous 1950s soul versions of popular songs. To stop this, Suno partnered with Audible Magic to use their content identification and rights management technology to block any user uploads that are not original content. Apparently, Exploders found ways to bypass that system, and were making AI covers anyway. So Suno pushed an update recently to add more filters, breaking their upload system, and now blocking many user uploads even if they are original, messing up the feature for genuine users. As of today, there is still no ETA on when this will be fixed. Suno AI is in chaos. With so many lawsuits happening right now, it gets confusing fast. So let's pause for a second and actually map out who is suing who, who settled, who is still fighting, and which lawsuits actually matter right now. The big three record labels — Universal Music Group, Warner Music Group, and Sony Music — sued both Suno and Yudio, claiming that these AI music generators were trained on copyrighted music without permission from artists or labels. That's the core battle. Now let's talk about settlements. First, Universal Music settled with Yudio in late 2025. And this changed everything. You can watch this detailed video on our channel to understand how it all started. Because after the settlement, Yudio basically became what people call a walled garden. Meaning, you can still create music, but you can't download or export songs from the platform. Instead, Universal and Yudio are working on a platform called Starstruck, which will be a mobile app used by everyday fans, rather than producers or professional artists. It will have four modes. Cover mode. Imagine a Lady Gaga song in the style of Billie Eilish. Reimagine mode. Same lyrics, completely different composition. Remix mode. Genre swaps and stylistic transformations of existing tracks. And Create mode, where users write their own lyrics, but pair them with an artist's approved AI voice under strict guardrails. But there's a catch. The generated song would be owned by the rights holder of the participating artist. In other words, the fan would not own what they create on Starstruck. They would be paying for the right to create and listen inside Yudio's controlled environment. Basically, it feels like, pay us to remix our catalog inside our sandbox so that we can make money for record labels. Let me know in the comments what you think about Starstruck. Then, Warner Music settled with Suno in November 2025. Big, multi-million dollar settlement. Licensing partnership. And Suno even acquired Songkick from Warner as part of the deal. I actually made an entire deep dive video breaking this down, so you can check that out after this video. Warner also settled with Yudio at the same time and signed a licensing agreement with them. Now, here's where things get serious, because not everybody settled. Sony Music is still actively suing both Suno and Yudio. Universal Music is still suing Suno. And these cases matter a lot. Because what happens next could literally define the future of AI Music. One of the biggest legal questions right now is this. Can AI companies legally train on copyrighted music under fair use? That ruling could decide what every future AI Music company is allowed or not allowed to do. If courts say training without permission is legal, AI Music companies explode. If courts say it isn't, the entire AI Music industry changes overnight. And here's where things get messy. Universal Music reportedly wants Suno to move toward a more controlled, permission-based system similar to what they did with Yudio. Suno apparently isn't thrilled about that idea. And if Suno agrees to lock down, they will instantly lose all users, making it a graveyard in just a few weeks. And to make it worse, there's a new lawsuit against Suno from Poseidon Wave Media. But the bigger story here is that this isn't just a battle between AI companies and record labels anymore. Independent musicians, collecting societies, and artist groups around the world are now filing their own cases. The fight isn't just labels versus AI anymore. The world's largest musicians union, the American Federation of Musicians, is now suing Universal and Warner, claiming they allowed Suno and Yudio to train on musicians' recordings without paying the musicians themselves. So what does all this mean to you if you're using Suno AI? And what should you do? Suno AI is running version 5.5 right now. Great model. You can watch the full breakdown on the model in this video right here to see everything that it can do. But because of the deal with Warner Music, Suno is training a model with their artist catalog. And once that model is trained and ready, Suno will eventually sunset all the older models to remove all the unlicensed training data in those models. In Suno's recent announcement of $400 million in series D funding at a $5.4 billion valuation, the company said that over the coming months, it will begin rolling out its first music model developed in partnership with the music industry. So if there's an older Suno model you absolutely love using right now, make the best use of it. Because those older models may not be around for much longer. And obviously, not everyone is using Suno AI to generate AI slop. Some are using Suno with their own lyrics, their own melodies, using it just like any other tool for music production. And if you're looking for a music distributor for publishing your Suno AI songs on Spotify, YouTube Music, TikTok, and all other platforms, I would suggest that you use DistroKid. They're the only major music distribution platform that's allowing AI music. They have added some new options to allow you to choose if the song was entirely AI generated, or was AI just a part of it, like vocals, melody, etc. Other major music distribution platforms like TuneCore are actively blocking Suno AI songs, citing they don't allow music created with unlicensed AI music generators. And finally, I would suggest that you don't put all your eggs in one basket, meaning don't depend entirely on Suno AI for your music production. There are many Suno AI alternatives out there that you can try right now. And you can watch this video next to see all the new features in other AI music generators and open source alternatives. Hope you like this video. Comment what you think of this chaos. And subscribe to the channel if you want to stay updated on everything that happens in the world of AI music.
Testing. Testing is uh stream is healthy. Okay, cool. All right, so I've never done this before, so uh I think it'll be cool. So basically, you got any questions about Suno AI, instead of leaving a comment, you can just call me at this number. This is not my actual number. This is a Google number, but um or Google voice number, whatever the hell um There we go. And I'll pin that. And let me just go up on the YouTubies. Make sure everything is good. Yo, what's good? Um other otherism authorism authorism. What do you know about AI farm? Farms AI farm. >> Okay. Yep. Audio is working good. All right. Uh what the [ __ ] is that? So uh good luck for the live call. Appreciate that, buddy. Okay, let's look up. Yeah, yeah, yeah, yeah, yeah, yeah. Yeah, I think this will be fun. Um, [snorts] yeah, I have the most ghetto setup right now, though. But what's good, El, whatever the [ __ ] your name is. And this guy's talking about Suno AI Farm. Best prompt generator. Oh, what the hell? What the hell is this? What the hell? Craft radio ready. Prompt made for American artist. Prompt studio. Write and generate song prompts. Melody to lyrics. What the [ __ ] is this? All right. Somebody just take my Sununo GPT and turn it into a website. Uh, I gotta get on the bow. How do I stop Sunno from adding live guitars to all of my beats? Love your channel and got one of your books and GPT. Cheers, man. I appreciate that. Otherism. Um, I don't know with your beats. Are they [ __ ] Are they beats that would have guitar in them? What What type of beats you working with? You never take me to Bangladesh. talk about ac cappella, please. Um, back in 1954, the first ac cappella was generated by Abraham Mson. Uh, Abraham Mson is not real. I just made him up. [clears throat] >> [snorts] >> All right, let's uh yeah, we'll go one at a time. So, let's talk about shot by TT asks, "How do I stop Sununo from adding live guitars to all of my beats? They are a trap soul, but even if they are synth or strings, Suno adds live guitar." Uh, my first instinct would be to just [ __ ] like um hold on, let's let's go here. To just do the exclude styles thing would probably be the easiest thing. And uh since yeah, we'll put that at instrumental. So, let's see. Trap soul. I honestly don't even know what the [ __ ] would go into trap soul to be real. Um [snorts] Okay. The perfect ac capella is on. Okay, let me uh let me see what we got here. No, no, no, no, no, no, no. Wrong button. Wrong button. Crunk. Uh what what's like an example prompt that you use for making the the trap soul beats? Do you have a prompt you could like put in the comments [clears throat] cuz I do have the 10,000 pseudo prompts ebook which is by the way 50% sale because I had surgery that took out part of my rib and uh I haven't been able to put out as much videos. So, you know, I need some money. So, [laughter] it's 50% off. Uh but here's a [ __ ] Well, everything is 50% off, but uh like here's a trap soul prompt. And let's see if it mentions uh we'll get rid of we'll get rid of all the vocal stuff and everything. All the vocal stuff. Don't need that. Oh, >> hello. >> You're on air with Chill Panic. Uh, you're not on speaker yet because I got to make sure you're not a crazy person. >> Oh, no. >> Hell yeah. All right. you have gained access to my ghettos uh version of a live [laughter] call. >> Hey, Google voice it does the job. So, >> yeah. Um, I just had like um a question and um something that I've noticed like since the introduction of voice before the introduction of voice, the best way to get vocals would be to mash up an ac cappella with whatever an instrumental or have it you have to put two songs obviously, but um because the voice just really doesn't seem to uh I mean it's hit or miss. Have you ever I guess my question is have you ever mashed up a ac capella with an instrumental to to like and actually liking that how it sounds versus >> voice? Because I get so many mixed results with voice. It's kind of driving me crazy. [snorts] >> Okay. So, you're you're talking about like trying to generate a like instrumental and ac cappella separately or like you want to like make >> No. So, let me clarify. Sorry, I thought it was all over the place there. But um so like basically what I would do is like I take my own vocals cuz I was trying to get my own voice, whatever. And I would upload an ac capella and then an instrumental that I liked and then I would just mash those to do a mashup instead of like a cover. >> Oh. Oh, okay. Okay. >> The voice sounds exactly like me whenever I do it that way. But when I use voice mixed results. [snorts] >> Okay. Okay. So you it sounded more like you whenever you did a mashup instead of doing voice. >> Sounds exactly like me every time. >> Damn. That's interesting actually. >> Mashing up that ocatoa with the instrumental. So it's not really mashing up because it's just taking aca not another song. It's just you know the voice >> the beat that you choose. [snorts] >> Okay. That's [ __ ] that's interesting. I've actually never tried doing it that way before. >> Yeah. I mean, I think you'd be surprised if you try that, but um and then I guess my question is any tips on using voice other than what you put in your video, like uploading 16 of the same acapella and all that other stuff. >> Uh >> they're still not there yet with their like where they're at with it. Yeah, I definitely think they're still not there yet completely with the voice because yeah, doing that whole custom model thing has really been the only thing that I've seen that helps it at least be a little more accurate. Um I have noticed like if my vocals are processed better then it'll work a little bit better but um >> yeah for sure like no effects like I don't like I've uploaded like um stems from an actual song that I uploaded to signal and then just >> actual like ac capellas from my own export on a doll and those results were way better like way cleaner but [snorts] >> I don't know I guess it's not there but >> the matchup trick, bro. Try it out sometime. You'll be you'll be impressed with how good it is. [snorts] >> Okay. [ __ ] yeah. I'm definitely going to try the mashup thing cuz uh that's they they keep me going. >> I appreciate it. I appreciate it. Yeah. >> Keep me good. I'll see you. >> All right. Thanks for calling. We got severe thunderstorm. Woo. Uh we got we got somebody else. Okay, let me let me go here. Yeah, have y'all ever tried doing the uh the Oh, what's the hell? I hope this isn't calling with my like actual [ __ ] number. I think this is with Google voice. I'm new to this. >> Hello. >> Hello. >> Hello. >> Hey. What's going on, man? >> What's up? You're live with [ __ ] Chill Panic. Woo. >> First off, love the content. Um, I'm a musician down here in Florida and I started using Suno as a way of kind of bridging the gap and help finishing songs so I can break them down and re-record them myself. My question is, when I type in prompts, I want a certain feel. Like if for example I want something that sounds very similar to like Drowned by Bring Me the Horizon and it's got like the Sith Pad keys. >> You can't really get that same style. >> And as far as like the going into the chorus, like even when I try to get really detailed with the metatags, it still kind of overlooks it. >> Mhm. Yeah, I definitely have noticed that [ __ ] I think I think a lot of problems with Sununo is uh unfortunately it's just not there yet all the way with like the having the complete control and all the songs that they've trained on. [snorts] You know, they probably haven't used much Bring Me the Horizon and stuff like that cuz um I guess for metal core that's actually pretty popular. Maybe they have. Um but but you said you like re-record it yourself. >> Yeah. That um like that. Um I'm a DJ here at a club here in uh Florida and one of my songs is so good. It's like like that. I've got 700 people dancing to it. So I contacted real studio musicians and said I want to re-record this and put my vocals on it. >> Oh, okay. Okay. Um Okay. But you don't like play guitar or anything? You want to like get like uh you like get other musicians to do the guitar and drums and all of that >> and you just do like the voice. >> We'll have like the rest of my band sit there and say this is what we're learning. >> Mhm. >> And I'll use like I'll try to use like some of like that the stems although like that the bass stem isn't exactly where it should be. So you can get the bass but you also get the bass to drum with it. >> Okay. >> Yeah. Yeah, the only >> maybe that'll uh >> maybe that'll clear it on up. >> Yeah, [snorts] the only thing I could say is just to [ __ ] prompt it a lot and try using different combinations of words. Like you can even like use chat GBT as like a like the the source or whatever it's called. And um [snorts] I guess just try different ways of describing the same thing like uh I guess bring me the horizon metal core >> like uh >> who knows what it is like it falls into [laughter] that same category as like Sleep Token and and Dayseker and all those cats. >> Oh okay. Yeah. [ __ ] love Dayseker. So yeah that that's definitely uh >> Oh yeah. Dayseeker is sick. Yeah. Oh god, I love that [ __ ] Um, so yeah, if you're going for that, I would definitely use uh metal core for one of the genres. That'll probably help lean it towards that. >> Oh, okay. Okay. Yeah, that um and I know kind of tightened up really really hard on like the whole cover thing. I was on a sweet roll for a little while where I was able to strip songs down and then >> upload them and then put them in a different genre like that. I mean, I never did the whole soul thing because it's like everybody was doing that. So, I took like >> Taylor Swift and made it like progressive rock and it turned out like really really good, but now you but now you can't even touch that sort of thing now. >> Yeah. Yeah. They've really tightened up on that [ __ ] And I've noticed even with like my own melodies, um, they've tightened up a lot on like it'll say that it's copyrighted lyrics or something. >> Oh yeah, no doubt. Well, thank you so very much, man, for taking my call. I appreciate the words of advice. >> Yeah, doing what you're doing. >> Well, good luck with everything. >> Thank you, sir. >> Yes, sir. You never take me to Bangladesh. This is fun. I never done some call-in stuff. Uh I got a call at 9:20. Uh or is that the one that just ended? I don't know. Uh how how did you call Suna? >> [clears throat] >> Hello. >> Hello. >> What's up? Uh, Victor, is that what it said? >> Yeah. Uh, I want to play you some samples I made and see what you think about them >> using. >> Oh, okay. Yeah, I guess we could do that. Um, >> like if you can give me any advice or stuff, but this is like kind of like a tailored twist on I mean using >> you think about it. >> Uh, yeah. Would you be able to send me like a link somewhere? Uh either >> I haven't uploaded anywhere, bro. So like I still have them like a sample on the my computer. >> Mhm. >> Yeah, but like I've been watching your videos and like uh that's how I started like doing this. >> Oh, [ __ ] yeah. >> For a while. [snorts] >> Yeah. But uh yeah, I don't know like if you can give me any your thoughts on this sample I just made. Um, it would be it would be difficult to do it like over the speaker phone because uh like I won't be able to hear it. Well, >> can I send can I send it to you through Instagram? >> Uh, I don't have Instagram. I do have email and I have uh Discord. Um, >> Discord. Okay. What's your Discord? >> I think it's just Chill Panic. Let me uh let me look. Yeah, I'm uh I'm pretty pretty sure it's just Discord. I mean, chill panic. [snorts] >> You never take me to Bangladesh. Um I'm [ __ ] stupid and don't Yeah. Yeah. Uh it's just chill panic. Nice. >> It's chill panic. >> Yep. Uh same way it's spelled on YouTube and like uh together. >> Yeah. And then uh Okay. Send a request. special word. Okay. Yeah, I sent you one. I'll send you some stuff there. But yeah, let me know what you think. >> Okay. Yeah. Uh >> All right. >> Little message. Damn, I got a lot of message request. [ __ ] >> Catch you later on there. All right. >> Okay. Sounds good. Take it easy, Br. >> All right. Bye. >> See you. Is there a way to? And yeah, sorry. If somebody if one of y'all call like while I'm on the while I'm on a call, would somebody just uh feel free to call back? Um, is there any way to upload MIDI to Sunno? Um, not that I've seen. I don't think so. Um, I've never really tried uploading MIDI to Sunno, though. But I'm like 90% sure you can't. Let me uh let me uh make some MIDI. Oh, you can't even see it. Uh my bad. I'm being a being a terri terrible streamer today. [laughter] >> Buzz killboard. Buzz. Hello. >> Yes. >> Yes, sir. >> What's good? >> Sorry I interrupted your MIDI. >> Oh, no. You good? >> You >> It's good. This is fun. >> I got a I'm up all the way up in northern Ohio. I got a small studio up here. Um, and I found, you know how you uh go to upload even an original song in the suno and you'll get a copyright strike. Mhm. >> So, I found a couple workarounds that have been working for me. >> Oh, yeah. >> Take the track and raise it or lower it a couple semmitones. And even if you change the tempo a little bit and then you upload it and it will accept it 90% of the time and then you do your thing, you know, you add your tracks or whatever and then download it and then you just take it back to the original key and speed and put it in your doll. I've had to do that several times lately, man. It's driving me nuts. >> Yeah. Yeah. That's uh I don't know what happened. And I guess with the lawsuits, they had to get like really strict on it or something. Um, but damn, that's that's good. Do you just do it like one or two semmit tones or like uh like you have to change it a lot? >> No, usually I drop it maybe a full step or raise it and then uh I have taken the tempo up quite a bit >> on a couple tracks and then it will accept it. And uh I don't know, man. One time I just recorded some goofy guitar like it didn't make any sense. >> Basically noise on the strings and tried to upload it and it went and accept it. >> Uh >> um I don't know if it was just glitching out that day or what. But also uh this isn't really a question. This is another suggestion. I have a couple buddies that are DJs up here and I suggested to them and they go that's a good idea but neither one of them has tried it yet. I I suggest on their laptop at like weddings and things they do like a improv like improv comedy but you'll you'll take you'll pull someone's name out of the hat, a subject, >> maybe a hobby or something and you'll ask them what kind of hobby they like or whatever. Anyway, you prompt that all in sunno and you hit create a song and it's going to pop out, you know, something like a funny song for them right there on the spot. >> Oh yeah, that's >> I just kind of thought that I thought that would be kind of a fun thing to do for some of these DJs out there. >> Yeah, that would be interesting. It'd be like a crowd work for DJs. >> Yeah. So, I I'm all the time doing that, goofping off on my phone, making funny songs for my friends and stuff. But anyway, I wanted to suggest that raising or lowering the track [snorts] a couple semmitones, maybe speeding it up 8 to 10 beats up, then it will upload, bring it out, put it back down the original. >> Okay. >> Drag it on over into your doll. So, I appreciate >> [ __ ] yeah. I feel like y'all are teaching me [laughter] more than I'm doing. So, uh, I appreciate that. >> A lot of frustration on it. [laughter] >> Yeah, I'm definitely I'm gonna have to steal that for a video. >> Yeah, do it, man. Do [laughter] it. Buzz Kilgore out. >> All right. I appreciate it. >> Call from Silicon To accept, press one to send a voice. >> Scooby-Dooby Doobydoo. >> Scooby-Doo. What's up, CP? >> What's good? [laughter] >> Hey, man. Bad nickname. >> Uhhuh. >> I said CP is a bad nickname, but um >> Oh. Oh, yeah. I guess that that might that might suck. >> [laughter] >> Uh but yeah, so I I've been uh watching your channel for a while, man. And I've never seen you take any calls or anything. So I was like, you know what? I want to let him know that, you know, I watch a lot of stuff and and I've helped a lot of people with a lot of pseudo stuff, and I've watched a lot of other creators. >> And I'm going to tell you, you have an amazing and creative way of bringing information to people. And you can tell that, you know, you are chill panting. You know that you may put a little bit on it, but that's who you are and that's how you deliver it and that's why people get so much from your content. And uh I just appreciate the hell out of you, man. And I like how you stay on the cutting edge of everything that's going on with Sunno. And you know, I've gotten into a lot of the local stuff now that I've got the the uh >> dual 5090s and stuff and doing local processing of music. >> Um it's not up to par with Suno yet, but it's getting close. >> Yeah. Yeah. >> And uh but I I enjoy your stuff a great deal and more importantly the way you teach. And I was thinking, I hadn't got any of your books, but I talked to a guy um that had sent me like some screenshots and he he had got one of your books. >> Mhm. >> And um I was like, you know, I had I had thought about putting out a book like that. And he was like, "No, you got to see how he's got it lined up." And he sent me some stuff. And I got to thinking, you should put out um a a video or, you know, not a video, a book. I'm sorry. you should put out another book or an add-on um where you're doing a lot of these style control configurations. >> Mhm. >> Because uh some of the stuff that that you put together there is is really tough to do. And I mean uh I tell people all the time the the results that I get. I I you know I use AI for you know almost all of it. And uh you know I've built the library over time. So to see you do it on the fly in your videos. I don't know if you make notes for yourself and say I'm going to do this this. I mean I'm sure you have notes but I mean I don't know if you do everything on the fly but some of the stuff that you produce is amazing. Um and I think you should do a side book and I would definitely buy that even if it was small. um teaching people how to apply the different styles, es especially style shifting >> like inside the track, you know what I mean? Like in the middle of a song going from one style directly to another. >> Mhm. >> And u because it's really difficult to do and uh when you do when you do hit one and you get lucky, you're like, man, that sounds great. But you can never make it do it. And uh I had so much respect for the video of yours. I watched the other day where you're like, "Look, people, you're never going to make it do it the same way every time. It's never going to listen to everything you say." And I wish that everybody would listen to you when you tell them that [snorts] because I think that's one of the the truest statement there is. It's just very difficult to get the same results every single time. >> Yeah. >> No matter no matter what formula that you use. But uh yeah, man. That that's my soap box. I just want to tell you I appreciate [laughter] you. You do a great job and uh you know I watch your videos every single day and if if I do happen to miss something I always come back and check them out and I think you're really good instructor especially for Zuno. Um and that you got a lot of passion for it. I believe that a lot of people are learning from you and that you're going to you're going to see a lot of uh improvement in your numbers and your visibility uh pretty soon. I'm sure of it. Damn. Well, [ __ ] [laughter] I appreciate the kind words. Um, >> yeah, absolutely, man. And and I mean them. That's that's more importantly, I mean it. So, you know, I I a lot of people when you when you've got a big channel and stuff and uh you know, people say, "Well, put me on your channel or do this or do that." And it's like, it's not that simple. You have to, you know, I could say your name all day long and that's not going to make people come to your page. and you have that special sauce and I think you need to stay to stay to it because um I'm sure you watch other people and I know you're probably your own worst critic >> but you kick ass. >> Well, I appreciate you you you do you kick ass at teaching. That's what I'm talking about specifically teaching people. >> Mhm. And uh I think that should be your focus if I told anyone anything. I mean I'm an old dude you know I'm in my mid50s and you know and I I still jam as much as I can but um you know luckily my my my my young beautiful 36 year old wife that says I don't look or act old. So um I I just think that you do a great job and I'm around a lot of people that that do a lot of high-end professional stuff in the music genre. um and in the industry itself and uh you know you you really do a really good job and I think that if you focus more on the teaching aspect of it instead of just the because a lot of your videos are just fun and kind of blah >> but the ones where you're really instructing people you do an incredible job and the thing where you're selling the slots keep doing that man I don't know how well you're doing but keep keep doing it. [snorts] >> I appreciate that. People are >> Go ahead. >> Yeah. No, I appreciate all of that. And uh yeah, sometimes hard to do the educational stuff. Uh which is why I like go back to entertainment sometimes because it just it kind of feels like most of it I'm just kind of repeating the same like main things over and over again. >> But then when I do those videos, I get a lot more positive feedback. So I, you know, it's uh it's hard to balance like new people coming in and then uh people who have been around for a while and trying to make something that's valuable for the new people as well as people that have been here for a while. But um >> I try to tell I try to tell people, especially people like yourself that have passion in a particular area, it's it's a different thing. You got to remember a lot of the times um you know people ask you a question and to you it's just like total redundancy. You've said it a thousand times >> and most most instructors most instructors and teachers when they teach things they're really not teaching. They go hey look you know pay me to watch me do something you won't be able to do when we get done. >> And that's not what you do. you take even [clears throat] even on your videos, I've never done your calls, so I I don't know about those, but I have a feeling that they're very good because even your videos where you're just saying, "Hey, this is how I do this. This is how I do that." >> It's to the point. It's it's direct. It's fast, and you don't use a lot of big words when it's not necessary. You don't point people in complicated directions. You have personality that goes along with it, which makes it fresh and relaxing. And I think that uh that you could really win in that. You may think, "Oh, you know, that's a drag. I get tired of this, the monotony of it all." But I think that if you start focusing your videos, your time, you know, professionally for yourself, not so much others. >> Mhm. >> But I I think that's where you'll shine because you really you really have a knack to say, "Hey, you know what's up? I'm Chill Panic and I'm gonna teach you how to do some stuff today." >> And and they actually learn it. >> [snorts] >> Well, I appreciate all the advice and all the kind words. I'm going to keep that in mind. >> Yes, sir. And uh and I'll I'll drop in time and say hello to you if you do any more of the uh any of the phone gigs anytime. And I'll see how you doing. I uh I've been working on a lot of AI stuff lately. So, I haven't been able to play much with my music a lot lately, but I still enjoy watching your videos. So, stay stay strong and stay at it. And I I hope to see your numbers come up real soon cuz I looked at your analytics earlier and they're, you know, before I called and they're improving. They're improving and I can see you're about to hit this curve that you're going to be shocked. So stay at it, man. >> I appreciate it, Broki. And good luck with everything with the music. >> Yes, sir. Thank you so much. Have a good one. >> Yes, sir. You too. >> Take care. Uh yeah. So if you called before when I was on that phone call, feel free to to call again. It'll be fine. Uh let's see. We were talking about MIDI before. So let me go back to Streamy. Let me go to this one. [sighs] Uh I forgot to ask you who you were barricading out of your room with that bar on the doororknob. Um, I just, uh, you know, my door doesn't lock. Like, it does, but it's it's got one of those locks where you can just use a [ __ ] coin in it. So, uh, I do that just to keep everybody out. And plus, I watch true crime a little too much. So, I'm a bit paranoid. We got a severe weather threat. What? No, we don't. screenshot. Uh yeah, somebody's locked in in there for sure. It's me. I'm locked in, [ __ ] By the way, by the way, uh quick little promo. I got the 10,000 SEO prompts ebook and the 500 prompts ebook [ __ ] uh tagged or whatever on this live and everything is 50% off if you use the code rib like R I um and this is what it is. It's just like it's just three PDFs because uh I literally hit the character limit on Google Docs which I didn't know was possible. And it's basically got 200 genres with 50 prompts per genre. And then every [ __ ] genre has like its own descriptor page. So it can give you like ideas for the genre tags, the mood, and the instruments and vocals. And I broke it up with the genre, mood, instruments, and vocals cuz, you know, it follows that whole uh GM formula with genre, mood, instrument, and vocals. Just like the the basic like prompt formula. All right. Done selling you. That's all. Just wanted to let you know that. And uh [ __ ] What are we doing now? MIDI. We were going to [snorts] Oh, how does Sununo handle key changes? Um poorly. I don't know. It doesn't handle music theory very well. It's [snorts] uh yeah, that's the thing with a lot of the [ __ ] like a lot of the times with the lessons I'm mainly telling people like it just [ __ ] just will not listen to you is the gist of what I tell people on [laughter] these lessons. Um, and there's also usually people have questions about changing lyrics and stuff like that cuz that's a whole [ __ ] rabbit hole trying to edit songs and stuff. But yeah, a lot of it is just that Sunno just will will not listen to you half the time, unfortunately. But um learning the the genres and understanding what different genres sound like can help a lot because I was trying to do something that was uh it was like a Creed type of song for like a a video I'm working on for like AI music videos and stuff and I like it was not working because I just got the genre wrong. Like I knew it was post grunge, you know? Uh, but I had it as postgrunge rock, postrunge hard rock, and um, it didn't really seem to work until I just did post grunge without adding any other type of thing to it. So, uh, you can get more control just experimenting and really trying to hone in on the genre because I think that's the one that it takes into consideration the most and it's going to put most of its uh, [ __ ] like processing power towards. Uh, what instruments do you actually play? I play guitar. Like I would I would consider myself like uh like uh, Damn, that that's pretty [ __ ] good. He can play the guitar level of guitar and I can sort of play piano. I [snorts] can play piano well enough that I could trick you into thinking I was good at it, but I'm actually not. Uh what's with the shades? Uh I had a I had an alter ego named Hip Seed Broomstick. He's a rapper and he raps like this. And um these were his shades because he's a mentally handicapped rapper who puffs a lot of trees is that's kind of his whole thing. And uh then I never did anything with that. And I needed some way to brand myself and I had the the shades laying around [snorts] and that's how they became my thing. All right, I guess uh nobody wants to call anymore. Anytime. J4Z666. Uh yeah, let's try to make some Let's do some MIDI stuff. [snorts] And I I don't even know why I'm doing it. I'm like 99% sure that you can't do any MIDI in Suno, but why not? Why not give her a little shot? I just made this up on the spot. That's probably going to sound like [ __ ] [music] Nope, it doesn't. That's actually Damn, that's kind of nice. Oh, [music] [music] I don't even know why I'm messing with sound design. We just We're just doing MIDI. export as MIDI file. Um, we'll just put it here. MIDI test. [snorts] But yeah, I'm I'm uh Yeah, there's no way you can use MIDI. No way. Uh, yep. Yep. Nope. No MIDI files. Uh I could try like where the hell is it? I could try just dragging it in, but uh unsupported file. Yeah. Yeah. No MIDI. No MIDI for Sunno. No MIDI for Sunno. The poison. The suno poison. The poison for suno. Your content is fire. Keep it up, man. Thanks for all your videos. I appreciate that. You're a sweetie pie. What a [ __ ] sweetie pie. There's been a lot of sweetie pies in the building today. Um, yeah, I got a piss, but I'm not going to. Yeah, you're damn right I'm not going to. Um, so yeah, I don't know what the hell I'm doing now, really. Uh, so I guess like a workar around for the MIDI thing is you could just put the MIDI just into an instrument and then upload that to Sunno. You know, you could just turn it into a audio file. When I play guitar, my neighbor threw a rock through my window so he could hear it better. That sounds completely believable and something that definitely happened. That's awesome. The buzz 66. Where's the buzz 67 at? Um Oh, we were doing some trap soul earlier. Can you try melody to lyrics on song AI farm? So, sorry if this disturb you. Um, no, it doesn't. It doesn't disturb me. I just like I'm like jealous of this site. I wish I had thought of this first because this looks like it was vibe coded. Um, so I feel like I would have been able to do this. Song AIM will write lyrics that match every beat, phrase, and syllable of your music. That's kind of That's kind of fire, actually. What the [ __ ] That's cool, huh? Would you look at that? It's been there since March 2025. Oh, maybe it wasn't vibe coded. I don't know then. Um, okay. You know what? I'm I'm down to give it a shot. Um, I use Sunno and FL Studio. Same, bro. Would you ever consider doing a video or series on how to create the SNO track in FL using the stems imported from Sunno as a starting point? Uh [ __ ] Yeah, I could definitely do that cuz I've had to do that quite a few times for uh for folks I produce for. So um it's [snorts] becoming like a a normal thing now for people to use Sununo for like their reference track. What's good, Chub? Damn, it's been a minute. [snorts] Uh, Chubbs was here uh back in the the uh I don't know, golden days with the RZ battles and stuff. This channel used to be something completely different. [laughter] It was like all FL Studio and uh it was a whole thing. [snorts] Love that. Ed Ed Bass Masters, would you look at that? Would you just look at it? Just wanted to say thanks. Watching your videos has helped me finish an album idea I had. Well, that's [ __ ] awesome. Yes. Okay. You looking good. I hope things at the crib are all good. And tell Slick I said what up. Oh yeah, I forgot you know who Slick is. Damn. [snorts] I appreciate that, Brusky. I hope you're doing well, too. I will definitely do that. Uh, nope, not bedtime yet. Uh, what's the possibility of uploading an instrumental melody and then having Sunno sing that melody from lyrics that we provide? Oh, damn. Damn, you [ __ ] with my Seven Lions remix. Thank you. I appreciate that. That's a That's a deep cut cuz that's just up on Soundcloud. Thank you. Um, let's see here. I'm moving out to Asia to live cheaply and make EDM music. Any advice or thoughts? Um, I don't I don't know if there's so many so many things in that that I could say. Ah, [groaning] but yeah. Um, yeah, I don't know. I'm starting to get tired for real. And, uh, don't, uh, we don't have any calls coming in. So, if you would like to do a call, uh, now is your chance cuz after this one, it's going to be my last one and I'm I'm going to piss and take a shower. Maybe even at the same time. I might eat, piss, and take a shower all at the same time. I might eat, piss, take a shower, write a song all at the same time. >> [clears throat] >> M [groaning] I might piss, take a shower, eat, write a song, and read a book at the same time. Thank God that was I [laughter] was going to keep going. >> To accept, press one to send a void. Hello. >> Hello. Hello. >> What's up, >> Mr. Canada? >> And I was like, "All right, [ __ ] it. I'm [laughter] going to call. >> Let's go." >> Uh, I was the one who asked you about uh Asia and you said you have a lot to say and it would mean the world to me if you give me any advice or any opinion. I would take it to the heart. >> Oh, yeah. I was kind of joking cuz just like uh that's a lot at once. Like you're moving out to Asia and make EDM music and it's like uh like I I can't give any advice on moving to Asia. I've never done that before. Um >> what's your idea about how uh EDM is going right now? Like do you think it's dying? Do you think it's evolving? Do you [snorts] think like it's a genre that like I should invest in because it's a bit like around EDM and big rooms will >> uh well I guess that depends on your goal. Are uh you're like trying to make a living with like being a music producer? Like are you trying to like ghost produce for people or are you trying to develop your own like sound and stuff and like be an artist? Yeah, ideally like I'm just finishing my bachelor in like cinema right now, but my heart is all about music. So, I'm moving out to literally follow that to be an artist. And >> for example, what um Kashmir is doing. So, I have originalities in um in Morocco or right now I'm in Canada, >> but I'm from there. I I'd probably like try to make the uh sounds uh whether it's Asian sounds, Arabic sounds with uh western sounds and [clears throat] try to build like something about that. But the goal is performately. >> Okay. Yeah. Uh yeah, Cashmere definitely did a really good job of blending those sounds together and uh his sample packs are crazy. He just came out with a new one too, I think. >> Yeah. Yeah. I got I actually the best one like I I get most of them illegally. >> Yeah, >> pretty [laughter] much. >> But is it is it something you wanted to do to perform and stuff like that? >> Um yeah, so the way I started out was um I didn't even start with EDM. I started um doing like rap beats just because I knew I liked music and I got FL Studio and started learning it and I just wanted to do something with music that got me paid. So, I figured I could sell rap beats and then I sold my first rap beat on Craigslist and um yeah, I don't know. Um, yeah, I started out with wanting to perform. Uh, as I got older though, that desire started to kind of dwindle and I didn't really want to perform as much. Uh, though I do still perform every now and then when, uh, one of my clients is in town because we have a few songs together. Um, but yeah, uh, I never like I never got successful in in that that route of things. I I what I got successful at is uh like being an educator and selling education, you know, so I I can't really give good advice on becoming like a touring artist or anything like that. >> Mhm. >> Um I will say though, I think it would be really good in the first few years or like like how how long have you been producing? um the last three years. I'd say I I wouldn't say I'm your I'm at your level. I'm learning a lot from you, >> but since I'm just something like my bachelor, all my focus was in cinema and how to shoot, how to film, etc. But I I probably sound a little bit uh [laughter] cocky, but I think I have something that is a little bit melodic that I want to share with the world. and I want to give it a shot. >> That's great. >> Yeah, that's why I'm like giving it a go. >> Um, and I think uh I don't I don't know if you share the same thing, but I go out a lot and I go to like these events when I see artist D and I'm like, damn, I I can give more, you see? And so I want to give it a shot. Yeah, you can you can definitely do it for sure. >> Um, and to answer your question from before, I don't think EDM is dying. I think uh just a lot of different genres are getting more like individualized now with like the whole era of having like a personalized algorithm and with music being so accessible, there's like so many more artists now. So, um I [snorts] think everything is getting smaller to a degree and um yes, uh J4Z666, I'm going to post this to the channel, but um yeah, so I think everything is getting smaller to a degree and a little more individualized, but it's hard to say for the future, but if it's like something you want to do, I definitely say go for it because uh uh I finally got to a point where I could quit my job, you know, and I do YouTube and music production full time now. And uh >> uh it was I never thought I'd be doing it like this, but um it was uh so it was like really messy and weird and I took a lot of different turns. Uh but it wouldn't have happened if I didn't start by like making music I wanted to make and things like that. And I got like my first like really highpaying client by because I was making music that I wanted to make. So, um, uh, and then that ended up just being like a serendipitous moment. So, but I mean that's like who knows that's [ __ ] I don't know if that's just magical thinking or or [ __ ] what. So, it's hard to say [laughter] like I feel like I can't really say anything with 100% certainty. But I can definitely say that uh I have had a more fulfilled life even before I quit my job by like trying to pursue uh doing it. >> A lot of people have done it. So you you could definitely do it. >> Thank you. Uh I think that's the message for everyone that is listening that is like never give up and continue and you never know what will be your message to the world. Like you said, you you want to help people, you want to teach them, etc., and you want to share wisdom and advice. Someone else would share it with a st, someone else would share it with experience, um, etc., etc., but am I asking too much? I don't want to I don't want to hold you too much, but I have a lot of questions and I probably have like three questions right now in my head. >> Oh, no. You're you're good. This is uh good for me, too. and getting an idea of where everybody's out >> perfect >> at. I mean, >> so a few days ago, um I almost switched to um Ableton >> Mhm. because I'm on NFL video as well, but and I almost switched and I downloaded the the uh the free trial and stuff and I literally hated it with absolutely my heart and [laughter] >> and I'm not sure about it because I love Kashmir and he uses Ableton. >> So I'm like, >> am I not gonna make it because I'm using Ableton? >> Oh no, that's [ __ ] stupid. you. >> No, that's uh I mean, sorry, no offense. You're not you're not stupid. I think that just that line of thinking is um I mean [ __ ] like Soulja Boy Studio. Seven Lions FL Studio Porter [ __ ] Robinson. >> Yeah. >> Um >> Martin Garrick. >> Martin Garrick Studio. Hm. >> No, but I'm like I'm I'm gonna ask all the stupid ideas I'm like insecure about >> to someone that is like has a lot of experience such as yourself just to get him out of my head. You know >> Yeah. [ __ ] uh have you heard of uh Herobust was that his name? >> Uh Herobust. He had like a few years in the hybrid trap era where he was everywhere. Um, he was making [ __ ] hybrid trap dubstep [ __ ] like with crazy sound design. He was in like uh I don't know, Reaper or Reason. I can't I can't remember. >> Oh my god. >> Yes. So like uh >> Okay. This software thing or just get out of my head. It doesn't matter. Right. >> Yeah. Yeah. That's what I would say. Um, >> do you think do you think like I'm a less um I'm a less of an artist if I make my demos with Junu Sununo because I I see you like using it as well. So I I do the same thing. >> Um I don't know. I guess uh I don't I don't think so. There's a lot of debate right now around that whole thing. That's uh I don't know. I think uh I think a lot of people would say that yes, that you're like less of an artist for using Suno, but I obviously don't think that. Um cuz I don't [ __ ] know. I mean, like, uh people said that people were less of an artist for using autotune. uh [ __ ] that Elvis was a demon for shaking his hips and like >> we got distortion in guitar from somebody I don't know I don't remember what it was some type of accident and then like they thought it sounded cool and kept using it and people thought that was stupid people [laughter] [ __ ] that that uh that Baptist church thought that uh Skrillex was summoning demons and [ __ ] with his music and people were like dubstep is robots having sex this shit's stupid. So, um I I think you get to decide whether or not whatever makes you an artist or not. And I think as artists, we all [ __ ] steal from our influences anyways. Like, um >> a lot of my old songs just sound like copies of Mayday Parade songs because I was really into them. And um you like develop your own sound over time by copying and interpret interpreting it in your own way. And I think Sunno can give you a lot of like ideas just from putting your demo in there with and it can speed up your workflow because you don't have to produce like 15 different variations of what can come next. You know, you can just get ideas quickly. >> So I don't know. I don't really have an well I guess I do have an opinion on it but >> yeah yeah but I I think we share the same opinion to be honest I think I think that same way as a tool to like develop your creativity and I've seen like Kashmir use it as well used it as well he used to track like for his intro in one of his pets >> oh really >> something yeah I mean I heard it I knew it was AI there's no way it was like he used the whole orchestra uh just for it to sound like that like it it sounded undermixed. You see, I I've watched like one of your videos uh on how to make um AI songs and that's one of one of the things like I'm going to go back to in the future >> because I'm like planning to steal some samples from the tuna because it's like really good. But um yeah, that's uh that's pretty much my my uh question. Uh, I probably have one last one. >> Okay. >> I'm thinking the whole year I'm not posting anything. I don't even have like an artist name yet that I keep dropping. So, I'm like I'm I'm not going to post on on like Soundcloud or YouTube channel or or whatever until like I have a whole thing and either start DJing and post [snorts] on social media or um hitting up labels. I'm not very sure about that on like that idea of hitting up labels. Is it something you tried before to contact labels? >> Uh, yeah. I used to submit a lot of stuff to label radar and it just uh it never worked out for me. Uh, the only thing I got on to labels was uh uh I did like some funk covers of uh popular rock songs and uh that got onto a label just cuz one of my friends um [laughter] yeah, I don't think he's watching. He can't be watching, right? I am watching. >> Oh. Oh, [laughter] okay. >> I'm just really like into the conversation. You have no idea, guys, how how important this is for me. Like, yeah. >> Oh, [ __ ] Yeah. I'm I'm glad. Uh I'm into the conversation, too. I just have to do something with my hands. [laughter] >> Yeah. But I also want to thank you so much for your videos. They were very helpful and there still are. And sometimes I slide like a comment and I always like expect an answer and you always answer and that's something like that is very very helpful. >> Oh [ __ ] I didn't know I was answering because uh there's a lot of comments I don't answer. I'll use it. >> Honestly, honestly, usually I don't really comment on on like tutorials and stuff, but like when I do, usually people don't comment because it's like old video videos. >> But like one of the comments that I u commented on you and you gave me like a whole advice about it, but then like there was something I'm like uh you were singing but like you were trolling singing. You weren't singing for real. >> And I said like, "Oh, but he can't sing." And then you dropped like a link of you for real. [laughter] And then I was like, damn, he can sing for real. And it was like old. [laughter] >> I was like, "Oh, okay. Well, he pays attention like to to the details." >> Yeah, that probably just hurt my ego a little bit. I was like, "Listen here, [ __ ] I can see >> I think it was sarcastic. I I think it seems like sarcastic. It wasn't like written as like seriously like oh please stop the thing in your head. >> Yeah. [laughter] But uh yeah, uh if I speak for a lot of people, I'm pretty sure you're you're helping a lot of people to be honest. And it's videos that we going to go back to a lot, especially with like you jump from um FL Studio to now with the AI and it's something like a lot of artists are unsure about >> and seeing you just >> confidently knowing how to use it and using it and you're like an expert in Apple Studio and production >> that takes like the stress of of the unsuress. I see someone said that's true. Hell yeah. >> [laughter] >> Thank you. >> Oh, and I appreciate you disciplination. Uh nine, sorry, somebody tipped 199 had to say thanks. Got to give it a like. Got to give it a heart. >> Um >> uh what you're doing, uh thank you so much [snorts] and uh of course I'm going to continue watching your videos and finding the comments. as uh one of many artists that watches you and um like when I w was watching you, you were around like 35k and now you're like at 72.5. That's just like amazing cuz you just continue. You're like a machine. [laughter] Like how does he drop so much and it's it's nuts. [snorts] Wow. That's amazing. Continue, man. >> I appreciate that. Um, >> I appreciate you. Thank you so much. >> I have to piss so damn bad. You too. >> I'll let you go. Good [laughter] night. >> You as well. See you. >> And uh you you uh you donated 199. So I feel like I need to answer your question. I feel obligated. So if that's what you were going for, you did it. If you were just trying to be sweet, then thank you and I'm sorry. Um to add two voices onto the same track into Suno. Um, I mean the only way I can really think of is you can do a duet. You know, that's the you can just prompt for a duet. That's the classic way to have two voices. Or you can create an instrumental and then create one vocal for it and then create another vocal for it and then stem it out and then take it either into a DAW or into Sunno Studio and you know, or you could just generate the vocals inside of Sunno Studio. But to get like two voices where you have like a lot of control over it, I think you'd need the Premiere plan to uh because you'd probably have to use Sunno Studio at some point if you wanted the most control. Uh but yeah, I don't know. What is the minimum jewels you accept? I don't know, man. I don't know anything about the jewels stuff. I've never really looked into it. Um, but you're a sweetie pie. Uh, just like that. Yep. Just like that. Um, I'm just a doodoo artist, but I'll be listening. I don't know what that means. Uh, with the just like that. Yeah, I'm really like that. And your best work is light pack. Boom. Um, yeah. I I have to pee so bad. So, I'm going to do that and I'm going to leave. Uh, sending you a gift or a jewel is just as difficult as creating. Damn. Well, I [ __ ] appreciate you trying Fenton Flawless. Um, you're a damn sweetie pie. And thank you to everybody who watched, everybody who commented, and everybody who uh called, and everybody who just watched like a stalker. Um, but I must leave and go to pee pe town and, you know, [ __ ] and eat and take a shower all at the same time and all of that stuff. Um, and uh, whatever else I'm going to do at the same time. You never know, buddy. Uh, but also before I go, um, I think I am going to do this again. This is kind of fun. I didn't really expect anybody to call. So, it's kind of nice. [snorts] Uh, but before I go, I got to tell you about the 10,000 Sunseo Promps ebook. Uh, let me tell you why. Why? Look, look. I had a rough month this month because because of the surgery [ __ ] Oh, I didn't record this. Ah, I can download it from YouTube. Anyways, so I had a bit of a rough month. So, uh, it's getting close towards the end of the month, so I'm putting a sale on [ __ ] everything to try to get some more bread. So, it's a win-win situation. It's 50% off on everything on the store. If you use this [ __ ] code right here, code rib, or if you use the link that's in the description, you can get all of it for 50% off. Um, but yeah, basically the 10,000 sunop prompts ebook, I called it the all genre pseudo prompt bible because it has literally every [ __ ] genre that I could think of and it separated out into three different PDFs because I [ __ ] it went over the limit in the first one on like the amount of characters I could have in the Google Docs. It was like a whole thing. Uh, so usually it's 97. Uh 50% off makes it some other number. And I know [snorts] I know the 97 sounds crazy for a [ __ ] PDF, but uh it took me like 3 months and I did literally test like these prompts to make sure they didn't just give some [ __ ] Um but mainly what what I think it's good for is exposing you to genres you didn't know existed. Like I learned a lot of genres that I didn't know existed just from doing this or genres like maybe I like knew the sound of but I didn't really know like [ __ ] Americana. And then you've got your genre, mood, instruments, and vocal tags. So you can use those to copy and paste and make your own prompts and whatnot. But yeah, just uh just that's there. It's on the damn store. It's tagged on the damn live stream and it's uh it's a [ __ ] 50% off and [ __ ] with the cold rib. If I try and send you a gift or jewels, I can't. Um I don't know. I don't know about the jewels. Sorry, I never got into it. And shout out to So Aai Farm. Yeah, shout out. I'm definitely going to check out So Aai Farm. Uh, that I didn't know that was a thing. No good deed goes unpunished. I think I have heard that before. Awesome book. I have it. I love it. Worth the investment. Damn, that's that's really nice of you. Teen Tinus. Tinus. Is that how you say it? Maybe I'll glaze in the sun pasture. [laughter] Oh, graze. Grace, not glaze. That makes more sense. Um, but yeah, Shakaron Macaron. Oh, Tina. Tina. That makes sense. That That's a name. Tinus. I was Okay. All right. Tina S, I guess. Anyways, y'all have a good night. I'm going to piss right here uh on my desk as soon as I end the live stream. I'm going to piss all over my laptop. And uh y'all have a good night. Watch out for thunder. Thank you. I I will enjoy the pee.
Does your AI music sound way too perfect? Like it's stuck inside a vacuum-sealed studio? Let's fix that. Here is the ultimate trick to force Suno to give you a raw live performance. First, wrap your section tags in brackets like intro, live crowd cheering. This tells the AI to build an acoustic space before the TRACK EVEN STARTS. >> [screaming] >> NEXT, WRAP YOUR SOUND EFFECTS in asterisks. If you just write applause, the AI vocalist will literally try to sing the word applause. The asterisks tell the engine to actually generate the sound. >> [music] >> Just listen to the crowd roar and the singer feed off that insane energy. Stop keeping your music in a digital box. Click below to watch the full video for my ultimate stadium rock prompt.
What an insane week in the world of AI
agents. If you want to know the latest
updates on Claude Fable 5, the latest
Codex feature that lets you record your
screen and turn it into skills, the best
open-source model in the entire world,
and if you want to know about the SpaceX
cursor acquisition and more, you're in
the right place. You're watching Agent
Native. I cover the latest updates and
news from Frontier agent platforms and
models so that we can learn about and
use AI agents effectively. My name is
Riley Brown, and if you want to become
Agent Native, hit that like button, hit
that subscribe button, and let's dive
in. Today, we're going to get started
with the most important news, in my
opinion, in the world of AI agents. The
company Z.ai released an open-source
model that I believe is like five or six
times cheaper than GPT 5.5, and some are
saying it's actually comparable and
almost as good as Opus 4.8 and GPT 5.5.
So, GLM 5.2 is a model released by Z.ai,
and this company is from China, and this
model is open-sourced, and it's much
cheaper than Frontier models. And by the
way, I'm going to show you exactly how
you can get this set up directly inside
Cursor in just 1 second. But, I first
want to talk about the benchmark. And
so, here are some of the benchmarks, and
this is what it looks like across the
board. You'll see that GLM 5.2 is
comparable to Opus and GPT 5.5.
Currently, I think the best model,
besides Fable, is GPT 5.5 with Opus
trailing just a little bit, but this
model actually held its own when I
actually tested it. Because normally,
when a new open model comes out,
usually, there are benchmarks that are
released. They don't actually tell the
whole story, or even in even close to an
accurate story. But, because they put
these cool graphs on Twitter, there's a
ton of hype, people make a lot of videos
saying that this model's actually
really, really good. And usually, when I
go to test that model, I just end up
incredibly disappointed. I actually test
the model and it does not pass the vibe
check. And as I tweeted earlier today,
this was not one of those times. This
model after spending a ton of time
actually using this model, I do believe
that it passes the vibe check. I think
that it's getting close to the frontier
labs, specifically Opus 4.8 and GPT 5.5
and I think this will actually cause the
frontier labs, OpenAI and Anthropic to
release even smarter models. I think a
lot of people realize that the models
that they rely on every day can be taken
away. However, with these open models,
you can actually download the weights. I
think a lot of people are taking this
time to test these open source models.
And so the best place to try out this
new model, GLM 5.2, in my opinion, is
directly inside Cursor and I'll show you
exactly how to set that up in just a
second. I use the Convex plugin and it
one shot a Trello app with basically all
of the different features that Trello
has with a database and authentication
and it works nearly perfectly or it
actually works perfectly. I also had GLM
5.2 go off to research about me, then
create a landing page, and then run it
locally. I also connected GLM 5.2 to my
Notion, to my Slack and to a ton of
other integrations and I was having it
just do general agent tasks for me and
it was doing a great job, just as good
as if I was using 4.8. And so yes, I
think the model was really good and
you're going to see a lot of people on
the internet saying the model is really,
really good. But you shouldn't take our
word for it, you should actually go in
and try it. So I'm going to show you the
easiest way to try this model directly
inside Cursor. So directly inside
Cursor, what I want you to do is follow
these exact steps. It should only take
you 3 to 5 minutes to get this model
directly inside Cursor. In order to add
the model to Cursor, we're going to be
using another tool called OpenRouter.
Normally, if you want to use a bunch of
different AI models, you need a ton of
API keys in order to access them.
OpenRouter allows us to only use one
key, so we can get access to GPT 5.5,
Claude Opus, DeepSeek V4, and in this
case, the most important one, GLM 5.2,
and then thousands of other models. This
video is not sponsored by OpenRouter. I
just want to explain why I normally use
this. And so, OpenRouter allows us to
add any model to Cursor. I'll show you
exactly how to do it. In Cursor, you're
going to go down to your plan here, and
you're going to click settings. Then,
what you're going to do is you're going
to come up here and you're going to
select models. You're going to come down
to API keys, and what you're going to do
is you are going to turn this on right
here. So, normally this is off, and you
are going to put in your own API key.
And then you're going to say override
the OpenAI base URL. You're basically
converting this OpenAI key into an
OpenRouter API key. And in order to
switch this from OpenAI to OpenRouter,
you're going to paste this exact thing.
I'll put the link in the description.
You're just going to paste this exact
thing in here, and then what you're
going to do is you're going to come up
to view all models, and you're going to
come down here and you're going to click
add custom model. Now, you can add any
model from OpenRouter here. And so,
we're going to go to OpenRouter, and
we're going to click models, and we're
going to look for z.ai/glm-5.2.
And you're going to see this little copy
button right here. You're going to click
copy, and now what you're going to do is
you're going to paste this model right
here, and you're going to click add. And
since I've already added it, it just
said it's already available, but for
you, it should show up somewhere in
here, and it should look exactly like
this: z-ai/glm-5.2.
Congratulations, you now have access to
the best open-source model directly
inside Cursor. Now, let's go test it
out. So, if you go to a new agent
session inside Cursor, and Cursor looks
very similar to Codex, you can select
any model. Here, I'm selecting
Z-AI/GLM2.
I can say "Hi, what model are you?" And
there you go. I'm GLM 5.2 by Z-AI, and
you are now ready to test the best
model. I want you to comment below what
you did with it and how good it was at
it. I want to know what you think of
this model. I'm genuinely curious.
Please let me know. And one of the
reasons why I think you should get into
using these open-source models and
testing them out is because the founder
of Z.AI, who created GLM 5.2, said that
they're going to get a Fable-level
open-source model, like a model that's
as good as Fable, that's open-source
within this year. So, someone said,
"What's the current timeline for China
to reach the Fable class or get as good
as Fable 5?" And Elon Musk commented, he
said, "Probably Q1." And then the
founder said, "Won't take that long."
So, that means he thinks it'll be done
by the end of this year. And so, that
means that in like 5 months, we could
get a model that is open-source that is
better than Fable, and it will likely be
significantly cheaper. I don't know
about you guys, but I think the best
place to use AI agents with my team,
especially for marketing, is directly
inside Slack. And the easiest way to
create cloud-based agents that runs
directly in Slack is with Hyperagent,
where the agent can actually become part
of your team. All you need to do is go
to Hyperagent, create an agent with your
favorite skill. This agent can watch all
of your channels, run on a schedule, use
integrations, and send updates directly
into Slack when something needs your
attention. For example, the first one
I'm building is basically a YouTube
researcher. It scans my competitors
using my YouTube researcher skill, and
it keeps track of what videos are
actually performing well. And it does so
automatically without me asking. Then it
suggests videos for me to make based on
the keywords and topics that are working
in my niche. And whenever I upload a
draft, it can generate 20 different
thumbnail options for the video, and my
team can quickly figure out which
direction is the strongest. The coolest
part is is that I don't need to remember
to open another AI tool and ask it to do
this every time. Because the agent lives
in Slack, my team and I can talk to it
where we already are working. It can
send us new ideas, run these workflows
on a schedule, and keep improving as we
add more skills and integrations. And
this is just one agent. You can build an
entire team of agents for your own
workflows. HyperAgent is giving away
$1,000 in credits to the first 1,000
people to sign up. Click the link below
to sign up. Claim yours now. So now I
want to move to the biggest super app
update of the week, and it involves
Codex. It feels kind of like a slow week
from Codex. They didn't really announce
anything that big, but they announced
one feature that I believe is incredibly
underrated, and it involves recording
your screen. Let me just show you how it
works. Directly inside Codex now, you
can use a plugin called record and
replay. I'm going to show you the
process for adding a Typefully draft.
Please make a skill called manual tweet
draft. So now, you could just tell Codex
by using this record and replay skill
that you want to show them how to do
something. So here, it's going to say,
"I'll use record and replay workflow to
capture the Typefully steps." Now, watch
this.
Look at that. It automatically turned on
the recording, and it says, "Recording
is now on. Show me the Typefully draft
process." So now, I'm going to go like
this. I'm just going to type, let's say,
"Comment." And now I am going to go
create a new tab. We'll go to
typefully.com,
and I'm going to switch to Riley Brown.
Hello, this is a draft by Riley Brown. I
can add images and videos. Now, I can
upload an image. And now we can do PNG
and here we go. And that is the basic
process. Once we're done, I'm just going
to hit stop. It automatically goes back
to Codex and it automatically enters I'm
done recording. And look at this. I'll
stop the capture now then inspect the
recording event. And now it's creating
this skill called manual tweet draft. So
then we should be able to just type
{slash} manual tweet draft and it will
show up here. It doesn't quite yet. Here
it summarizes exactly what I did and
this was a very short task. You could do
it up to 30 minutes. That was a 1-minute
task. You're allowed to upload up to 30
minutes for a task so that Codex has a
really good understanding of how to do
it because they have a really good
computer use. Okay, so it is now done.
And if you see here, we can actually
type {slash} we can type manual tweet
draft. There we go. Hey, can you please
upload the latest video to Typefully as
a draft? It's in my downloads, the
latest video there. And here we go. It's
off to the races and I believe we can
just open up Comet. Let me go ahead and
close this out. There you go. We can see
Look at this. Computer use is working.
That's its little mouse.
It clicked new draft. Now it should
upload a video or at least type out a
draft for it. There we go.
Upload an image.
Now it's going to find the last video.
There it is. Wow, this is crazy.
Wow, let's go. Ah, I need to upgrade.
Oh, no. That's super weird. I think I
just rightfully rejected it because I
says I need to upgrade. Okay, so that
video was just too big. Uh you can't
upload anything above uh 512 megabytes,
but you get the point. I recorded my
screen and I taught Codex how to use
Comet to upload something to a different
software and then I immediately turn
into a skill. And in order to do that,
all you need to do to get that started
is just use the uh record and replay
feature and say, "I'm going to record
something. Watch and make it a skill."
And you can tell it what you want to
name the skill, but it's literally that
easy. And so, potentially the loudest
news of the week came on Tuesday, June
16th, when SpaceX acquired Cursor. And
remember, the only thing that I care
about is becoming Agent Native and
talking about things that are actually
practical and useful to understand. I
don't actually care about this
acquisition except fact that I believe
that Cursor is going to be closing the
gap on both Codex and Claude Code. And
the main reason I think Cursor's going
to get so much better is now they can
afford to subsidize these plans if
they're able to train a model that's as
close to as good as GPT 5.5 and Claude
Opus. SpaceX is actually the fifth
largest company in the world, and so
basically, this $60 billion acquisition
means that Cursor gets access to
basically unlimited compute, basically
unlimited money and capital through
SpaceX, and they also, underrated fact
is they get access to the Twitter
distribution. I guarantee you Elon Musk
is going to be retweeting all of the
Cursor content trying to grow Cursor as
much as possible. And in return, SpaceX
sees this as a huge advantage to get the
best AI agent coding platform in the
world, arguably. They get all of
Cursor's developers who are very, very
good at what they do, and they also get
access to the training expertise because
Cursor did train Composer 2.5 and
Composer 3's coming out soon. So, the
teams are merging and I expect Cursor to
get significantly better. And I talked
about this in my full-length video when
I covered this entire story. I said,
"Notice here in the actual announcement
by Cursor that they didn't say for
developers. They just said useful AI."
And to me, this is an indication that
Cursor will likely become a direct
competitor to Codex and Claude Desktop
because they already have a really good
in-app browser. They already have
Composer 2.5, which is a fast, good
model. You already saw earlier in this
video that you can use open-source
models directly inside Cursor. This, I
believe, is going to turn into the best
general one of the best general agent
platforms. And so, the overall trend
from this news right here is I really
hope we end up with a very tight
three-way competition between Codex,
Claude Desktop, and Cursor. The more
competition, the more benefits they're
going to have to give to users, and the
better the tools are going to be for
everyone because they're going to be
fighting for all of the market share in
the world of AI super apps. And I
couldn't be more excited for Cursor to
get better. Okay, so to close out this
episode, I do want to talk about some
updates with Claude. And I think all of
us have kind of this weird taste in our
mouths surrounding Claude, and I think
we're kind of all in this Mythos or
Fable depression. And so, this is just
one of the tweets that I screenshotted,
but I've seen hundreds of tweets like
this. Something around the lines of, "I
don't know if it's placebo, but using
Fable for those days, it felt like it
just never gave up on problems and kept
trying crazy ways to get whatever you
wanted done. Now back on Opus, and it's
just kind of lazy. It thinks things are
too daunting and keeps asking if you are
sure. There was this sense when you used
Fable that you could basically do
anything. And one of the best benchmarks
for AI models is how ambitious can you
actually be? And one thing with Fable, I
felt that I literally wasn't smart
enough to even come up with an idea for
a thing that Mythos or Fable wasn't
truly capable of. And so for the past 4
months when I was in Silicon Valley,
right, I was talking to everyone and
everyone was talking about how good GPT
5.5 was. Now, they got access to Fable
for like 4 days and now they can't even
go back to GPT 5.5 or Opus 4.8. They're
literally in this Fable Mythos
depression where they just are waiting
for this model to come back because they
know that once it comes back, they're
going to be able to get done whatever it
is they're trying to get done in like a
fraction of a time. That's how good
Fable was. And so right now, it is 3:26
Eastern time on June 19th and Fable's
still not back in any of the Claude
products. It is still illegal to use.
And so right now, Anthropic is working
with the government trying to figure out
how they can get this model back into
our hands and we just have no clue when
it's going to come back. But beyond
being in this Mythos depression, there
are two updates I do want to talk about.
One of them touches on a theme that I've
been talking about a lot, which is agent
native apps. But the first thing I want
to talk about is Claude's new update to
their design mode. New in Claude design,
it stays on brand with your design
system across projects, lets you edit
directly on the canvas, syncs with
Claude code, and connects to more of the
tools that you already use. So for those
of you who don't know, if you go to
Claude
.ai and this only works on the web, not
on desktop, they have this feature right
here called design. So the first thing
that they announced is it says it stays
on brand with your design system across
projects. I haven't used this long
enough to test that. But what I can test
is that it lets you edit directly on the
canvas. So, I notice here there's this
edit feature. I think I can click Can I
edit this directly? The open-source
rival is here. Wow. A cheap Chinese
model that passes the vibe check. A
record GLM
5.2
is a very good model. Okay, this is
really cool. You can just edit things
directly on the canvas. This is really
fun, actually. And of course, you can
also do markups. So, I can say like,
"Don't have any of these here. I don't
like these." That's really cool. And the
next thing Claude added is they made it
really easy to share these and send them
to other tools. So, I can send them to
Lovable, Base 44, Gamma, Miro, and
Replit. So, I could in theory send it to
Lovable, and I could connect it, and I
could basically, if I designed a landing
page or a website, I could theoretically
deploy it on Lovable, or I could
actually just deploy it straight to
Vercel. I should have that already set
up, and we can send it to Vercel. And
yeah, I already have this set up. So,
now it's going to be able to deploy this
to Vercel. So, it can actually be on the
internet. And finally, something brand
new to Claude Code is artifacts. You
know that the Claude web app and Claude
desktop app already have artifacts when
you use the normal Claude mode. But, now
Claude Code can create artifacts, and it
can send little interactive pages. So,
here it's saying, "Research where users
are dropping off since the previous
release." And we can see here in this
video, it's just going to go off and
create a little mini app, or an agent
native app that you can share with other
people. And here it created this little
artifact. It has its own link. And now
it says propose a solution. And so, it
shows the current and the proposal in
this little mini app. So, you can get
the agent or Claude code to create this
little artifact or I call them mini
apps. And you can view them and you can
be like, "Okay, that's a good idea. But
here's what they're proposing. Okay,
yes, we can do it." And if you want to
share it with a team, you can just
easily press copy link and then you can
just send it to whoever you want on any
platform. And the example that they
showed was a phone. So, links made for
sharing. Here's the message and you can
if you send it to someone on your team,
they can very easily open it and look it
over. And so, you can very easily share
these little mini apps or artifacts,
whatever you want to call them. All
right, so those are the biggest updates
for the week. With Claude, we have
design mode and mini apps. With Open AI,
we have the record and replay to create
skills. Screen record to skills, really
cool workflow. We have the best open
source model ever created, which is GLM
5.2. And then we have these the
acquisition by SpaceX of Cursor. And the
main point of this is that Cursor is
very likely to get better and it will
likely become a better deal for their
$20 per month plan and $200 per month
plan. And we love competition between
Cursor, Claude, and Codex. It's very fun
and the open source models. We have And
And that's kind of the fourth bucket is
all of the open source models together.
And we just have so much competition
from all angles, multiple countries.
This is amazing. I'm very excited for
next week. Next week, these are some
things that I'm expecting based on the
rumors that I've been seeing around
Twitter and other areas. I think we're
going to see a return of Fable from
Anthropic or at least I'm really hoping.
There's been some rumors circulating on
Twitter about a new model by Open AI.
We could see some more open source
models being released. We're hearing
some of the other companies,
specifically from China, talking about
how they're going to be releasing more
open source models. Gemini might be
releasing a model, and then finally,
this one I'm really excited about,
Gemini may be making an announcement
regarding their super app. And so, I've
been somewhat harsh on Google when it
comes to them not deciding what their
super app is. They have way too many
products. Instead, I want them to pick
one, and here we have Logan saying,
"Feels like we are entering the super
app era." And I've been saying we've
been in the super app era for 100 days,
Logan. Choose your challenger. I'm
really excited for Google to just pick
one, whether it's anti-gravity, whether
it's Google AI Studio, whether it's
Jewels, whether it's their Gemini
desktop app. We don't know what their
super app is, so it's really hard for
them to compete because it's impossible
for me as a content creator to tell them
which tool to use. I don't know which
Google tool to use. Their models are
pretty bad. I really hope they catch up
because Google's one of the other
companies that can in theory be a
competitor. They just don't feel like it
right now. So, I really hope Google
comes back. Anyway, thank you guys so
much for watching this video. This has
been a really exciting week. Next week,
I will finally be in my studio filming
from New York City. Couldn't be more
excited. Anyway, I'll see you guys here
for the next one.
I've spent $30,000 on a launch video.
And I was told that that was cheap. This
entire 1-minute launch video for Spotify
is created by asking Pop Code with
Stable 5 and giving it spotify.com and
say make a launch video for [music]
spotify.com.
>> I've learned a bunch of little tips and
tricks to maximize my speed to getting a
good video. The main thing that I'm
adding now is assets. So, I'm either
adding screenshots [music] of UI or
examples from other things I've seen
online that I like.
>> I think a lot of the engineers here
actually can make their own launch
videos now.
>> [music]
>> They don't have to learn a tool. They
don't have to learn After Effects or
even CapCut. Everyone now has a coding
agent, so everyone can literally ask
your coding agent to make a HyperFrames
video.
>> All right, hey everyone. My guest today
is Ben, VP of Product Engineering at
HeyGen, as well as Jake, PMM on the
HyperFrames team. Look, guys,
HyperFrames is a freaking amazing tool
and it's completely 100% free for making
incredible AI videos just straight from
HTML. And Ben and Jake are going to show
us exactly what you can do and how it
works and how to prompt it. So, welcome,
guys.
>> Thank you. Thank you, Peter, for having
us.
>> Yeah, thank you so much. All right,
guys, well, why don't we just get right
into it and you can just show us how
impressive HyperFrames is. You know, I
can't believe it's free.
>> So. Sounds good. Sounds great.
So, this entire video end-to-end is
actually code. There are definitely some
assets, you know, for instance, me
talking, which is a piece of media that
we add to the construction. But, if you
actually look into our studio, you will
notice that it's really putting together
videos, audio clips, code, animations,
motion graphics into an HTML, CSS, and
JavaScript code base. And HyperFrames is
designed and constructed so that,
because of these data attributes that we
add to these HTML elements, our renderer
is able to turn such a code base into a
rendered MP4 video that you can share
anywhere. This is the power of
HyperFrames and very excited for people
to be able to do that. Everyone now has
a coding agent, so everyone can
literally ask your coding agent to make
a HyperFrames video just just like that.
>> But this is like super impressive, but
um why don't we start from step one?
Uh can you can you show us how do you
set up HyperFrames in like a Codex or a
Claude code?
>> Absolutely, absolutely. First and
foremost, I definitely encourage people
to go to hyperframes.ai.com/quickstart.
Here is the most recommended way to
teach your coding agent how to use
HyperFrames. And you just copy this and
you go to your
you know, for instance, you just go to
go to your terminal and then you run
this command. It'll pull in HyperFrames
skill and
you can it'll obviously take the steps
to do that.
Or you can also find HyperFrames in the
Codex plugin store. Go down in the
creativity section, there is HyperFrames
by Haijan. Just have to install it, try
it in chat. Lastly, if you don't have
those tools but you still want to
experience the capability of
HyperFrames,
you can go to Claude and in the
customize connect connectors or connect
your apps,
search for HyperFrames. HyperFrames is
right here. You can install it and you
can literally go to any chat in Claude
and say, "Make me a video." Something,
right? For instance, here I'm asking it
to make me a video about Taylor Swift. I
think it's
>> Got it.
>> not loading.
But yeah, that's uh that's the the easy
setup steps.
>> All right. And and and you showed uh
some pretty complicated code to generate
that amazing video that you showed for
Claude MCP.
Uh but you know, like I I don't know how
to write any code, so like what's the
easy version of trying to make something
like that?
>> Absolutely. I think this is the easiest
way for folks to start getting a sense
of how to work with hyperframes. We have
a skill which is fully open sourced here
in our hyperframes open source project.
You know, we have been working on many
skills uh to help people, you know, make
really good launch videos, uh motion
graphics, etc. etc. But this skill is a
very comprehensive skill that takes your
agent uh step-by-step to create a full
video. I think first we just see maybe
an outcome. We're also amazed by the
kind of content and uh video that Fable
5 is able to make using this skill. So,
this video I'll first mute it to talk
over it.
This entire 1-minute long launch video
for Spotify is created by asking Cloud
Code with Fable 5 and giving it
spotify.com and say make a launch video
for spotify.com.
And as you can see, Fable followed our
skill's instructions to take assets from
the spotify.com,
it wrote a full storyboard on how the
flow of this launch video should be, and
um
even the
audio.
>> Yeah, that was the same but okay, so it
just used that skill, the web website to
video skill, right?
>> That's right. That's right. So, quickly
kind of give everyone a high-level
overview. And and the reason why I want
to show this is that it's not necessary
to say, "Oh, everyone use this one
skill." But I think just by reading
through the skill, it gives you a sense
of how to use hyperframes together with
your agent. So, here this skill
essentially teaches your agent, all
right, in order to turn a website into a
video, here are the 1 2 3 4 5 6 7 steps
that you need to take, right? And it it
it details on every step. So, first
step, ask your agent to actually capture
the content from your website. So, we go
to the sub skill here. It actually
teaches the agent how to pull the
screenshots, the assets, and putting
them into
or a set of folders that, you know, your
agent then can use these images and SVGs
and assets and videos for its video. And
then you can see that step three is
storyboarding. You know, our skill
teaches agent to be more diligent.
Before building anything, build a
storyboard so that, you know, then it
knows to build the code scene by scene.
So, if we go back here
to the Spotify video, you can see that
the Spotify video is broken down into
scene one, scene two, scene three. And
this is how your agent uses the
Hyperframes skills, follows the
Hyperframes standards, and writes the
entire code base for this end video. So,
my suggestion is
literally, you know,
tell your agent, "Use my website, make
me a launch video." That's the simplest
way to use it. But,
if you actually want to get better at,
which I'm sure Jacob will talk a lot
more about, you know, other other tips
and tricks that we have, the study the
website Hyperframes skill, I would say,
is the best way for you to learn.
>> I mean, I can just straight-up copy the
skill, right? Like, I can just copy and
paste it.
Point code is like GitHub, be like,
"Hey, just go get the skill for me."
And [laughter] uh and then just like,
you know, so I I just have to type like
{slash} website to the video and paste
my website link. And then that's it,
right? Then they'll do it.
Yeah, that's incredible. How do you get
the voice? Like, is it through like
Eleven Labs or something or like some
some
>> That's a good question. Our goal here is
we obviously want to make it so that
hyperframes is accessible and almost
like completely free for people to get
started, right? So, we actually use a
local model that we have a skill for it
and your agent decides that it needs to
use audio, then it will download the
model and then actually use the local
model to do TTS. We obviously also have
ways for your agent to connect to
HeyGen, connect to 11 Labs, all these
different providers for your agent to
use text-to-speech or image generation,
etc. etc.
>> Okay.
Yeah, you know what? I I I I normally
don't like to do videos talking about
products like because it seems like
promoting a product, but like this stuff
is free, number one, and number two is
like so good. So, like
there's no reason why people should not
try this. It's like super approachable.
And Jake, you you've been quiet so far,
but like thing tells me that you're a
real expert in actually getting really
good videos out of this. So, maybe you
can share some tips or tricks that you
use.
>> Yeah, so I have um I mean, I think it's
important we lay out we've really only
been a team for 2 months. Um
>> Yeah.
>> So, so, you know, we're [snorts] all
getting acquainted with the tool as well
and getting better every day. I have had
the privilege of creating most of our
launch videos so far. So, through that
I've learned a bunch of little tips and
tricks to to maximize my speed to
getting a good video and then also, I
think, the output quality. I think the
key differentiator I'm going to
introduce here is that the types of
announcements we're doing, we don't have
like a website or something to to start
from. Um so, therefore, we have to do a
or I have to do a bit more of the
groundwork
um myself for getting like projects set
up initially, right? Because what we saw
with the website to hyperframe scale,
that first step is it's taking in all
this information from the site itself
and laying the groundwork. So, when
you're doing something that new, that
maybe you only have Figma screenshots
for or for us like a lot of our videos
are based around a a cloud code session
or similar, you're going to have to
build those yourself for now. Or, if
I've already made them, it's in our
launch video specific repo, which which
has the source code for all the videos
I've created. So, I'll talk more about
that later, too, cuz that's a a very
fast path to getting quality visuals
that you can then amend for your videos.
But,
essentially, I do I do these setup
steps. So, the first one is I create a
new project folder. And all I'm going to
throw in there is context, right? So,
some of the things we're we're
launching, all we have is like a read me
document about new feature that's coming
into our Hyperframe Studio or something
like that. So, I'll pull that file and
I'll add it into my project folder. The
main thing that I'm adding now is
assets. So, I'm either adding
screenshots of UI or examples from other
things I've seen online that I like.
Just basically laying the groundwork of
like,
I see a couple frames of this video
already in my head. Here they are.
And then, the key thing that you also
want to add is one aesthetic source. I
think most people might have heard of
design.md by now. We just released last
week something called frame.md, which
you can create on
hyperframes.dev/design,
but you just drop your design.md, and
then our agent will reformat it to be
better for video.
Okay. But, this is a really key thing
for matching the aesthetic of your
brand.
>> And design.md is just like a like a
bunch of like fonts and colors and like
style styling, right? Like kind of like
a brand guideline.
Yeah.
>> It's Yeah, it's basically just a brand
guideline, and because it's all just
like hex codes, it's more of a visual
direction, but because it's not written
into HTML, it allows the agent to take a
bit more liberty. and we kind of push
that a step further with frame.md giving
it the context of like instead of
building a a webpage which is what the
design.md is made for, you want to
maximize the frame and make things
larger and use motion and yeah.
Then my first step is the or the first
prompt that I give is I point my agent
towards the new project folder that I've
created. I point them towards either the
design.md or the frame.md that I've made
and then the last thing is I ask for it
to go through everything and then create
a as
a table of key events. So it just breaks
down scene by scene what this video is
going to be talking about, maybe a brief
explanation of what's on screen and
normally what I'm refining here is the
text copy, right? I really care about
what we're going to say and how it's
going to build into this video. So I
might take a couple shots back and forth
just being like, you know what? Actually
I think it should be this line or help
me brainstorm here, whatever. But this
is more just like
the the meat of the story
um
as opposed to any kind of visual
direction.
Then what I mentioned earlier is my next
step once I'm happy with like this kind
of overview of the video. For me my
files are all local so I point towards
thing the projects that I've made in the
past, but for other people they can go
to our
launch video repo repo and they can ask
their agent to pull specific elements
from the the videos that I've created.
So I want to give a quick example that
I'm going to share. We launch pretty
frequently and I try to not make things
net new if I can. So I want to give an
example in our last three videos I have
reused
the same Claude prompt box just with
different frame.mds.
This is video number two. And then this
is video number three, right? So it
looks vastly different. But ultimately I
pointed my agent towards the same
first prompt box that I had created and
then described the the relevant motion
for each video.
>> But the prompt box is just a is just a
image or or like it's some code?
>> It's like a code.
>> It's code.
>> It's code, right? Okay.
>> Dude, so are you are you going to open
source this stuff or is this
>> It is. It is.
>> Take our scale.
>> [laughter]
>> It's all there.
>> You know, I I do want to call out that
we are and we have open sourced quite a
lot of these components.
Not that we don't want to. It's just
that, you know, I I don't think people
know or we have not talked about it
enough. But in our hyperframes open
source repo, there are actually at least
a good like 50 components that we use
regularly in our launch videos.
>> All right, so like all the cloud
components and every everything else is
there?
>> Yes. Yes. And we actually open source
every single one of our launch video
because it's really just a code base,
right? We open source the entire code
base that would render to the video. You
know, it's really hard for you to read
actual code base, right? It's a agent
written like hundred thousands of lines
code, but I think it it'll be a great
resource for your agent to point to. You
can literally say, "Hey, pull that piece
from that code base because I want that
effect." Your agent will do that.
>> That's amazing. Yeah, I'm I'm I'm going
to call it right after this. Yeah, that
sounds amazing.
>> Yeah, an an example prompt that I would
use is like, "I I really love the text
animation from this is the name of the
video. Can you grab that one and pull it
for my intro of this video, right?"
Those types of moves are what your agent
is going to excel at. And then it's also
going to allow you to more easily, if
you're doing things in parts like this,
it makes it easier to apply to an
aesthetic. And that's kind of my point
here, right? So if you if you see the
prompt box and it's the right
interaction, but your your colors are
different, you want like a slightly
different pacing, whatever.
That is the beauty of it being code is
you get to just
take the like the structure and you
already start with this baseline. So,
when you introduce these design changes
or whatever they may be, it's a lot
faster and it's more likely to work that
first try. The next big unlock that I
think we're going to be adding in pretty
soon to to everyone's videos, but
I create a storyboard.html
from that markdown file that I had
previously created. And what this is is
essentially I'm asking the agent to
create one frame per scene.
And it's showing me the most visually
dense section of each scene. And then
I'm asking it to use the the references
that I've gotten from the the launch
videos, and then finally the kind of
design system.
>> You just want to review these things
before it gets too far, right?
>> Yeah.
>> Exactly. Yeah, so this one is so much
faster. Like it it inevitably is going
to take a little while for your for your
agent to come up with the full
composition, especially if you're making
a 45-second video or or longer. And so
by just doing the static frame, you can
align on aesthetic that much faster,
which I think for a lot of people is is
a primary concern. So, this is this is
definitely a big unlock for me. Most of
these videos I'm making within a day of
the the day we're launching it. So, time
is so important to me
um in in that case. So,
you get something back um that'll look
like this where it it kind of says like
this is the hook scene, this is scene
one, scene two, and then it'll have like
one frame. This is obviously
a very lightweight version, but
essentially, yeah. What I'll do is I'll
go back and forth on this. And then once
I'm pretty happy with these static
frames, I literally will just ask, "Hey,
can you turn this into a full
HyperFrames video and pull it up in the
HyperFrames studio or use HyperFrames
preview
um when done.
>> So, just like what Jake's mentioning,
our studio will breaks up based on the
your code composition, obviously. The
studio breaks up your video into
multiple scenes. And within the scene, I
assume Jake you were going to talk about
the inspector.
>> Yes.
>> So, essentially, each scene, you know,
is its own motion, right? Sometimes, you
know, sometimes the scenes are
overlapping with each other. You know,
for instance, if I I was like, "Huh, I
don't I don't know. I don't know about
about this, you know,
text. I don't know about its position,
either. Maybe I want it somewhere else.
I want to change the text." Uh you can
all do that in the studio. And the
beauty of HyperFrames is that
as humans who are editing this, one, you
don't actually need to understand code.
You don't need to go to the actual
index. You know, the HTML code and find
the place and change, you know, these
things. You don't do that. We build the
studio so that humans can uh can modify
it using this UI.
And but the beauty of that is that when
you make those changes, they become
code.
And so, your agent actually knows what
you changed. Because the agent can
basically do a code diff or something,
right? And what's even cooler is that
because of the fact that LLMs are so
good at HTML, CSS, and JavaScript, it
knows exactly visually what this change
will entail.
And that allows the agent to
continuously working with you, working
with the human on making this video
perfect.
>> And I can also just tell the agent to
change like what what do you want to
play to something else, right? Yeah.
>> Yeah, absolutely. You can always do
that. Sometimes there are just like
these tiny changes that you don't even
know how to describe, right? And and
that's where the last mile editing comes
in, and that's where the the studio
really helps. And you know, when you
come from where, you know, Jake was
showing his like a storyboard, his
storyboard is essentially turns into
each and every one of these scenes, and
then he can then go into each scene and
talk to his agent on how to you know,
how to make each scene work to the way
that he wants them to.
>> Got it. Okay.
And then after I finish I'll just I just
hit hit export to export as a video?
Yeah.
>> That's right. And then it'll turn it
into MP4. We also support a couple of
different configurations that you can
take, MP4, MOV, and WebM. Do ask if they
can export a transparent background
layer. For a lot of the more, let's call
it more professional video makers, they
want, you know, Hyperframes to be making
the motion graphics, and then you can
download the WebM and then get the WebM
to put it into your
I don't know, Premiere Cut. But, you
know, either way, uh we support many of
these configurations and locally.
>> Dude, this is incredible. And like this
is just like all HTML and CSS, CSS,
right? Or or like
>> All HTML, CSS, and JavaScript.
>> So some So theoretically I can also
export it as a website or or something,
right?
>> Absolutely.
>> [laughter]
>> Okay.
>> That's right. Peter, you're touching on
some things that we're also excited
about because of the fact that it's
HTML,
uh we can make interactive videos.
>> Yeah, exactly.
>> Our player can be interactive. Yeah.
>> And I feel like that the slide that
there you were showing, Jake, is also
HTML, right?
>> Yeah.
>> Yeah.
>> Yes, it was also made with Hyperframe
skills, yeah.
>> Dude, you know, like I got CloudDesign
to make a slide deck, and then I got to
make a video, and the video was like way
more impressive than the slide deck. So
I I I feel like you guys can also just
expand to like the slides market, too.
Like just
>> [laughter]
>> You know.
>> We will try We will try We were
literally talking about it this morning.
>> Yeah. Why why don't we talk about cuz
you know, I have a bunch of PMs and
engineers watching this. So why don't we
talk a little bit about how this stuff
actually works? Like maybe you can walk
through like, you know, you're working
on HeyGen, which is about like avatars
and stuff, right? But but then all of a
sudden you have hyperframes. So like
where did that come from?
>> Yeah.
>> And like
>> Yeah. Great question.
This might take a little bit longer, so
uh bear with me. Um
I I I do think we need to take a take a
quick step back.
You know, at HeyGen, one of the things
that I think we really focus on we don't
compete with, you know, cinematic
videos. So like we don't compete with
like C dance, Bale 3, like, you know,
Hollywood. Like that's not our focus.
You know, HeyGen has always been known
to have one of the best, if not the
best, avatar models, right? And the kind
of business problem that we help our
users solve is communication. We always
believe that video as a format is one of
the most effective communication format.
Because would you rather read a
five-page doc or watch a
one-to-two-minute video to understand,
you know, everything that you need to
understand? I have a lot of examples of
how our users and internally how we use
video as a communication format.
So, avatar was our first step because
many people are, you know, shy in front
of camera, they don't feel confident,
you need to do retakes. I'm sure Peter,
sometimes you do retakes, you know? And
our users even more. They're not even
like, you know, professional, you know,
communicators or video makers, right? So
they leverage our our avatar models so
that they can finally show up in front
of camera and talk to their audience.
Because the people-to-people connection
is really important.
But just the people is also not enough,
right? You you need to have the B-rolls,
the motion graphics, the the explanation
of your product, the all of these like
video editing that then even more people
don't know how to do.
Like I think majority of us don't know
how to do video editing.
So, ever since about like I think the
beginning of last year, we have been
really focusing on, "Okay, now we've
nailed A-roll. We've helped people make
their avatar videos and so that they can
show up. How can we help them make that
final video? Cuz people just take our
video and then maybe go to like CapCut,
Premiere Cut to finish that video,
right? They even hire a video editor to
finish the video, right? So, how can we
take them from end to end? So, and we
obviously believe that AI agent is the
way. So, we've been trying to build a
video agent to do that.
However, we learned the hard way that
agents are even though very, very
capable,
when they are working with JSONs and
like, you know, XMLs, which is obviously
the backing data model that all of our
video editors sit on top of,
it has no visual intelligence.
Like,
when a a an agent writes a JSON blob, it
can be accurate. It can be verifiable
and correct structurally. But, agent has
no idea whether this JSON is going to be
good-looking or not.
>> Yeah, exactly.
>> It doesn't matter. It doesn't know.
>> Yeah.
>> And And that is actually the biggest
problem that we ran into. And
furthermore, is like, you know, human
modify that JSON through the UI and then
the agent is like, uh what changed?
Like, you know, what happened, right?
And so, that's actually when we turned
to code. Because we believe that code,
especially HTML, is
LLM's native language.
LLMs can express not only accurately the
information, but also visual aesthetics
through HTML, CSS, and JavaScript.
And it's not necessarily fully true
before like, you know, Gemini 2, like,
you know, GPT-4 time, but it's
definitely true after like Gemini 3,
uh you know, GPT-5 and, you know, Opus
models. Like, these models are
incredibly good at visually expressing
something using code.
>> Yeah.
Yeah.
>> And that really unlocked how our users
can just talk to our agent and edit a
video and the video will come out
visually interesting.
And that's how we uh so we started by
like actually just having agent write a
very small snippets of code and slap
that on top of like our our you know
video editor to all the way be like why
can't
it just all be code? And then our
footages will sit on top of it, you
know, any images, you know, assets, SVGs
can sit on top of it and then the code
just becomes that foundation layer for
agentic video making.
>> That's incredible. Yeah, I always
believe that code is the foundation of
all knowledge work and it clearly like a
a bunch of creative work, too.
>> Absolutely.
>> But how do you how do you like is there
some sort of verification loop that is
actually like you know making beautiful
scenes and stuff or
>> So, there are a lot of things that we've
noticed that agents are already very
good at, right? Like agents are already
very good at writing a landing page,
right? But you know, you hear a lot of
people using hyperframes and be like,
oh, I'm making a PPT video.
Uh which is fine, you know, PPT videos
are useful in some use cases, internal
use cases, it's totally fine. But when
it comes to launch video
it's not going to cut it, right? People
are not going to watch your PPT uh for
more than 5 seconds. So, what we also
found is that LLMs today,
especially with HTML, CSS, we call it
spatial aesthetics. Um spatial
aesthetics means like you you look at a
you know, a a landing page and your eyes
move from top to bottom, left to right
and you scroll down, right? All the
information are laid out and spatially
it looks great.
But videos don't work that way.
Videos we call it temporal aesthetics.
So, your eyes are always looking at the
camera or you're looking at the the
video
and the information is fed to you.
Like your eyes don't move that much.
>> Yeah, cuz of the time element. Yeah.
>> Exactly. There's a time element to it.
And and that we found is not
you know, is actually not something that
LLMs are very good at. Because it's not
being trained on top of that. So, we
internally build evals, benchmarks, and
also self-check loops so that our own
agent gets better and better at that. We
open source a lot of that ideas into our
skills so that your agent gets better at
that as well. But we're actually working
with Frontier Labs as well on how they
can train LLMs to be better at at this.
>> Got it. And and is the primary use case
right now like product launch videos and
kind of like Cisco real
like tech product stuff?
>> You'd be surprised. There are so many
different use cases. Product launch
video for us is the holy grail.
Because, you know, as a founder myself
from from previous to to HeyGen, I've
spent $30,000 on a launch video.
And I was told that that was cheap,
right?
>> Yeah.
>> And launch video is like so so so
important and we see that as like the
the top quality video type that we need
to get to. But there are many many
videos that today people use us for like
real estate videos,
you know, educational videos obviously,
internal training videos, or just motion
graphics that, you know, you make a
motion graphic for specifically
something. And internally I'll share one
last example, PR to videos or commit to
videos. Internally we literally ask
Cloud Code to look at my commits for the
last
seven days
and tell my team what I did for last
seven days. And it was really fun. It
was really useful. It gets everyone, you
know, a sense of like what everyone is
working on. And it's a fun like Friday
afternoon event where we just watch like
10 videos together.
>> Yeah, I mean like cuz I I I do a lot of
like product reviews and stuff and like,
you know, people share like documents
and slides and this stuff is just boring
to go through. And like some like people
started doing more prototyping and
stuff. But yeah, just having a really
nice launch video or like a video you
can share in Slack. Like people can
understand what the hell you're trying
to do in like 2 minutes. It's like super
use- useful.
>> Yeah. One of the things that we really
want to also tap into and we work with,
for instance, Hermes agent, we have a
deep integration with that with that
team. You know, one thing that we found
and I'm sure, Peter, you run many of
your own agents, right? Agents are
extremely verbose. They come back with
walls and walls and walls of text.
>> Yeah.
>> Um and I have gone to a point where I
just don't read them. All right, I just
[laughter] like All right, sure. Yeah,
maybe you did my what I asked you to do.
But, you know, we turn that into videos,
too.
We ask agents to be like, all right,
when you're done,
make a hyperframe video, tell me what
you did in 30 seconds.
>> Wow.
Wait, so so like cuz I'm actually using
Hermes agent right now. So, it's
natively integrated into Hermes or do
you have to install anything?
>> Yeah. It's a it has a hyperframe skill.
Uh you might need a run one command of
like adding a hyper, but it's already in
there.
>> Okay, you know what? You know what I
want to do? I I want to like give it a
bunch of pictures of my kids
and then make some sort of a reel based
on that. You can probably do that,
right?
>> Absolutely do that, yeah.
>> And dude, let me out let me try to push
you a little bit further. If I have a if
I have a separate video of my kid like
on the playground or something,
like I said before,
can I play that video within the
hyperframe video?
>> Yes, absolutely.
>> It can, right? Is it just called?
>> Yes.
>> Yeah, just called. Okay. All right, all
right, man. All right. I'm I'm going to
be playing with this a lot.
>> Yeah. What What we found, too, is that
especially for the frontier models, they
are actually also very good at visual
understanding. So,
>> Okay.
>> Fable 5 might be able to clip out that
specific like timestamps of your kids'
video to highlight in a hyperframe
video. Cuz, you know, hyperframe can
also just make it so that the that this
MP4 is played from like the third second
to the
>> Oh, I see. I see.
Okay, let's talk about Okay, so what
happens if I can afford fable 5? Like
what's the next best model for doing the
doing the stuff?
>> Great question.
>> Gemini or
>> Yeah.
>> Gemini.
>> [laughter]
>> Gemini? Really?
>> So, we have been doing a lot of testing,
a lot of evals. You know, if you want to
ask for like the top of the line
quality, absolutely like GPT 5.5, you
know, fable 5, these are the top tier.
But, Gemini definitely brings a, you
know, a quality to cost balance. Our
internal agent is built all on top of
Gemini.
>> Okay, great. Okay.
Yeah, I'm very excited. Very excited.
Okay, cool. So, let's just kind of recap
the whole process, right? So, I guess
step one is to go to the GitHub for
hyperframes and just like clone it or
like
install it.
Uh and then maybe install the website to
video scale. I think it one shot, right?
So, that's like a one shot scale.
>> Yeah, one of the things I I wanted to
quickly show is that we also have a
bunch of templates that you can
>> work off.
>> Yeah, if there is like one that you feel
like it's like close enough to you, but
you know, you want to change the colors,
you know, you can click on fine tune.
You can change the palette. It'll like,
you know, preview immediately.
You can like define your own colors
directly if you have them. Uh you can
also change the type typography. If you
don't like the original one, you can go
for a more, you know, funky or, you
know, a more I don't know. You can You
can change all kinds of fonts. And then
you just just download the design pack.
It'll get you a frame.md that works
really well with hyperframes.
>> Oh, that's perfect. Yeah, this this
looks like a slide. You should You
should definitely support slide slides.
>> Absolutely. [laughter]
>> Yeah, you should definitely support
slides.
>> We will We will work on it.
>> [laughter]
>> Okay, great. And and and Jake, like, um
you know, you were you were you were not
born to be a amazing
HTML video editor, right? So, so how do
you learn this stuff? Like how did you
become a good at this stuff over the
past 2 months?
>> You know, I started with small projects,
right? Like only a 5-second video where
I wanted to have some kind of uh motion,
and I learned how to describe it. And
then from there I just kept refining.
And then when I got like text effects
and other things that I liked, I would
turn those into a skill that I would
point my agent to for any of my my
videos after that, right? I reuse text
animations, I reuse prompts I create
just amending them to the specific video
at hand.
>> Okay. Well, I I guess it can use this
app. People can just clone the re-point
card, copy what we've done. So,
>> Exactly.
>> Yeah. Yeah.
Yeah.
Cool.
>> Yeah, I I I do think that what this was
really freeing for builders like
ourselves is that many even though Jake
did make almost all of our launch videos
in the last like 2-3 months, and you
guys wouldn't believe it, but we had
like at least 20 to 25 launches in like
2 months. Um and each launch comes with
I in my opinion a very good launch
video. Uh maybe not at the top of line,
but good enough. But I think a lot of
the engineers here actually can make
their own launch videos now because they
understand the process, they understand
the product, they just work with the
their agent, the agent makes everything
after a lot of iterations, right? But
they don't have to learn a tool. They
don't have to learn After Effects or
even CapCut, right? Which I think is a
is a huge unlocking for builders and
entrepreneurs who are, you know, who are
trying to build videos for their
businesses, but, you know, finding it
extremely hard to learn.
>> Yeah, learning a new like who wants to
like go to web website and learn new
tools? It's a pain in the ass. Like I
just want to get my agent to do it.
>> [laughter]
>> Yeah. So, so uh
okay, cool. So, then what are you guys
planning next? Like is going to be
another 25 launches
in the next month or so?
>> Yeah, there are quite a lot, you know,
Jake showed off the storyboarding. We
believe that it's a important step in
the video making, so we likely will open
source a lot of that. We're also
building what we call media use. We're
actually going to build a ton of skills
inside of the hyperframes so that your
hyperframes learn how to use background
matting, how to add sound effects,
music, all of those things. A lot of
that actually at you know, HeyGen we we
will offer a lot of that for free. While
some of them you will likely cost you,
right? For for more expensive models. So
yeah, we want hyperframes to be able to
do honestly
anything and everything that a video
editing tool needs to do.
>> Okay, I think that's amazing. I think
Okay, look, I think number one I I see a
lot of like AI builders now posting
hyperframes videos to launch their like
new GitHub repo or something. And that
that just feels like really empowering
because otherwise you would have to pay
like $300 to do it. Um and the other
thing is like you know, I've been on
YouTube for for a while and I I I still
haven't learned how to edit a video my
myself. It's just like I I don't want to
learn like Premiere Effects or whatever.
Like I don't want to learn. So yeah, so
I've been paying my video editors to do
all my stuff. And I'll probably still do
that, but like you can do little little
things with hyperframes. It just feels
like really empowering. Yeah. So so I I
guess I I want to thank both of you for
putting this stuff out there and making
it free for all of us to use. Uh where
can people find you guys online?
>> We're on Twitter. Find us through the
HeyGen X account. We post almost every
day about hyperframes. So find us at
HeyGen.
>> Got it. And the great thing about
hyperframes videos is like it just helps
you go viral on Twitter so you know, if
you don't go viral
>> It's true. And we will always retweet
your hyperframes videos if you
uh tag us.
>> Okay, cool. I'll definitely tag you
guys. All right guys, well thanks so
much for your time, man.
>> Thank you so much, Peter.
>> Thank you. Cheers.
Today on the 5 minute AI weekly recap, why this week was realignment week. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right friends, back with another [music] 5 minute weekly recap for very very busy people. Hopefully this helps you regular listeners who were particularly busy this week catch up. And if you have friends, colleagues, family who need a view into what is happening but don't have time for a daily show, send them this one. Now very rarely do we have weeks that have as consistent and clear a theme as we did this week, which was the realignment of the entire AI industry. Two big things happened last Friday, right after the time that I was recording the weekly recap. The first was the SpaceX IPO, which we had seen an initial bump as it went public on Friday afternoon, but then the second and arguably much bigger deal was Anthropic suspending access to Fable 5 and Mythos 5 in response to a new US export control directive. Both of these contributed to or were part of the realignment this week, and for sure the dominant theme was Fable fallout. Now to fast forward to the conclusion, throughout most of the week we haven't necessarily had all of the best signs that Fable 5 was coming back anytime soon. I think a lot of people expected that with the effective banning happening at the end of business on Friday, the White House had an interest in getting it back online by Monday, but that certainly wasn't the case. Now as I record this on Friday, June 19th, we are getting some positive signals, but at this point there is no resolution. Instead, what this week was mostly about was another lesson of why people and companies need to think about their relationship with AI models differently. Now this had already started because of growing token costs at the frontier. One of the biggest themes for the last few weeks has been people exploring alternative models and alternative model architectures such as routers. The fact that now models are seen as powerful enough that they can be shut down at random by the government as a whole new category of risk of overbuilding your strategy around one single model. And a lot flowed into that vacuum this week. One category of that was Chinese models. Indeed, one of the big critiques from people who are worried about this move from the White House is that it seems to be a complete boon for open source or open weight Chinese models that people were already looking to because of cost benefits, but now are potentially looking to because they can run them locally or have more control. Z.ai meanwhile timed their release of GLM 5.2 perfectly. It did well on all the benchmarks, but more than that, it seems to for many be passing the vibe test. Latent Space summed up the average experience with these new buzzy Chinese models writing, "In the AI news business, there's a bit of trepidation about talking about open models. They come out guns blazing looking pretty on notable benchmarks and then a month later they fade into disuse like they never existed. GLM 5.2, however, they say seems to pass the vibe check of being a frontier model that just happens to be open." They pointed to a tweet from Jeremy Howard, who is as they put it not one given to hype, who said, "GLM 5.2 is a marvel. It is at least as good as Opus 48 and GPT 55. It's super fast, inexpensive, and not too verbose. It responds with nuance and judgment. It handles long context very well. I've never experienced an open weights model like this before." Matt Pocock wrote, "Folks who are running GLM 5.2, how are you doing it? What harness and provider are you using? Getting FOMO about an open weights model for the first time." AI educator Riley Brown wrote, "Spent a lot of time using GLM 5.2. I've always been skeptical of the open models as they've never lived up to the benchmarks and announcements. This is the first model that passes the vibe check. This feels like a deep seek R1 moment that will push the frontier labs into releasing even better models. Time to buy a beast computer to run these models on." But as I said, it wasn't just Chinese models that were filling in the fabled gap, but also new model architectures. OpenRouter, for example, released their new Fusion API, which they say can achieve fabled level intelligence at half the price. Basically, the way that Fusion works is when a prompt is sent into Fusion, it's fanned out to a panel of models in parallel with a judge model that reads every response, and then selects the right model for the job. This is an example of the type of approach that people were already exploring because of token efficiency and cost needs, but now in the days of government AI shutdowns seems even more valuable. Something up the feeling of the shift overall is Mike McNally from USV who writes, "For the first time in around 3 years it feels like the AI table has been flipped over. Yes, the labs and hyperscalers will have the highest chance of resetting it before everyone else, but there is now a window for a new ecosystem to emerge. A rebel alliance, basically anything that gives people and enterprises powerful intelligence while maintaining tight incentive alignment." Now, this is interestingly where SpaceX story intersects as well. SpaceX's big pop on Friday extended into this week and holding aside the merits of the company's valuation, it gives them leverage and Elon is taking advantage of that. Specifically, SpaceX actually followed through with the acquisition of Cursor, which could have some pretty big implication for models. Cursor indicated that it's got a full model, not just a post train of a Chinese version coming, so I'll be looking for signals about how they plan on competing. Now tied up with Elon and SpaceX they could go two very different routes. They could continue trying to live at this Pareto frontier between efficiency and performance, or Elon's eyes might get big again and maybe they try to actually compete for state of the art even if it's expensive. Meanwhile, the one other area of Fable fallout that was on display this week was in geopolitics as European leaders at the G7 were caught between begging for access to Mythos/Fable while also trying to plan a new AI sovereign path. What to watch for next week then? Well, of course there has never been a more obvious what to watch. Everyone just wants to know if we will get Fable back. And in terms of what to work on or build this weekend, the conversation about loops and the different way of interacting with AI that represent is getting louder once again on Twitter. Future Forward's Matthew Berman just launched something that he calls Loop Library, which I'll link to in the show notes, and gives you a bunch of different copyable loops including for functions outside of engineering that you can go try and play with. So, that's it for this week in AI for very busy people or the 5-minute AI week. Hope you have a great weekend and see you back here for an operator's cut tomorrow. >> Mhm.
So, something is happening in AI. Companies are actually returning to humans and scrapping replacing workers with AI. And more information is telling us that this is likely to be the trend for the rest of 2026. The companies that swung the hardest at full automation are now actually walking back the layoffs, switching the bots off, and they're calling humans back to the desk they told they would never need again. Now, this is actually happening at a variety of companies. Starbucks, Klarna, McDonald's, IBM, Air Canada, and dozens more. Replacing humans with AI simply isn't working, and the truth is now too loud to ignore. So, essentially, this started with a Starbucks because the story is the cleanest. I mean, in May 2026, a few days ago, Starbucks killed its big AI inventory system. The system was called Nomad Go, and this was supposed to use cameras and computer vision to count every item in a store automatically, so that no humans would ever have to do inventory again. 11,000 stores ran it, and after a full pilot, Starbucks pulled the plug. And the reason was very simple. The AI got the counts wrong, and humans had to recount everything by hand anyways. So, the company is going back to manual counts done by baristas, and the promise was zero human effort, but the reality was actually double the human effort because workers were now doing their job and the AI's job. Now, you have to think about this logically, okay? The guy, Brian Niccol, who took over the company as a CEO in 2024, he actually pitched investors on a turnaround called Back to Starbucks, and AI was at the center of that turnaround. The story was that smart cameras and predictive software would handle inventory, scheduling, and forecasting, so the company could run with fewer and fewer hands, and investors loved it. Wall Street analysts wrote it up as the future of retail. The very first big AI deployment under that plan is in the bin. The CEO has not announced a replacement strategy. Company has gone quiet on AI specifically, and is talking about giving baristas more time with the customers. Now, what's crazy about this, okay? And this is where we have this problem in this industry that I've seen a time and time again, is that the Nomad Go demos worked in a controlled environment and with perfectly lit shelves and clean product. The AI hit 99% accuracy, and that's what Starbucks bought. But real stores are messy. Coffee bags get stacked sideways, syrup bottles get half hidden, lighting changes, and once systems meet the real world, the accuracy collapses. And workers reported that the AI constantly miscounted and they had to recount every shelf to fix it. The replacement plan died because the demo and the deployment were two different products. And you see this happening in AI all the time. How many times have they released a different model from the one that they demoed in those videos, and it simply cannot do the same thing? I mean, this is the unfortunate reality when it comes to working with AI. Often times, what you're finding out is that unfortunately, AI actually doesn't save time in many industries, it adds time. And at Starbucks, a barista's job was supposed to shrink because the AI was counting for them, but instead, the barista had to count anyway and also fix the AI's mistakes and explain to the manager why the numbers did not match. And this is the time tax of fake automation. The company buys the software, pays for the software, the worker still does the work, and then the worker does the extra work on top to clean up after the software. That's not replacement when you think about it, that's just replacement theater. Now, when you think about this, okay, people are finally starting to wake up to this, okay? After this Starbucks news broke a couple of days ago, analysts started to write that AI as a labor replacement strategy is now being repriced inside earnings models. For the last 2 years, any company that announced AI-driven layoffs got a huge boost in their stock price. The market just assumed that those layoffs would stick and that the savings would flow straight to the bottom line, but the problem is is that now the market is seeing layoffs reverse, AI tools get switched off, and rehiring is quietly beginning. And that financial story is shifting from AI saves money to AI costs a lot of money, and then we're going to have to hire everyone back anyways. Now, the most important number in this whole story is coming from MIT. So, in August 2025, MIT researchers, you probably heard about this before, they ran a study called the Gen AI Divide, and they found something pretty staggering, and that's that 95% of generative AI pilot programs at large companies were failing to deliver any measurable revenue or cost impact. Not bad results, just no results at all. And almost every replacement project, every chatbot, every automated workflow sat there burning money without moving a single business number. And only 5% of pilots actually worked. The other 95% were quietly burning the company money. Now, this is interesting because of course studies are running concurrently, and we are going to see more studies come out in 2026. But one of the things I think we should pay attention to is one of the most famous ones, which is of course Klarna. And that was one of the the most famous AI stories, okay? The buy now, pay later company. In 2024, the CEO Sebastian went on every podcast, every magazine cover, and he was just saying, "Look, Klarna has essentially built a customer service agent that did the work of 700 humans." And they actually froze hiring across the entire company and said that AI could do every job at Klarna. Their headcount dropped from over 5,000 to around 3,500, and he was the poster child for the replace everyone with the AI thesis. Every other CEO on Earth was being asked when they were going to do their Klarna moment. But the thing is is that in 2025, Klarna quietly reverses. Customer satisfaction crashed, the AI agent couldn't handle easy questions, but anything complicated, anything emotional, anything where a customer actually needed help, it failed, and people hated it. They called for a human and got a bot, and they got the same wrong answer 15 times, and then complaints started to pile up, and Klarna started hiring humans back, calling them gig customer service workers, and then started to route the hard questions to them. And then they actually had to admit on stage that they'd pushed too far on cost-cutting, and that the quality had suffered as a result. The poster child for AI replacement just became the poster child for AI rollback. Now, and so the worst thing about this entire scenario is that like AI replacing the easy part actually creates a bottleneck on the hard part because now the humans who are actually working there, they're only encountering the cases where the humans are super frustrated. So, they're going to get burned out and then the customer service scores are getting worse, not better, even after hiring the humans back because those humans are now only drowning in calls. I mean, this chart clearly explains it clearly where AI can do 60 to 70% of jobs, but it just isn't there for that, you know, 30%. Now, this is not the only story. We had McDonald's removing their AI-powered ordering technology from their drive-thru restaurants in the United States. And whilst this one is, you know, in early 2024, this is still happening across many different chains. So, McDonald's tried this same playbook and in a partnership with IBM, they ran they ran AI voice ordering, and they actually started started earlier in 2021. Now, the pitch was that AI could take your order faster than human, never get tired, and of course be there all the time. But 3 years and millions of dollars later, McDonald's just decided, "You know what? We're killing that product." And, you know, I'm pretty sure you've maybe seen videos of viral customers screaming at the AI as it added nine iced teas to one order as it refuses to understand a Big Mac, and as it tries to charge people for orders they never made. And, of course, drive-thrus are now back to humans, but that replacement experiment ran for 3 years, ended with the original workers walking back into the headsets. And, like I said, this wasn't just McDonald's. This has happened across many different industries, but the problem is that those edge cases still aren't covered by AI. AI isn't there yet. And you have to remember, okay? A lot of times people are going to say, "Well, you know, AI is going to change and stuff like that." But, remember, guys, that this is essentially built into how these AI models are. So, it's going to be really difficult to remove these things in the future. Hallucinations are part of the model. And when you talk about hallucinations, Air Canada tried to replace its customer service with an AI chatbot and they got sued. So now, not only are the hallucinations bad, they're opening people up to litigation. In 2024, the chatbot told a grieving passenger he could get a bereavement discount after the fact and the airline tried to argue in court that the chatbot was its own legal entity and the airline was not responsible. The judge said, "No, Air Canada was forced to pay." The story became the legal precedent that if your AI replacement tells a lie, your company is on the hook for that lie. >> [music] >> That is a big precedent, okay? Think about it. Companies are now going to have to think about the fact that since these LLMs hallucinate even a small percentage of the time, do they want to have to continually pay out in those small cases where they're going to be on the hook for whatever that chatbot hallucinates? That is pretty crazy, okay? Think about it. Companies are going to be thinking twice whether or not they want to be using these chatbots because is that cost saving going to outweigh the decision of potentially getting sued or having to pay up because your AI agreed to some ridiculous claim. Now, one of the most damaging admissions from 2026 has come from Uber. In December 2025, Uber rolled out Anthropic's Claude code to 5,000 engineers. They built an internal leaderboard to gamify usage. Adoption exploded and by April 2026, Uber's Chief Technology Officer told The Information that the company had burned through its entire AI coding budget in 4 months. Individual engineers were running up between $500 and $2,000 a month each on AI tokens. 70% of Uber's code now originates with AI and yet in May 2026, Uber's own Chief Operating Officer went out there and on the podcast, which I'm about to show you guys, admitted that he could not draw a line between the AI usage and any actual improvement in features shipped to customers. And his exact words, which you're about to see, is that it just isn't there yet. >> Uh our last quarter were AI driven um or you know, our token usage went from X to Y, or percentage of employees who, you know, all all these sort of numbers. Um, and it's amazing, and I think it's like this massive transformation of society, but then you sometimes go and you talk to your senior engineering leaders, and you're saying, "Okay, how many projects that were on the cutting room floor got moved above the line because of the, you know, productivity gains? Because 25% of our code commits were via Claude Code last last quarter." That link is not there yet, right? Like you you're not I mean, I think maybe implicitly there there's more that is getting shipped, but it's it's it's very hard to draw a line between one of those stats and, "Okay, now we're actually producing like 24.5% more useful consumer features, right?" And and that line is hard to draw. And I think over the over the coming quarters and years, like maybe that will become clearer, but I think today it's hard even if some of the underlying metrics are like trending in a really astronomical direction. >> And now, you want to know something crazy? In May 2026, the company Microsoft that invested $13 billion into OpenAI and another $5 billion into Anthropic, they actually banned the Claude Code for its own engineers. And the crazy thing about all of this, they were told to cancel all of Claude Code licenses after the engineers were reportedly using it too much. And so, the real reason this actually happened was because it's just really expensive. Now, I'm I'm going to be making another video about this, but essentially most people don't realize as well is that even the AI that is effective, things like Claude Code, it is remarkably expensive. So, when you have almost a trillion dollar I'm not even sure if Microsoft is a trillion dollar company. I'm pretty sure they are at this point, but when you have a company of that size that says, "Wait a minute, Claude Code is too expensive," that is pretty crazy when you think about it. If they can't afford it, who can? The company that sells AI to the world told its own people to stop using AI because they could not afford it. And I found another cost here that just goes to show that many of these agentic systems, even though they might be effective, they're still more than humans. Here, an Nvidia VP essentially says that their, you know, token usage actually cost more than their team. And if that's the case, why would you actually start to use these models if they're more expensive? The whole point is to save money here. >> I spoke with a VP at Nvidia who first flagged this to me. He said, "Oh, yeah, for months our costs for my team have been more for AI than humans." So, that was the first flag. And then we started to hear this coming out in droves. Uber's CTO said he already blew out his whole budget for 2026 just on AI-related costs. And obviously, that means he's spending more on that than he's spending on human workers. And now I'm starting to hear especially from startup founders, they're bragging about their AI bills being high because the kind of >> Now, do you want to know the craziest statistic of all of this? Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027. So, this is the crazy scale of it. This is the crazy scale of this. 40% of agentic AI projects canceled by the end of 2027. And one of those key reasons is escalating costs, as we've spoken about, unclear business value, or inadequate risk controls, according to Gartner. Now, when it comes to inadequate risk controls, I actually just found this tweet. So, you can see it says, "CEOs are quietly realizing their AI replacement plan has a problem. Two problems, actually. One, the token cost for running AI agents is now exceeding what they were paying the employees they fired. And two, when those tokens run out, the AI stops. Just stops. No continuity, no workaround, just a spinning wheel where your workforce used to be. You fired humans, you fired humans to save money, and bought a subscription that builds you into a corner. The employees you let go knew what to do when things broke, and the AI just invoices you for the outage. And then there's the permission problem nobody wants to talk about. To do its job, the AI agent needs access, full access, your systems, your patents, your contracts, and your future plans. Everything you spent years building, handed over to a process that has no loyalty, no discretion, and no skin in the game. You didn't hire a replacement, you gave a stranger with no soul the keys to everything you own. And all of this is true. The tokens are super expensive now, and when the tokens run out, you don't really have someone to manage them if you're just trying to replace them. And even if you do, you're going to actually have to pay for someone who understands AI, which is not only expensive, but you also got the token costs on top of that. And then of course, to do the job, the AI agent needs access, like full access. And there are so many cool use cases, but because you can't give an AI full access to your systems, those use cases, I do wonder if they will ever truly be fulfilled. Now, I think companies need to take a look at what IBM has done because they're the only company that has seemingly been doing this well. And they never tried to replace humans with AI in the first place. So, IBM, they built two internal AI tools called Ask HR and Ask IT. Ask HR handles 94% of routine HR inquiries, and Ask IT cut IT service interactions by 70%. And here's the main thing people are missing. They didn't lay anyone off. They redeployed the savings into hiring more engineers and sales people. Head count went up, and the CEO said that AI augments humans, it does not replace them. And IBM is essentially the rarest company right now because a company that used AI to win and didn't pretend that humans were the problem. And I think that's what the companies need to start realizing. Whilst it sounds good to just replace humans with AI and just maximum profits all the way to the moon, that doesn't work in reality. AI simply isn't there yet, and using humans in an augmented fashion is going to be 10 times better than just replacing them. Here you can see it says, "Bradford said that while AI is better at tasks, humans are still better equipped to be strategic partners to the business and help unlock people's true potential." And that is the real lesson from every failure in this video in one sentence. AI can replace a task, but it cannot replace a job. A job is a bundle of dozens of tasks, and only a few of them are automatable. The barista is not just counting inventory, she's reading customers, calming complaints, training new hires, fixing the espresso machine, and noticing when something is wrong. The customer service rep is not just answering questions, she's reading tone, building trust, and handling the case the scripts never anticipated. Companies that confused the tasks with the jobs are now the ones writing the rehire emails, and companies that understood the actual difference are the ones actually winning. So, now, when you think about if you're trying to use AI for yourself, what you need to be able to do is know AI a lot, like you need to verify everything. You can't just trust the flashy demos when they come on stage and say, "Hey, this new AI tool can do XYZ." Don't trust that at all. You need to be able to test it and verify it, okay? You need to run the pilot in your worst examples, not your best ones. You need to measure the time it takes for you to fix those mistakes, and then you need to add the time of the cost of the AI. And if those numbers still work, only then can you deploy it, okay? And if it doesn't, then just walk away. You're simply wasting time. And most of the companies in this story, they did not do that. They just bought into the demo and the hype, and then they got the disaster. And so, now, we're in this essentially cleanup phase of the AI replacement era. Boards are quietly re-asking CFOs how much AI inventory they have on the books that are not earning any earning anything. CEOs are firing the AI consultants who promised them the replacement Utopia, and HR teams are rehiring the workers they were told they would never need again. And Gartner has predicted that more than 40% of this stuff is going to be canceled by 2027. So, the replacement story for now seems to be over. And now, here's an interesting thing that nobody had considered. Now, I was browsing Twitter and I came across this, which is essentially breaking news because it was only a few hours ago, and it talks about the fact that Senator Elizabeth Warren just urged to tax the AI to give free services to people. So, essentially, what she's saying in this video, which I'll show you now, is that the wealth creation from AI is going to be so crazy that we need to actually impose some kind of AI tax so that those displaced by it actually can still benefit. I mean, it's pretty crazy when you think about it. Now, when you're going to be deploying it in a company, is there going to be an AI tax when you're using those goods and services? It's going to be really interesting for the future. >> Only scary. Tech execs are warning that AI could lead to a level of wealth concentration that will break society and create a permanent underclass. Those are their exact words. I refuse to accept that. There is no doubt that we need to regulate AI and consider bigger and bolder options to rein in the technology. But understand this, if we're going to build an AI future that works for everyone, then we need to tax AI and invest in people. Taxing AI raises the money we need to deliver universal health care. So, if millions of workers get fired because of AI, those workers don't go bankrupt just from a visit to the doctor. And if AI transforms the future of work, then taxing AI means we'll have the resources to invest in things like free college and apprenticeships programs and a jobs guarantee so that all Americans can have good paying work. That's what taxing AI promises. And here's what it could look like. Right now, companies pay taxes on their workers and get tax breaks for investing in AI. Woah, it's time to make corporations pay their fair share and make sure they're no longer incentivized to fire workers and replace them with AI.