Anthropic Claude Mythos: надпотужна ШІ-модель, яку тримають у секреті через ризики кібербезпеці

I just got back from the Human X event out in San Francisco last night. It was an event totally focused on the people and the companies building in the AI space. I met a lot of amazing people, but while I was gone, a ton of AI news happened. And well, this is your weekly deep dive into everything that you need to know that happened in the world of AI from the past week. I'm not going to waste your time, so let's dive right in. Let's start with Claude Mythos and Project Glass Wing. This is the story that literally everybody in the AI space is talking about. I've seen probably six videos now of people claiming that they're having meltdowns as a result of this model and the implications of it. So, let's break it down real quick. First, we'll talk about what Mythos actually is. So, Mythos is a model that Anthropic built that is supposedly the most powerful AI model anybody's ever seen. In fact, here it is in Anthropic's own words. Cloud mythos is a generalpurpose unreleased frontier model that reveals a stark fact. AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities. Mythos Preview has already found thousands of high severity vulnerabilities, including some in every major operating system and web browser. Given the rate of AI progress, it will not be long before such capabilities proliferate potentially beyond actors who are committed to deploying them safely. The fallout for economies, public safety and national security could be severe. If we scroll down the article a little bit here, we can see some benchmarks. So, cyber security vulnerability reproduction. We can see the previous Opus 4.6 model got at 66.6%. This new model quite a bit bigger, 83.1%. When looking at the various software engineering benchmarks, Opus 4.6, previously pretty much the state-of-the-art GPT 5.4 and Opus 4.6 were pretty comparable. Mythos preview scored 24 percentage points higher than Opus 4.6 at Swebench Pro. On Terminal Bench, it scored 17 percentage points more than Opus 4.6. On Swebench multimodal, I mean double pretty much. So based on the benchmarks, this is the best coding model the world has ever seen. Essentially, they put out a 245page system card for Claude Mythos here. And you don't have to scroll too far down this paper to see it has demonstrated powerful cyber security skills which can be used for both defensive purposes and offensive purposes. Designing sophisticated ways to exploit vulnerabilities. It is largely due to these capabilities that we have made the decision not to release Claude Mythos preview for general availability. So again, the big implications here being that this model is so good at understanding and writing code that it is finding vulnerabilities in software that nobody's found vulnerabilities in for years. Mythos preview found a 27-year-old vulnerability in OpenBSD, which has a reputation as one of the most security hardened operating systems in the world. It also discovered a 16-year-old vulnerability in FFmpeg, which is used by innumerable pieces of software to encode and decode video. The model autonomously found and chained together several vulnerabilities in the Linux kernel, the software that runs most of the world's servers. So again, they're saying that they've developed a model that is so powerful that if they were just to release it on the world, bad actors will be able to use it to essentially hack into like any website and find vulnerabilities and crack any software on the planet. That's kind of the implications here, right? So instead of releasing it to the public, they've decided to take a different approach, which is where this project Glass Wing comes in. Basically, what they've decided to do is give access to a handful of companies and not like just anybody in the company gets access to these, but specifically like the cyber security specialists at these companies. The idea being that well this model will likely eventually be released out into the world, but also models that are just as powerful, maybe even not from Anthropic, will be released out into the world as well. And they want these companies to get out ahead of those vulnerabilities. So, they're giving the companies you see here on the screen access to Mythos and saying, "Go use it to find vulnerabilities in your products, find vulnerabilities in your software, and patch them up quickly because models this powerful and even more powerful than this one are not only coming, but they're actually here." They also released this video along with this announcement. And this is the part I found the most interesting. There's a kind of accelerating exponential, but along that exponential, there are points of significance. Claude mythos preview is a particularly big jump along that point. We haven't trained it specifically to be good at cyber. We trained it to be good at code, but as a side effect of being good at code, it's also good at cyber. So because these companies are just trying to make their models better and better and better at coding skills, they're also as a consequence getting better and better at cracking cyber security vulnerabilities as well. Now this isn't the first time we've seen this sort of narrative of this model is too powerful, we cannot release it to the public. In fact, here's an article I recently came across last week. The nonprofit research group OpenAI revealed that it had developed a new text generation model that can write coherent, versatile pros given a certain subject matter prompt. However, the organization said it would not be releasing the full algorithm due to safety and security concerns. Instead, OpenAI decided to release a much smaller version of the model and withhold the data sets and training codes that were used to develop it. If your knowledge of the model called GPT2 came solely onto headlines from the resulting news coverage, you might think that OpenAI had built a weaponsgrade chatbot. Here's some of those headlines. Elon Musk founded OpenAI builds artificial intelligence so powerful it must be kept locked up for the good of humanity. Here's another one. Muskbacked AI group. Our text generator is so good it's scary. And here's the best one. AI can write just like me. Brace for the robot apocalypse. That original article that I mentioned from Slate was released on February 22nd, 2019 about GPT2. Fast forward to 2022, a mere four years ago. Google engineer put on leave after saying AI chatbot has become sentient. The point being here is that some of these headlines are starting to feel a little bit like the boy who cried wolf. No pun intended. There is some benefit to these companies to going out and saying, "We have the most powerful model the world has ever seen. It is so scary. we need to lock it up. Nobody can have their hands on it because it is that powerful. Because well, that helps them raise capital. That sort of postures them as the company building the smartest, best thing ever in the world of AI. And it gets people pretty hyped up for when they finally can get their hands on a model like this. Like they're building pent-up demand. People are like, "This model's that strong. I want to play with it." Saying all of that, I do actually think this time it's kind of warranted. I mean, it was kind of warranted in the past as well. I mean, the the reasons they were saying GPT was too powerful was they were saying it could be used to flood the internet with fake information and propaganda and basically slop. And well, it kind of did. Fast forward to today, now they're saying, look, if this model gets released into the world, it is so good that hackers and bad actors are going to have a field day with this. and no products are safe. So, I actually, if I'm being totally honest, I do think there's a little bit of a marketing play here, but I don't actually think that's their intention. I do think, and this is just my gut feel on it from everything I've read and everything I've heard from others, is that Anthropic is legitimately scared to release this into the world, and they are doing the thing that they feel is the most responsible approach. They're going to all of the biggest tech companies in the world, not named OpenAI, and saying, "Go use our software. find the vulnerabilities before models that are this good get released on the world and get them fixed because almost everybody on the planet uses tools that have at least one of these companies behind the scenes. So, it is pretty damn important that these companies get locked down before new vulnerabilities pop up. So, while I personally want to get my hands on the latest and greatest AI models whenever possible, I actually do appreciate the decision that Enthropic made here. I know there's going to be a lot of people in the comments who disagree with me and just think that this is pure 100% marketing hype and I do again think there is an element of that but hell I want to know that my Apple products and my Microsoft products and my products that are powered by Nvidia and Cisco and Crowd Strike and the various security layers behind the scenes are locked down before stuff like this gets released on the world. I actually think that's kind of a good idea personally. All right, I've already said more about this than I was planning on saying. There are a million videos about this Glass Wing and Mythos thing. I even made previous news videos where I broke down Mythos back when it was leaked the first time. So, this isn't even the first time I'm talking about this. But, if you want to dive deeper on Mythos, there are a lot of channels talking about it right now. You'll have no shortage of options to go dive even deeper on this. We also got a couple major large language model releases this week that we actually do get to use. There's actually two that I'm going to talk about. The first one from Meta called Muse Spark is the one that has the most people talking. The one I'm going to tell you about in a minute is the one that I actually think is more impressive, but nobody's talking about it. But let's talk about Spark first. And as a quick side tangent, it seems like all of these model makers and companies just keep on copying each other's naming conventions. Like OpenAI has a model called Codex Spark. My mini supercomput that's sitting in my corner running my openclaw bot right now is called a DGX Spark. I don't know. It's just interesting to me that like these names all kind of keep repeating themselves. Anyway, sorry, side tangent. So, this new model called Muse Spark is actually Meta's first model in quite some time and it's the first model out of Meta Super Intelligence Labs. So, if you don't remember, Meta's AI used to be run by Yan Lun. They brought in Alexander Wang from scale and a whole bunch of high-profile engineers that they sort of plucked and poached from all the other AI companies to build this super intelligence labs. Since that whole team came together and Yan Lun left Meta, we haven't really seen a lot out of Meta. And this new model seems to be a pretty big leap over their previous llama models. Now, their previous Llama models, they were making a big deal over the idea that they're open source and that they believe that these models should all be open source. And well, this one is not open source. We've got some benchmarks here for this new model, comparing it to Opus 4.6, Gemini 3.1 Pro, GPT 5.4, and Grock 4.2 Reasoning. Pretty much the state-of-the-art models from all of the Frontier Labs. When it comes to figure understanding, it beats all of these models by quite a bit. I mean, GPT 5.4 is pretty close, but it still beats that out for multimodal understanding. It's, you know, middle of the pack for the most part. Looking through a lot of these benchmarks, it's like a middle of the road, right? Like I don't think anybody's going to replace it for coding. When we look at the coding benchmarks down here on SWEBench, it kind of falls below both Opus and Gemini 3.1. and is pretty much on par with Grock 4.2. On SWEBench Pro, pretty much falls slightly below all of them outside of Grock 4.2. On Terminal Bench, it pretty much lower than all of them outside of Grock 4.2. It doesn't appear like it's going to be anybody's new coding model. When it comes to health related stuff, it's the new best model for healthbench hard open-ended health queries, but for everything else, it kind of falls around middle of the pack. So, it's a good model. is not really the best at anything, but it is almost caught up to the, you know, models that truly are the best right now. And for the first release, as a company that's only existed for like maybe a year now, pretty impressive how quickly they got to these sort of benchmarks on their first release as this new super intelligence labs. artificial analysis here has a good post over on X that does a pretty decent job of sort of putting this into perspective. So they have the artificial analysis intelligence index which incorporates 10 different evaluations across you know multiple different fields and we can see that Meta's last model Llama for Maverick scored way down here. here. It was pretty much the last one on this chart. And with their new model, we can see it leaped all the way into fourth place as far as the artificial intelligence index. Hasn't quite caught up with Gemini 3.1, GPD 5.4, or Claude Opus 4.6, but it is closely falling in right behind those three big players. It's also very, very token efficient, which means it theoretically should be a lot less expensive to run this model. On this chart, smaller is better. We can see this one falling all the way down here at the bottom of the graph. And when you compare it to those same big models that we were just looking at, GPT 5.4 and Claude 4.6, well, those are the two least token efficient. If you want to play around with this new model yourself, you can play around over at Meta AI. They say they're opening a private API preview soon, and I'm sure a wider API rollout will be coming soon after that. But it sounds like it could be a fairly inexpensive to use model that is up there with some of the state-of-the-art models in performance, albeit not quite the best at really anything, but good at most things. And then we have that other large language model that I mentioned that's kind of more exciting in my opinion, and that's GLM 5.1 out of ZAI. Now, what makes this one exciting to me is that GLM 5.1 is an open-source model under the MIT license. You can actually grab the model weights over on HuggingFace right now. But when we look at the actual benchmarks on this, it's beating out most of the state-of-the-art models. I mean, for SWEBench Pro Software Engineering, it's got a 58.4 compared to GPT 5.4's 57.7 and Opus 4.6 57.3 and Gemini 3.1's 54.2. too. We have a model that's getting basically state-of-the-art coding performance in an opensource openweight model. Now, it doesn't score quite as high as GPT 5.4 in realworld terminal tasks or agentic coding, but it's in second place when it comes to things like math. I mean, it's pretty on par with the state-of-the-art. Not quite as good as Gemini 3.1 and GPT 5.4, but pretty up there. And just across the board, this model is a really really impressive model for being an openweight model. Like you can download this model, you can fine-tune it for your own use cases. You could run it locally and theoretically it could write code about as good as GPT 5.4 and Opus 4.6, which is just mind-blowing to me that open models have sort of caught up already. And also kind of crazy to me that not more people have been talking about this model specifically. Now, this is something I've said in the past, and I'll say it again. I have a really hard time demoing and testing new large language models because they kind of do most of what I need them to do. They write decent code for me. They help me brainstorm ideas. They run my open claw agents for me. They pretty much do the things I need them to do. I don't have a real use case for really complex math or really complex science or things like that. And so I struggle to go and say like here's proof that this one is better through giving it text prompts and proving that I want to work with somebody to like build a really cool benchmark or some tests that we can do that are sort of like the average Joe tests of like here's how I prove that this is better for just like everyday people using these models. I just don't know what those benchmarks look like yet. I actually put a post on X saying I'm looking for somebody to help me with that benchmark. So, if you're somebody that wants to help work with me to create some sort of really cool test that we can build together to compare new models as they come out, hit me up over on X and let's talk because I really want to find some cool new ways to test and compare new models as they come out. But I'm not necessarily interested in testing in the most complex math problems or the most complex science problems or things like that. I want more of the like average Joe everyday use cases that we can benchmark against if that makes sense. Anyway, hit me up if you're interested in exploring that with me. We also got a couple updates out of Google this week. They're not new models, but they're new sort of updates that you can now do inside of Gemini. The Gemini app can now generate interactive simulations and models. So, very similar to what we saw OpenAI and Anthropic both release a few weeks ago. Gemini now has similar features where you can create these interactive sort of visuals and experiments and it'll generate, you know, sliders and things that you can adjust to get the visuals to change in real time for you. And again, looks very similar to what we saw out of both OpenAI and Enthropic. You can do these kinds of things in those models already. To do this, head on over to Gemini. Make sure you're using the pro model and then you say something like, "Help me visualize the threebody problem." All right, it gives me a quick explanation and let's click show me the visualization. And we can see it created this sort of complex visual with all sorts of sliders and presets. Figure 8 orbit, chaotic dance, dominant sun. Let's do figure 8 orbit. It just like immediately goes off. It doesn't really do much of a visualization. Let's drag some sliders around and see if we can get it to do something. And yeah, it's not quite working right for this one. Let's just do the tried andrue. Help me visualize compound interest over multiple time frames. Show me the visualization. And we've got a visualizer here. Let's see. Starting principal, $1,000, 5% annual rate. Over 10 years, we would have 1,629, but over 20 years we would have 2,653. If our annual rate was 10%, we'd have 6,727. If our principal started at 10,000, we'd have 67,275. And we can obviously see our chart is changing as we adjust these various options down here. And we could change to compounding quarterly. And that changes our number. And if it was at 30% annual rate, starting with 10 grand, in 20 years we'd have $3.2 million. Could somebody help me find uh something that gets a 30% annual interest rate? That'd be great. Thank you. Another new feature inside of Gemini is the notebooks feature. And not to be confused with notebook LM. Notebooks are essentially what projects are inside of both Claude and ChatgPT. They give you a dedicated space to organize your chats and files. And because they sync with Notebook LM, you can unlock even more efficient workflows directly from Gemini. With notebooks, you can keep conversations about a topic organized in one place. You can move past chats into notebooks, give Gemini custom instructions, and add relevant files like documents and PDFs to give Gemini more context. Basically the same idea as projects in those other platforms. As of right now, it's just rolling out to paid users. So, if you're on an AI Ultra Pro or Plus plan, you should be getting access to it. And I imagine it'll be rolled out on the free plans in the near future as well. But let's jump over to Gemini here. This is one that I've had early access to for a little bit. But if I open my sidebar, I could click on new notebook. Give our notebook a name. I'll just give this one example notebook. And then once in here, I can come and click on these three dots. Click on notebook settings. Give it its own set of custom instructions. Have it use its own notebook memory. So it will remember past chats you had within this one notebook. Great for if you're using it for like a journal or you're doing research on a very specific topic and you want all of your chats to reference your past chats, but you don't want it to reference chats outside of the notebook. This is great for that. So, you can really sort of dial in how you want this notebook to work here. And you can add sources by adding files, adding from Google Drive, adding URLs, and copying and pasting text in. And when you create a notebook here, you can actually push it over to Notebook LM and do all the cool Notebook LM stuff with it, you know, podcasts and videos and mind maps and all of that kind of stuff. Pretty cool new quality of life feature. I'm actually quite surprised it took Google this long to roll this feature out if I'm being honest. There's an AI video model that people have been raving about for like months now called Seed Dance, but nobody in the US has had access to it until this week. It's finally starting to roll out in the US. You can actually use it inside of the Runway app as of this week as well as inside of Bite Dance own Cap Cut app. Now, I'm going to test it inside of Runway since I have a Runway account, but from what I understand, a lot of the features that were sort of making this go viral early on have been nerfed at this point. Like, you can't generate trademarked IP or actual videos of like celebrities and things like that. It is still a really, really, really good model. It's just the things that caused it to go viral for the past month have all been sort of stripped away. So, if I log into my runway account, down at the bottom where it lets me select my model, I can swap this to seed dance 2.0. Now, the timing probably couldn't be better. We just lost Sora. And so, this one comes along like 2 weeks later and is kind of the new best model for the most part. There's another one I'll talk about in the rapid fire section here in a minute, but this is the best model we have access to right now. I'm going to give it a pretty detailed prompt here with multiple scenes in it to see if it'll switch between the scenes properly. And we'll go ahead and generate. I'm putting the actual prompt on the screen right now if you feel inclined to pause and actually read the whole prompt. But let's see what it generated. Okay, that's actually pretty impressive, honestly. Like, it actually generated pretty fast as well. I was surprised how quickly it generated. It's a lot faster than Clling 3.0 is right now. So, a really, really great video model. Get in, experiment with it. It's available inside of both the Cap Cut app and inside of Runway right now. So cool new video model to go play around with. The company Hey Genen had some news this week. They launched their new Avatar 5 model that captures your identity in just 15 seconds. And well, if you couldn't guess, you just saw exactly what it looked like. That was a video that I generated with my new Hey Gen 5 avatar. Now, it did only have me record 15 seconds of video to get that clip. Pretty impressive that you only need to record 15 seconds and it gets that close. I think the voice is still a little bit off and the lip-s syncing is still a little bit off, but we've definitely come a long way with cloning. There's some other cool features inside of this hen platform. Like if I go to avatar here and then click on avatars again off of my 15-second recording that I did, it generated a bunch of different potential visuals that I can use. Like I can take this version of myself sitting in this white room wearing a red shirt and use this in a video instead. And then I just type whatever I want up here. Click generate and about a minute later I have this. If you haven't already, make sure to subscribe to Matt Wolf. Help me get to 1 million subscribers. I'm so close. They also have an editor called AI Studio that gives you the option to remove the background, which would be cool if I was able to like edit that a little bit better. As you can see, it cut off my headphone and I have no way of bringing it back. But if I wasn't wearing headphones, it would probably work pretty well. All right, I've covered so much already and this video is going way longer than I had anticipated. I spent more time talking about mythos than I planned, but there's still quite a bit more I want to talk about. But let me share the rest of this stuff in a rapid fire and we'll run through it real quick. Starting with some news out of OpenAI. They just opened up a new plan. Before they had the $20 a month plan and a $200 a month plan. Now they just added one that's kind of more in the middle at $100 a month. We're introducing a new $100 a month pro tier. The new tier offers five times more codec usage than plus and is best for longer higheffort codec sessions. The new pro tier still offers access to all pro features including the exclusive pro model and unlimited access to instant and thinking models. They've also increased codec usage for a limited time through May 31st. So, if you're on this $100 plan, you get up to 10x usage of ChatGpt Plus on Codeex. Personally, I've been finding myself playing around with codecs more and more. So, look for more videos of me showing off codecs in the future. Most likely, Enthropic launched a new Claude managed agents feature this week. Now, you can already sort of build agents. Claude co-workers is kind of an agent itself. Claude Code is kind of an agent, but this one seems to be like to tie your agent to tools like, you know, notion and ClickUp and things like that. I haven't actually played with this a ton, so I can't speak too intelligently about it yet, but if you go to the Claude console over at platform.claude.com, there's this new managed agents feature over here. You can click on quick start and they have a bunch of agent templates already pre-built and you can see that they're designed to like tie into existing products, Notion, Slack, Intercom, Asauna, etc. You can also apparently just sort of describe what you want your agent to accomplish and it will build out your agent through prompting. So the types of examples I've seen are like in notion you can have like a canban board and when you select items from one section of the cananband board and drag them to another, it actually fires off a bunch of agentic tasks because you drag it from one to the other. So again, it really seems like it's designed to tie like an agent to things like notion and a sauna and like task management software where when you move things around within the task management software, things happen with your agent behind the scenes. Now, if you really want to go down the agent rabbit hole and build agentic workflows for tasks that you do on a daily basis, I still think something like Claude Co-work or OpenClaw or the new Hermes or Hermes or whatever it's called are probably better options to look into. And since we're speaking about Anthropic and agents, this was something that popped up this week that kind of made a lot of people frustrated. Boris Churnney here, who heads up Claude Code for Anthropic, they made the announcement this week that starting on April 4th at noon Pacific time, Claude subscriptions will no longer cover usage on third-party tools like OpenClaw. You can still use an Enthropic API key in your OpenClaw. you just can't use your like $200 a month claude max plan to power your open claw. And a lot of people were very very frustrated by this. Now I understand why they did it because things like openclaw burn through a lot of tokens really really fast and it was probably costing Anthropic more than the $200 that people were spending on the Claude Max plan. But this was definitely something that a lot of people were frustrated by who had built like a whole open claw system around anthropic models like Opus and now all of a sudden these agents that they'd built and been working with for a while were either a no longer going to work or b going to get a hell of a lot more expensive. Personally, my openclaw is running using a local model. I'm using the Neotron model locally and then when it does have to do more complex tasks, I'm having it use the OpenAI API because they are still letting you use your subscription to cover the use and things like OpenC cloth. Perplexity announced this week that you can now connect your financial institution data to Perplexity using Plaid. With Perplexity and Plaid, users can consolidate their financial information into a single secure, highly personalized hub. The Plaid integration provides readonly access and user data never touches perplexity services. So you can do things like link credit cards to track spending by category with full detail of the transactions, connect mortgages, auto loans, student loans, monitor the balances and payment history, combine bank investment and other accounts for a complete view of net worth. Here's an example of like a dashboard that was built that almost has like a GitHub look to it. The company Factory AI released a desktop app. Factory in the past was primarily used through the command line interface and now they have a simple desktop app where you can kick off agents or droids to go and do tasks directly from within the desktop app. Definitely something I'm excited to play around with a little bit more. Full disclosure, I do have a small investment in factory AI. Cursor rolled out a new quality of life update this week. You can now run Cursor on any machine and control it from anywhere. So you can actually kick off agents from your phone to run on your dev box. Very, very cool update. I use cursor a lot still myself as well. A handy little feature to be able to prompt it from my phone and get it to go do stuff on my computer. XAI released some new features this week in their photo platform. You can now edit images with text. So, it generates an image, you give it a text prompt to edit the image, it will make those edits. You can add blurs and redact parts of photos. And it's currently available on iOS and coming to Android soon. Jumping back to OpenAI for a second. Apparently, rumor has it that we've seen a leak of GPT image 2, which means we're probably getting a roll out of their new image models soon. Over in Arena AI, there was some new image models called masking tape alpha, gaffer tape alpha, and packing tape alpha. And here's some of the images it generated. According to Peter Levels here, it appears to be really, really solid for like infographic type images. Here's a map that it generated a storefront looking through a door, which I mean looks really realistic. I think is this a reflection up in the top off the glass here? And it even made like a mockup of the YouTube homepage here with a whole bunch of different thumbnails on it. Now, I can't confirm 100% that this is actually GPT Image 2, but if it is, we can probably expect it to be rolling out fairly soon. Now, earlier I sort of teased another video model that I'm going to talk about. Well, there was some mysterious model that dropped on one of the leaderboard sites called Happy Horse 1.0. And according to the leaderboard website, it's the number one ranking model above the new Seed Dance 2.0, above Clling, pretty much above everything. And if we look at some of the examples that are being shared here, it looks pretty good. I mean, it looks like a pretty realistic video. Here's another one. They look like stock video to me, but it looks very, very realistic. Now, I don't know if this is 100% confirmed or not, but according to the information, this new model is actually out of Alibaba. The Happy Horse 1.0 model has topped an AI model leaderboard, beating Bite Danc's Seed Dance 2.0. Alibaba Group has anonymously released a new AI model generation called Happy Horse 1.0, which rose to the top of the AI model leaderboard. Again, I don't know how they have actual confirmation that this came out of Alibaba, but I mean, it wouldn't surprise me. Google actually released an app on iOS this week that's a dictation app that works completely offline. It's called Google AI Edge and it works similar to like Whisper Flow, Super Whisper, and Willow. And it uses their new Gemma model to basically listen to your speech and transcribe it to text automatically for you fully on device. So you could be completely offline and it will do this for you still. And finally, Spotify rolled out a new feature for their AI generated playlist before you can tell it, "Hey, generate me a playlist of songs to get me in the mood for a workout that are poppy and only have female vocalists." And it would have generated you an AI music playlist for that. Well, now they've released the podcast version. You can go and tell it, "Hey, I want podcasts about business and how AI is impacting business. Go find me a bunch of podcast episodes." a feature I will definitely be using because I'm a fan of both Spotify and podcasts and am constantly looking for new podcasts to listen to. The biggest problem with podcasts in my opinion is discovering interesting podcasts from shows you've never heard of before and this sounds like a great sort of discovery mechanism for that. So really really excited about this feature as both a podcaster and a podcast listener myself. So yeah, it's been a busy week and most of this happened while I was at Human X. So there's a few little tools and things I haven't had a chance to play with, but if any of the things that I kind of briefly brushed over turned out to be much bigger or exciting than what I made them out to be, I'll probably make a follow-up video and talk about them a little bit more. But these weeks are only getting crazier and crazier. I don't anticipate it slowing down. on. We might have some up weeks and some down weeks, but going forward, we're just going to see more and more. And the acceleration is going to keep on picking up. And that is the reason I've sort of slowed down on trying to make videos about every single topic that comes out because it can be very, very overwhelming. It's easy to feel like you're falling behind and not keeping up with the new tools and news that's coming out. And if you're trying to keep up with it on a daily basis, well, you're probably going to get burned out. which is why I decided to put most of my focus on making these Friday roundup videos where I sort of look through all 300 pieces of news that came out in one week in the AI world and narrow it down to just the things I think you would be most interested in. I'm trying to filter through and find that signal from the noise and just bring you the most important, most interesting, and probably most useful stuff to the most amount of people. Now, I will make videos in between with some tutorials and some commentary and stuff, but when it comes to news, I want to try to minimize the overwhelm. I'll drink from the fire hose every day and then give you more of a trickle every Friday and try to help you stay looped in with just one video a week on all the AI news you need to know and do my best to keep you up from feeling overwhelmed. That's my goal. If that's something that interests you, give this video a like and consider subscribing to this channel. I will make sure more of that kind of stuff shows up in your YouTube feed and I'll do my part to keep you informed with ideally not making you feel overwhelmed. I'm getting closer and closer every day to that million subscriber milestone and I'm hoping to hit it this year. So, I'd really really appreciate it if you would click that subscribe button. It really does help me out and you know, it's just a milestone I've been shooting for for a long time and I'm like this close. So, really really appreciate it. Again, I record these on Thursdays, publish them on Friday. So, if any news came out late Thursday evening or on Friday, it will likely make next week's video. But again, that's what I got for you. Thank you so so much for tuning in. I have so much fun keeping my finger on the pulse of all of this stuff. It's literally become my full-time job to just stay looped in on the news, and I absolutely am having a blast trying to stay on top of it and try all these new tools so I can tell you what's worth your time and what's not. Hopefully you find this helpful and uh yeah, again, that's what I got for you. Thanks so much for hanging out with me and nerding out with me today. Hopefully. I'll see you in the next one. Bye-bye. What is up everybody?