AI News: A Flood of New Models (Here's What Matters)

There was legitimately a ton of news in the AI world this week. We got tons of new models, and I'm going to break down what I think matters. So let's get right into it. I'm changing things up slightly this week, and instead of talking about what was probably the biggest news of the week, I'm going to talk about what I found to be the most interesting news and the coolest news this week. For me, probably the coolest thing that I saw this week Is this new WorldCLAW model from Tencent. Now, unfortunately, this isn't a model that we have access to actually play with yet, but I just found it really, really cool, so I wanted to share it first. It's a tool where you can actually prompt an entire 3D world into existence, and every little individual element of this 3D world here is actually an editable individual asset. So for example, generate a snow-lined village. It generated the entire environment, but it also generated every little individual asset, the trees, the houses, The carts, the little lamp posts—everything you see in this was generated separately as individual assets. Here's another one: a desert battle scene. You've got the desert here, and you've got all these assets. There's a depth map, there's a normal map, there's an instance map, and every single one of these is a separated asset, which I just think is super, super cool. There are a ton of different examples here on their website of different worlds with the generated assets. Like this one has dragons generated inside the world here, and I don't know, I just find this really impressive. This reminds me of some like, you know, Diablo vibes or something. And these aren't like 3D Gaussian splats or Nerfs or anything like that. In fact, if we take a quick look at the research paper— don't worry, I'm not gonna get too into the weeds here— they actually have a terrain planning agent. And the tool planning agent can actually do a web search to get more details about that specific terrain. When the scene specification contains concepts that require external knowledge, the agent invokes a search tool to retrieve relevant references. And then to actually generate the terrain, it uses GPT Image 2. So the same image model you get when you prompt an image inside of ChatGPT. It also employs the model out of Meta called SAM-3, which takes an image and segments out every element of that image. So if, let's say, GPT Image 2 generates an image with a whole bunch of different assets in it, it will then use SAM-3 to segment those all out. What this Tencent Hunyan model is most known for is actually taking 2D images and converting them to 3D images. So it will generate the images with GPT Image 2, use the Segment Anything model to break out everything from within the images and then use the Hanyan technology, which converts 2D images into 3D images, to create all of the 3D assets that go into the world. I find it really, really cool. I feel like research like this has huge implication in like the gaming world, obviously, for building things like levels. But I can also see this being used for like creating little 3D animations, you know, like the worlds that people use to make their own version of like a little, you know, Toy Story animation or something like that. I also believe this could be used to develop virtual environments that you can train robots in so that robots could see every possible scenario and train in virtual environments before being put into the real world. That's kind of how they're already starting to train robots now. This seems like another pathway into that. Now again, unfortunately, this isn't something that we have access to. They do link to their GitHub here, but when we pull open GitHub, there's really nothing here yet. But it's actually something I'm pretty excited to get my hands on. All right, now let's move on to the thing that probably has the most people talking about it this week, and that's the new Grok bot out of Space XAI. Now interestingly, XAI or Space XAI, I'm probably just going to call them XAI to keep it easy. They recently acquired Cursor, and this is a bot that has agentic use cases similar to what you can do in Cursor, but it's not Cursor. It's almost like a competing product to Cursor, where Cursor is probably a little bit more designed to be used specifically by coders. This new Grok bot seems to be for people that want to use agents like an OpenClaw or a Hermes or that type of thing without needing to think about or or touch or look at code. Probably closer to like a Claude Cowork or Codex Work competitor than it is to a Cursor competitor, I guess. So I'm gonna go ahead and download it for macOS. Now, once we're inside GrokBot, you can basically create little agents that are designed to do different things or have different focuses. They even give you some suggestions down at the bottom: inbox triage, a chief of staff, a negotiator, a prototyper, researcher, shopper, apartment scout, et cetera. The chief of staff, it even suggests you use this one to manage other other bots. Let's create a research bot. You could name them whatever you want. I like to go pretty literal with my names. So it knows it's named research bot and it says, I'm research bot. I dig into things, pull sources, and come back with clear takes so you don't have to wade through it all yourself. What's the first thing you want me on? A topic, company, paper, et cetera. So just by naming it, it knew its role. So I'll just say, I want you to keep on top of the latest AI news for me and tell me what I need to know about it and why it's important. And while it's doing that, let's go ahead and create a new bot. I'll click create new. I'll go ahead and name this one email triage. While it's thinking about its next steps, let's see what Research Bot says. All right, weekday mornings at 8:57 AM. I'll send a tight brief. What moved, why it matters, nothing extra. First one is coming as soon as I've checked today's news. So it's looking it up for me right now. The Email Triage noticed I named it Email Triage and is checking to see if it has access to my email. It noticed it doesn't and asked me if I wanna connect my Gmail. We can go ahead and authorize. And now it says Gmail's connected, pulling your inbox now. Let's create a Chief of Staff. It's gonna know its role just by how I titled it. While it's doing that, I can jump back to Research Bot. See what news it found for me. We got news out of Nvidia, Gemini 3.7 Flash, Grok 4.6, news out of Anthropic. So it found a bunch of news for me, and now it knows at 8:57 AM to send me new news. Jumping back to my email triage here, it did look through my emails. I'm going to blur it because it's obviously some sensitive stuff here. And I can have it, you know, draft an email to one of my messages, sweep the pitches from today, set a morning digest, things like that. And our chief of staff is set up up here. What's the main thing you want me Calendar and priorities, email, running the other bots, or something else entirely. I'll just say running the other bots. So let's take a little bit more of a look around Grokbot here. If I click on plugins, you can see all the different things that we can connect it to. So Gmail, Google Drive, Google Calendar, Granola, Notion, Slack, various agent orchestration tools, customer support tools, data analytics. So tons and tons of tools that you can connect this to and give it access to. Now another interesting thing that Grokbot does is each of these agents has its own little like computer that it has access to. We can actually see what our agent's computer sees right now. It's got access to Chrome, a file manager, and the terminal, and we can actually teach it a task. So if you have something you wanted to do on a recurring basis, you could come up here, click Teach a Task. It'll start recording the screen, and then you can just walk through the task that you want it to do. Once you're done doing the task that you need it to remember, you hit Stop, and then it learns from the demonstration and creates a new skill to do the thing you just did whenever you ask it to. They've also got what's called routines, which is basically like cron jobs or automations or recurring tasks. I can hit plus on here, give the routine a name, give it instructions, and then tell it when to run the routine. So it can run on a schedule like every hour, every day, every month, you know, whatever schedule you want, but also based on other events happening. So if it's connected to Slack, anytime there's a new Slack message, it could trigger a prompt. Every time there's a new Git event, it can trigger a prompt, things like that. So you can set these routines that happen on an automated basis. So again, this is for people that just want it to go and do stuff for you. You give it the tasks you want it to do on a regular basis. You show it how to do the tasks. You schedule the tasks up so they happen on autopilot, but maybe you don't want to have to think about code. You don't really need to see all of the thinking process behind the scenes. You just want to create a bunch of sort of automated agents that do things for you. That pretty much seems like the main use case for this as of right now. They do also have a mobile app. So once you create your agents, you can manage these agents from your phone. And they can continue to do things on the computer because again, they have their own computer. The computer is in the cloud. It's not actually doing things on your computer. It's doing them on its cloud computer. So that's the new GrokBot out of xAI. Another big update comes out of Artlist. Now, if you've seen the new SeedDance 2.5 model that just dropped a couple weeks ago, it's an insanely impressive model. And Artlist just announced that SeedDance 2.5 is now available directly on their platform. Now, what makes this model really cool is that it can generate videos up to 30 seconds in just one go. So you can build out a whole scene without needing to stitch together a bunch of tiny clips. And you can also use up to 50 reference images for, say, a character or product or location or visual style, and it'll stay consistent from shot to shot. So if you want to try it out for yourself, you can access Sea Dance 2.5 right inside Artlist alongside the rest of their creative tools. Just head to the link in my description box to sign up. And thanks so much to Artlist for supporting my channel and sponsoring this portion of today's Now, one of the big issues with AI and AI just being everywhere now is that you really can't tell anymore what's AI-generated and what's not. AI is getting better and better at looking and sounding more real and more human in pretty much all areas— in text, in video, in images, in music— and all of it is getting harder and harder to tell. And now a lot of these companies are trying to solve for that, and a lot of people aren't very happy about it either. So for example, Claude is starting to put little invisible watermarks on its images. Oh, no. Marks into the content that it generates to support transparency. Anthropic is working to include machine-readable marks in content that Claude generates. Claude markings cover output from supported models everywhere you use Claude, including Claude Platform API, Claude, Claude Code, Claude Cowork, and Claude Tag. Embedded watermarks will apply to all generated text. Marking will apply to output from supported models wherever Claude is offered worldwide. So when a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won't see it and it doesn't change the meaning, quality, or readability of Claude's response. Because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere and may persist through some editing. Now, it does go on to say that a detected mark provides a signal that content was processed by Claude, but it's not fully conclusive. They also say that lack of detected mark doesn't mean the content wasn't AI-generated. And again, a lot of people really are not happy about this. They want to be able to write their Substacks and their blog posts and their newsletters with AI and not have anyone notice that it's AI. But also, it It seems like the easiest workaround is just letting AI write the initial draft, and then you go through and edit it and tweak it and sort of reword it yourself, which you probably should be doing anyway. I also think like a flurry of tools are gonna pop up that will figure out how to strip these watermarks anyway. If tools pop up that can detect them, there's definitely gonna be tools that pop up that will help reword things, pass them through the detector to make sure it's no longer detecting the watermark, and it will keep on sort of editing the text until it passes the watermark test. So I don't really think that this is going to work long term, but I am actually kind of happy, believe it or not, that a lot of these companies are trying to make it more and more obvious when stuff is generated by AI. And along similar lines, this is actually news from last week, but it's so relevant to what we're talking about, I'm sharing it this week. And that's that if you generate music with AI, specifically with Suno, well, a lot of the fun there might be ending as well. Suno released this article called How We're Building the Future of Music Responsibly. They do say, while we don't think it's our role to decide what has artistic value, we have an important responsibility to make it harder to abuse AI tools in ways that undermine a healthy music ecosystem. In the coming weeks, we'll be adopting new audio watermarking and fingerprinting technology so we can partner more closely with distribution platforms on combating fraud and misuse. So basically, Suno songs are going to have watermarks in them, and if you upload these Suno songs to YouTube, Spotify, Apple Music, you know, wherever the music's distributed, it's going to signal to those platforms that this song was AI generated, and then those platforms, it's up to them to decide whether they want to pass that information along to the end users. They also announced they'll introduce a new download policy designed to limit the ability to mass distribute songs on streaming platforms. And well, what that actually looks like in practice, if we check out their billing page now, starting on September 3rd, if you're on their less expensive Pro plan, you can download 20 songs per month. So you could generate, you know, whatever your credits will allow, but you're only allowed to download 20 times, and then it won't let you download anymore. If you're on their more expensive $24 a month plan, you can download 60 songs per month, again starting September 3rd. I know a lot of people that generate music with Suno are freaking out about this and canceling their Suno plan, and I understand they're watermarking it and you're— they're limiting how many songs you can download. But I also know that a lot of people are frustrated that they'll listen to a song on Spotify, find out it was AI generated, and not like the fact that they were listening to AI generated songs, you know, for their own ethical reasons or whatever reason you have against listening to AI songs. I feel like people should know. I also understand the counter-argument. Sometimes I do know people that make a lot of YouTube videos and they just want like subtle sound beds under their video, and they'll use Suno to make songs that have no lyrics, that it's just like background music. And if you put out a lot of videos and you use Suno to generate just like subtle background music in those videos, well, this is going to impact you as well because now you're going to be limited to how many downloads you can get. And maybe the rest of the video isn't AI generated, but just that underlying sound is. And is Watermark gonna cause that whole video to be marked as AI? You know, there, there are like some subtleties there. The first time I built the original Librarian game, I wanted to make some background music on the, the game, and I wanted a few different songs to sort of cycle through so it didn't have the same song every time you played the game. I generated lo-fi beats to be the background music of those games using Suno. Now personally, almost every use case I have would probably fall under me downloading 20 songs or less a month, so I don't think it's gonna have a huge impact on me. But I do understand that a lot of people are frustrated, and this is probably going to push them to some of the more open models or to some of the other companies that have song generation that don't have watermarks or download limits. So it is what it is, but Suno is still the best when it comes to generating music. I mean, it's the best quality music generator. I don't know, I got mixed feelings about this one. But along the same lines, Spotify, they've decided they're going to make sure it's very clear when you're listening to AI or not. So right now you can voluntarily share if your profile represents an AI persona through Spotify for Artists. We hope that artists will be transparent parent, so we're giving them the opportunity to disclose, but we won't rely on self-disclosure alone. Spotify will also review artist profiles and identify those whose public identity— the identity an artist presents on their profile, including name and imagery— appears to represent photorealistic AI-generated identities. Starting in mid-September, you'll start to see the AI Persona badge appear on artist profiles. Also, if you are determined to be an AI artist, you will not be included anymore in editorial or algorithmic recommendations, so you won't see any music from AI personas in your personalized recommendations unless you follow those AI personas. So Spotify is really starting to crack down on this. And again, I'm a fan of the idea. Now, if I like a song, I like a song. However, I do want to know if it's AI generated or not, just because I do have a bigger appreciation for music that I know like a band created or individuals put a lot of time into without using AI. Like, I want to know one way or the other if it was AI or not. That doesn't mean I'll never listen to AI generated music. There might be some songs that I I just like, I really dig, but I do want to know. Probably an unpopular opinion among the AI world. We also had like a flurry of new model releases this week. But before I start breaking down all the models that came out, I do want to say something real quick. Yes, I do have 8 different model announcements that I'm going to break down pretty quickly here. But if I'm being totally honest, most of them feel like pretty marginal updates. Every once in a while, a really big model comes out that just blows everybody away. Fable was one of those. I think GPT-5.6 SOL was one of those. Prior to that, like Opus 4.6 was one of those big leaps. Prior to that, it was maybe GPT-4 or Gemini 3 Pro. It feels like lately we just get tons and tons of new models just thrown at us all the time, and most of them are just like teeny tiny marginal improvements that 99% of the world's never really going to notice the difference on. Like maybe a model got slightly cheaper or slightly faster or like one point better on a benchmark. And I feel like most of what we got This week is a lot of that kind of stuff. But I also know a lot of people tune into these videos to make sure they stay fully looped in on all the latest AI news. So I don't wanna skip over 'em. I don't wanna act like all these new models didn't come out this week, but I also don't really super care to go too deep into every single one of them either. They're just not gonna change anybody's life in any real meaningful way. Let's just be totally honest about that. All right, I'm done rambling. Let's check 'em out. So the model I've heard most people talking about this week, the one that feels like the biggest leap out of all of the models that were talked about this week is Grok 4.6. Now, the benchmark I tend to pay attention to the most is this DeepSWE benchmark. This is the one that most people that do any sort of software development will say has the closest correlation to you actually noticing a jump in AI coding. And this follows along to pretty much how I use the models right now. GPT-5.6 SOL is pretty much My favorite coding model, the one I use the most. Fable 5 is probably my second best, the one that I use the second most. Well, this Grok 4.6 kind of is almost as good as those 2 models. Now, when it comes to like knowledge work type stuff, people look at this GDPVal, and Grok 4.6 is the new leader in this one here. Harvey Lab, which looks at sort of like legal use cases, it also is beating everybody in that one as well. So I mean, it's a good model, and the pricing is $2 per million input and $6 per million output, which If you compare that to GPT-5.6 SOL, I mean, it's a huge discount. $5 input to $30 output. So $2 in and $6 out. You're gonna save quite a bit of money using it and get fairly close to the capability. But again, when I'm generating code, I mean, if it's me personally, I'm just gonna jump to the model I like to use the most for coding. And I am on a Pro plan, which does cost me $200 a month, but I still manage to get most of what I want to do done without actually having to spend more money on top of my $200 So I'm just gonna go ahead and use the best model I have access to. Here's how Grok 4.6 performed on BucyBench. You can see what it generated here. It costs a little less than 8 cents to generate, used 13,000 tokens, and took about 3 minutes and 5 seconds to generate. And my LLM as a judge, which uses 3 different LLMs to judge this, ranked it at number 43 overall on the leaderboard, which interestingly, it ranked lower than Grok 4.5. It actually thought it was better. Thought 4.5 did a better job. Again, take these rankings with a grain of salt. It is LLM as a judge, so very subjective. So while this is a pretty good model at a better price, there are still models that are going to be quite a bit better. However, Elon did announce that 4.7 is significantly better than 4.6 and should be ready in 3 to 4 weeks. So we'll test that one out and see how that one does in a few weeks here. Google also released another model this week, not another pro model, not Gemini 4, but Gemini 3.7 Flash. So we got 3.6 Flash a few weeks ago. We've been hearing we're getting Gemini 3.5 Pro for a while, but it keeps on getting delayed. People are still expecting that. But instead of some big state-of-the-art model, we got another Flash model in Gemini 3.7 Flash. When we start looking at some of these benchmarks here, they're not even really comparing this one to any of the state-of-the-art models. We can see 3.7 Flash here on DeepSuite got a 65.3%, but it's comparing it to Sonnet 5, not Opus 5, not Fable 5, Sonnet 5, and 5. So not even like the best GPT models. Where this one really shines is again in cost. 3.7 Flash is available through the end of the year at an introductory price of $0.75 per million input tokens and $3.75 per million output tokens. If we look at this chart here, we can see cost going down to the right and then the DeepSuite score on the left. So we can see Gemini 3.7 Flash is pretty far to the right, meaning that it's pretty inexpensive to use, and it's pretty high up here as far as intelligence goes. Now, it's not at the level of Opus 5. It seems to be sort of in line with Fable 5, but a little bit lower, and a little bit lower than Kemi K3. So we can see the path that these Flash models have followed, where they're getting cheaper and cheaper to use while getting big jumps in intelligence. Now, here's the rest of the benchmarks if you want take a quick pause and look at them. But interestingly, it is really good at generating SVGs. Here's what it did on BuceyBench, and we can see it cost a little less than 9 cents, used 23,000 tokens, and took 1 minute and 19 seconds to generate. But if we look at the leaderboard, my LLM as a judge here ranked this as the new leader of BuceyBench, saying that this is the closest looking one to Bucey so far. All right, so the next model that we got this week actually comes from Meta. This model is called MuseGlimmer, And it's an open agentic model that runs on your device. So it's a smaller model. It's not going to be the most state-of-the-art intelligence because again, small on-device open weight, usually not going to be the most state-of-the-art. So when we look at the benchmarks down here, they're benchmarking it against other similar models, Gemma 431B, Quin 3.627B, models within that same open weight but smaller parameter models. And in its class, it pretty much outperforms all these other models in a lot of the benchmarks. GDPVal for, you know, knowledge work type stuff, not as good as Quin 3.6, but still not too far off, I guess. But again, these ones are designed for local deployment. It's meant to be downloaded and run with fairly consumer GPUs. It does say at full precision, this model would require 55 gigabytes of memory. So I mean, like, you can do it on a consumer PC with like the highest end GPUs. Like this would run on a DGX Spark, but they also have a quantitized version that can run with less RAM. So, you know, you'd get this one to run on like a 5090 or something. But again, the big deal with this one, it is an open weight model, so it's designed to be downloaded, run locally. You can fine-tune it, things like that. I did run it through BUCyBench, of course, and well, yeah, that's what you get. It generated for, you know, half a penny and used 4,289 tokens and only took 25 seconds. But yeah, you see what I see. And we got another open weight model this week, this one out of NVIDIA called Nimotron 3.5 Lightning. And this one I think is in pretty much the same class as the model we were just looking at. It's a 30 billion parameter model, just like Glimmer, and it is an open weight model. And here's some benchmarks for that model if you're interested. The light green on the left is this new Nimotron 3.5 Lightning. But if you're wondering how it did on BUCyBench, yeah, not a whole lot better than Glimmer. These smaller open weight models, not the best at generating SVGs, but it did do it for a third of a penny and only took 40 seconds and used 18,384 tokens. DeepSeek out of China also released a new model in DeepSeek V4 Pro. They did previously release a V4 Pro preview model, but this is a generally available release here, and it does have some updated benchmarks over the previous model. When looking at DeepSWE here, it scored a 62. 2.7, which is better than GLM 5.2 and even better than Opus 4.8, but not as good as Fable or Kimmy K3, and definitely not as good as GPT-5.6 Sol. Running this one through Busey Bench here, this is what we got. I don't think it knows what Gary Busey looks like, but it looks like a face at least, right? You can see it scored decently in the face coherence area, but it took about 8 cents, 25,000 tokens, and 3.5 minutes to generate. Microsoft released a new model called MAI Code 1.0. And I don't know if they really released much benchmarks or details about it. And it's not inside of OpenRouter, which is what I use for BucyBench to run all of my benchmarks. So I couldn't even benchmark this one on BucyBench. And then in the last bit of model news, I know there's been a lot, OpenAI announced GPT-5.6 Sol with ultra fast mode. Now this is using the Cerebras GPUs and generates 750 output tokens per second. So it is going to use their best, best model, be really good at coding and be ultra fast. But they only released this to a few, like, trusted testers. I'm not one of them, so I don't have access to this model as well. But we do see a demo here where this is generating something ultra fast on the left, and this is the sort of standard 5.6 SOL where it's generating much slower. Like, this one's already done and active while this one's still generating forever. I'm really excited to play with these really fast models, especially like state-of-the-art level intelligence like 5.6 SOL. 5.6 SOL, but like practically instant results using those models. So when you think about it, if you have 5.6 SOL, which is really smart, and you let it think for longer and longer and longer, but it's also doing that many tokens per second, so every minute it's doing so much more thinking than it would have done, you're just gonna get like insane results out of this in way less time. All right, so I know I've covered a lot already, but I did say there was a ton this week. There's still more I wanna cover, but I'm just gonna break it all down really quickly from this point on. So let's jump into a rapid fire. Pew pew. Pew pew pew pew pew. Let's start with some news out of OpenAI. If you really like to use Codex or the ChatGPT app, but you're a Linux user, well now the ChatGPT desktop app, which includes ChatGPT, ChatGPT Work, and Codex, is now available on Linux systems. Jumping over to Anthropic news, if you use Claude Code, auto mode is going to start being the default now. So starting August 14th, the day this video's coming out, new sessions on Pro, Max, and Team plans will run in auto mode. So basically for most things, things, it's going to automatically approve them. So you don't need to jump in and approve every little step unless Claude deems that it's something that could be harmful. Now you could already turn this on, but now it's gonna be default. And I think most people are gonna really like this. You're not gonna have to approve everything. You can just prompt something, walk away, and ideally come back like an hour later to a finished project. There's been an update to the Claude Chrome extension. It's now using Claude Cowork and actually can take more control on your behalf. So in their example here, they open the Claude sidebar extension inside of Chrome, and then they have it do work with all of the open tabs to generate a report. And it actually goes through, screenshots all of the various open tabs, and then once it's done screenshotting all of the open tabs, it continues to work inside of Claude Cowork on everything that it just screenshotted, then going on to generate a report directly inside of the Claude app. If you use Claude Code, your various Claude Code sessions can now message each other. This is This is actually kind of similar to what we were looking at with GrokBot earlier, where your different agents could be working on stuff and then message other agents that are working instead of having to re-explain yourself in another session. You can now tell Claude to do it. It sends a summary and the other session picks it up mid-task. We got a new video model in LTX 2.5. It can generate 10-second AI videos from an image in just 6.8 seconds if you're using NVIDIA's Super chips, and it's an open weight model. So here's an example of a video clip that it generated. Here's another clip of some drones carrying some goats, and it does say that it runs locally on NVIDIA RTX GPUs. Alibaba also released a video model last week called Wan 3.0 that generates some pretty decent-looking videos as well, and these ones can generate up to 30 seconds per clip. So another really, really good video model that you can use. Jumping back to Microsoft, they also released a new image model, MAI Image 2.6. And as far as the text-to-image arena leaderboard goes, it's currently ranked number 2, just behind GPT Image 2. If you're not familiar with the arena leaderboard, it's basically a blind test of generations. So you give it a prompt, it gives you 2 image outputs, and then people vote on their favorite output, not knowing what model it generated. And then based on that blind voting, that's how these leaderboards are created. So based on just taste test, essentially, people ranked this one second, only behind GPT Image 2. If you're a Twitch streamer and you don't want AI models to train on your stream or on the comments on your stream. You have to manually turn that off now. As of right now, Amazon can train on your stream, your video on demand, and even the comments that come in through your stream. So just something to keep in mind if you do use Twitch. Speaking of streams, I'm gonna start doing some streaming. I think the next one I'm gonna do is like next week, but I'll mention that again in a minute. And finally, Google DeepMind released this this week, which is an AI model that could read sign language. It's called Sign Language to Text. Or SL2T. And we can actually see it in use on some of their demo videos on the website. This person is doing sign language without any audio, and it's actually transcribing the sign language directly onto the video. That closed caption you see is reading the sign language as he is signing here. So a really valuable, helpful use of AI in my opinion. Anyway, that's what I got for you today. Again, it was a really, really busy week. A whole bunch of sort of like little Updates that combine to make it feel like a really, really big week. But that's what I'm here for. There is a lot flying around right now, a lot of new AI announcements and launches and products and companies shoving AI into everything. It gets really, really overwhelming. And that's what this channel is for. I drink from the fire hose all week. I overwhelm myself and keep myself looped in on everything that's happening and then turn around and make a video for you once a week that helps break it all down and tries to tell you what matters and what's important and what's not. There's some channels out there that are going to try to make a video about every single announcement that comes out and convince you it's the biggest thing that's ever happened. That's not what I'm trying to do. I'm trying to simplify it for you and just give you one video a week that helps you understand everything that's going on and get the full picture. If that's something that sounds interesting to you, consider liking this video and subscribing to this channel. That'll make sure more stuff like this shows up in your feed. Also, as of this recording, I'm only about 5,000 subscribers away from a million, so it'll also help me hit that milestone that I've been shooting for. Just a fun goalpost that I've been working on for 17 years now. Also, I did mention a couple minutes ago that I'm planning on doing some live streaming. So next week, I think on Tuesday, I'm going to go live on this channel and we're just going to chat. I'm going to bring on one of my buddies, we're going to talk about everything that's going on in the AI world and just have some conversations, chat with the comments, pull up some tools and test some things, and just have fun on a live stream. I think it'll be a really cool experience. Hopefully you'll join me. Uh, I'll set up a link where you can sort of set up a reminder if you want, but that should be Tuesday. We haven't picked a time yet, but I imagine it'll be Tuesday midday sometime, maybe 11 AM Pacific is kind of what I'm leaning towards right now. But again, thank you so much for hanging out with me, nerding out with me about all this AI stuff. The AI world has been an interesting roller coaster. I started talking about AI a little over 5 years ago when I started talking about it 5 years ago, the sort of discourse and the narrative around the AI world was, this is just the coolest thing ever and everybody needs to see this and it's exciting and fun and everybody just thought it was cool. Fast forward to today and the narrative's more like AI is destroying everything. The internet is worse because of AI. Data centers are destroying everything. And if you use AI, you're lazy and you suck. So it's been a weird ride to like be in the mix of that narrative arc over 5 years. I'm still excited by cool Cool new technology and trying to figure out the limits of what I can do with a lot of this stuff, and I'm still having a lot of fun playing with it. But it's definitely been a bummer to see that narrative shift, and I definitely also understand where people are coming from and why they're frustrated. So anyway, I don't know why I'm rambling about that. That's the kind of stuff we're gonna go deep on in our live stream, so make sure to join us there. Thanks again for nerding out. I really, really appreciate you hanging out with me. It's been a wild roller coaster over the last five years in this journey to a million, and I appreciate all of you that have been along for the ride. So thank you so. Thank you so much, and, uh, hopefully I'll see you in the next one. Love you. Bye.