AI NEWS: Did These Models Just Beat Google?

Well, it's Thanksgiving week here in America, which usually means a pretty quiet week in the world of AI. But it is my solemn duty to dig deep and find every single tidbit of AI news that deserves your attention. Or my name isn't Dr. Matthew Jarvis, AI Agent Wolf the third Esquire I did my doctoral thesis in the Gratuitous use of Memes of Digital media Commencement the end of one thing, the start of something new. But anyway, much to my surprise, there were some pretty crazy launches and updates that are actually pretty impressive. So buckle up Ron, we're going in. Sorry, were you talking to me? Let's start with Anthropic because they just released the most impressive coding model I've ever personally tested. On Monday they dropped Claude Opus 4.5 and this launched less than a week after Google dropped Gemini 3 on us. And I'm pretty convinced that this is now the best available model on the market for anything development related. So here's the breakdown. Claude Opus 4.5 is Anthropic's newest flagship model. It's dramatically smarter and is now considered considered the state of the art for coding agents and real computer use tasks. It handles ambiguity, it reasons about trade offs, and it can walk through multi system bugs like it's nothing. And pricing is actually cheaper now too. Opus level performance used to be way more expensive, but Opus 4.5 is now $5 for input and $25 for output per million tokens on the API and it's available across the entire Anthropic ecosystem. It's in the cloud app, the desktop app, the browser extension. It's in Excel now, it' all the major cloud providers and it's likely in whatever your preferred development IDE is. Lately I've been using Claude code a lot and yep, it's in there already. But the bigger story here is how many upgrades came with this new model. They updated the whole Claude developer platform. There's a new effort parameter that basically lets you pick whether you want speed and low cost or maximum brain power mode. At medium effort it matches Sonnet 4.5 while using 76% fewer tokens. And at max effort it surpasses Sonnet while still using less output. So you get more power, but you also get to spend less. And like I mentioned, cloud code also got a pretty big upgrade. There's a new plan mode that asks clarifying questions and then builds a plan markdown file and then executes on the plan. And in the desktop app you could literally run multiple cloud sessions at the Same time. So one could be fixing bugs, one can be searching GitHub for you, and another can be updating the docs. And for normal users who just want to use the Claude AI platform, you're now going to see longer conversations that don't hit a wall anymore because CLAUDE is automatically compacting older context. So you can basically have like an infinite chat. CLAUDE for Chrome is now available for all Max users, and CLAUDE for Excel is now open to Max team and Enterprise users. And when it comes to safety, Anthropic says this is their most aligned model yet. Whatever that means. I'm sorry, Dave, I'm afraid I can't do that. But I guess it's harder to trick with prompt injection than any other Frontier mod. And the safety evaluations seem to show the same. And I've been using it myself personally for coding, and I am not exaggerating. This thing is insanely good. For me, it feels like the first time an AI model actually understands how all the pieces connect across a bigger system. It's not just completing the code. You can actually see it's reasoning a lot better over the code. All right, so that's basically the overview. Now let me show you a few real examples of what I tested with it, because this is where it really shines. So one of the more fun things that I love to test whenever these new coding models come out is how well they code up little simple games. I'm not the most technical person in the world, and for my little ADD brain, having it make little fun games is my way of going, oh, this works really well or this kind of sucks. And Claude is pretty good at making little games. So I gave it the prompt to create a beautiful and colorful clone of vampire survivors. Make it visually appealing and put your own spin on it, but keep the gameplay itself the same. And it created stellar survivors. So it's like a space based vampire survivors. Here's what it looks like. So I've got this like little space ball thing here, and I got these little planets trying to attack me, but I can destroy them. And then as I destroy them, I gain xp. And then as I gain xp, ideally I can level up here at some point. Let me kill a couple more. All right. And now I get to level up. I can level up. Either my pickup range, do more damage, or I have homing stars. Let's do the homing stars. We could see my little planet thing is gotten a lot better. All right, let's add star beams. So now I have even more weapons shooting out of my guy here and it gets even easier to level up. Well do cosmic heal so I actually gain health. And you can see I'm collecting these little blue dots as my xp. So it pretty much modeled vampire survivors perfectly. Just like in in space version off of this two sentence prompt here that's always so impressive to me. I can give it two sentences and it's like, oh, I know what you want. Here you go. When I tested Gemini 3, I asked it to make like a Minecraft clone. So I figured, hey, let's test a Minecraft clone again here. And well, Claude didn't quite nail it as well on the first try as Gemini did. You could see it dropped me into this world that I actually think looks visually impressive. A little bit more visually impressive than the version Gemini made, unfortunately. See, it sort of put me inside the world instead of above the world. So I can't like actually jump on top of the blocks. But I actually think the world itself looks impressive. It just didn't get the physics right when I first tried it. But from like a cube based voxelized world generator, it looks kind of cool. And then I went and asked it to reimagine Super Mario Brothers 1 as a quarter view Voxelized game recreate level 1 of Super Mario Bros. But with voxel character art, voxel world and quarter view perspective. It should be playable and recognizable, but with a new twist. This is what it created for me. We could see it's not perfect and for whatever reason I keep getting like these loading errors. And look how fast he jumps and moves. That's also not quite ideal. He's like super Super Mario Mario on crack here.

It's really easy to beat the level though because he just hyper jumps over everything. So it put this like little voxel twist it making all of the characters out of a whole bunch of little cubes. It didn't really follow my perspective prompt and the speed is insane, but that's what it built. But the most impressive thing I built wasn't something I tried to one shot. It's a project I've been going back and forth on and working on for a while and Claude just figured it out. I was running into bugs and problems and all sorts of stuff and Claude just got it. And that's this journaling app that I made. I'm really, really big into journaling. I have for most of my life and I used to use an app called Day One, but it had some things I didn't like for the longest time. It wasn't Available on Windows computers. And I tend to switch off between Windows and Mac computers often. So I want to be able to journal wherever I journal. I also like to handwrite journals every once in a while, just in a physical notebook. And that didn't really work for day one. And sometimes I like to audio journal where I speak out my thoughts. And so I wanted to create the journaling app that was the absolute ideal perfect journaling app for me. And that's what I created with this one. You can see all of my journal entries are here on the left side. Some of my journal entries are just photo entries. Like we're working on remodeling our pool in our backyard. So I have this pool progress entry. You could see my image entries. You could get a quick preview of the images in the sidebar here. I've got my written post when I click into one that I could read here. And then I wanted to add some random fun features. So I have this AI generated art feature where based on my journal entry I can actually generate an infographic or generate a collage. Based on this entry I can regenerate the title with AI. So if I click this, it's going to read my journal entry and give me a new title for it. So Rainy Estes Park Brewing Stanley Hotel and Cabin Barbecue Day. And then down here I can generate a collage. It's going to read my journal and then generate a collage based on that. And then I also have this generate AI insights. So it'll read my journal entry and then give me some insights around what I journaled on. So here's the collage. You can see it sort of did this like a handwritten collage page. And it does this for every journal entry I do just for fun. This is using Nano Banana Pro here. And then it gives the AI insights, key observations, reflections, looking forward. It gives it a mood, rating themes that I talked about, key topics, and then patterns. It actually looks at recent journal entries around this one to spot patterns among multiple days of journal entries. So I've got AI analyzing my journal entries as I put em in here. I can also filter by sentiment, was it a more difficult journal entry where I was pissed off? Or was it a more joyful or positive or was it more neutral or reflective? And I can filter down to the mood of the posts. I have all these tags that AI is automatically generating for every post. So I can filter it down by tags and then I can filter it by ones I've already done AI analysis on versus ones that haven't When I go to create a new entry, I can give it a title, but if I leave it blank, AI will create a title for me. I could plug an entry date in, type out my journal entries. But I also made it so I can upload audio files as a journal entry and I can upload images as a journal entry or I can upload scanned handwritten pages. So if I upload a scanned handwritten page, it will actually OCR it, read all the text and then transcribe the text into a journal that I can then AI analyze and you know, get an AI generated title for. And if I upload an audio file, it takes that audio file, transcribes the audio file and saves that as the post. So I could type directly into this journal. I could handwrite in a notebook and scan it in and it will save it to the journal. I could speak into my phone or my watch and audio journal and it will save it in here and transcribe it. Or I can upload photos and it will just save those photos as an image entry. I've also made options to like bulk import entries. I have a page where you can see all the tags and how often they're used. I created a calendar view so you can actually see how frequently I journal. And they're color coded based on my mood. So the blue ones AI analyst hasn't run yet. Green are more joyful posts, gray are more neutral, orange are more reflective so I can at a glance see how I was feeling at any time during that year. And I can go back through years. And I imported all of my journal entries from over the years as well. And you can see most of them I haven't done AI analysis on. You can also see when I got onto some good streaks. There's also a month view here so I can flip over and see how often I'm doing in each month. Again color coded based on my mood. I built a media page so it just shows all of the media images that I've loaded in and if I click into one of them, I can quickly like scroll through them all using the left and right arrow keys. I built an inbox here so I can just toss photos in from my phone and then organize them later. These aren't saved into my journal yet, but I can click into one of these here. Give it a date, give it a title, or let AI pick a title and it will save it as a journal entry for whatever date. So they're just like an inbox of images that I haven't processed into my journal yet. And then I can bulk edit journal entries by clicking select and then selecting Multiple Entries. I can delete multiple entries or I can even merge entries together if they're both on the same day. And of course I gave it Dark mode and light mode. And this was all built with Cloud Opus 4. 5 over the last like 72 hours of me just going back and forth saying add this feature, add this feature, remove that feature, Here's a bug. Add that feature. And it nailed it. This isn't a single prompt. This is a couple days worth. But I built like the ideal journaling app for me, not publicly available for other people. I did save it on a cloud server so I can access it from any computer anywhere. But I just love making these sort of apps that are designed to solve one problem for me. So from my testing, if you're wondering whether or not to use Gemini 3 or Opus 4.5, this is my opinion on it. So Gemini 3 is better at one shotting great apps. A single prompt usually results in something pretty impressive. I also think Gemini has like better design taste, so what it creates often looks better than what Claude might create on its first attempt. However, Claude feels so much stronger to me. For once, you're more in the weeds with a project like it seems to solve and fix bugs way more reliably than Gemini and it seems to get stuck in a loop way less. Like Gemini will often try the same solution over and over again, even if it didn't work the first three times it tried it, where Opus will actually seem to continually try new things until it works. So if I had to pick just one, I'm finding Opus little bit more valuable at the moment. But since I don't have to pick just one, I find that Gemini is pretty good for creating that initial bones of a project and like designing a great front end. And then Opus is great for picking up where Gemini left off, fixing bugs, refactoring code, adding new features and things like that once you're deeper in the project. I mean, right now we're living in the best time ever to just develop cool stuff that you need for yourself and to let AI code little fixes for the bottlenecks you have in your life. Like Cloud Opus 4.5 just made that a little bit easier. And I'm honestly finding myself getting kind of addicted to prompting little apps and scripts into existence. It's pretty freaking cool. All right. So right on the heels of that Opus 4.5 news, we have another big drop in the AI world this week. And this one comes from Black Forest labs. It's called Flux 2. It's their new frontier visual intelligence model. And if you're wondering, yes, it seems awfully similar to what Nano Banana Pro can do. And yes, it also came out less than a week after Banana Banana did. But here's the headline features. It can handle up to 10 reference images at once, so characters, products, styles, outfits, things like that. And it'll keep everything consistent. It excels at generating photorealistic images and complex infographics with actually legible text. Very similar to what we saw at Nano Banana do. It understands structured prompts, multi part instructions, brand guidelines, layouts, logos, lightings, all of that kind of stuff. And it can edit Images up to 4 megapixels with without everything turning into like weird mush. They built the underlying architecture around the Mistral 324B vision language model, so it has real world knowledge baked in. Remember how we were looking at Nano Banana and it would sort of do research for you before generating the image and it would include what it found? Well, yeah, this has that also. It's also got better spatial reasoning, better lighting logic and much better typography than the earlier Flux models. Now when Black Forest Labs releases these models, they do it in an interesting way because they usually release a whole bunch of versions. And this one has four different versions. You've got Flux 2 Pro, which is their top tier, best quality, fastest speed model. This is the one for production teams and people who want closed model performance but at lower cost. But then you have Flux 2 Flex, which is the same core model, but you get full control. So you get control over the number of steps, the guidance scale, the quality, the speed, things like that. And it's probably one of the best models right now for text rendering in fine detail. Then you have Flux 2 dev, and this one is the one that opens open source people are going to care about because it's a 32 billion parameter open weight model which you can run locally. It supports multi reference generation and editing right inside of a single checkpoint. And it uses Nvidia's FP8 optimization so you can actually run it on consumer grade GPUs. And then finally you have Flux 2 Klein, which isn't available yet, but is going to be fully open Source under Apache 2.0. It's going to be smaller, faster and trained from the main model. So it'll be capable, but a much smaller size. And pricing is pretty competitive as well. Flux2Pro and Flux are available through The Black Forest Labs playground and through the API. And the open models are on hugging face fall replicate runwear together deep infra pretty much everywhere. And if you want to run dev locally, the weights are free and commercial licenses are available on their site. So the big takeaway here, Flux 1 is already the most popular open image model. Flux 2 is basically the production ready version with better realism, better prompt following, better text, better consistency, higher resolution, and they're openly making the tech behind it available. Now I haven't tested this one yet, but I'm about to. So now that you've got the overview, let's jump in and test a few things and see what it can actually do. All right, so the first prompt I tested here, high resolution product photo of a matte black electric skateboard on concrete city street, early morning light, cinematic shadows, 5 millimeter DSLR, shallow depth of field, realistic textures with minimal logo on skateboard deck. And I mean yeah, it did it. We got like a boosted board looking skateboard, concrete city street, early morning lighting. Pretty much hit it all pretty well. Next I want to try with an uploaded photo. So I gave it this image of a skateboard that I just found on Google Images and said add a realistic urban street scene plus lighting. Now it left these little like marks in the corner. I wish it didn't do that. But you can see it took those and it put them out on a city street and made it look nice. So decent for turning a plain product shot into something a little more visually appealing. I want to try to cram a bunch of stuff into a prompt here. So modern poster for a sci fi Movie Festival 24 by 36 poster layout, bold sans serif headline, galaxy nights, neon glow, spaceship, silhouette, et cetera, et cetera. You can pause if you want to read the whole thing, but you can see the images it gave us here. Pretty much nailed it. These all look pretty decent here. I like that one a lot. But they're all, you know, pretty good shots that followed the prompt. They all look like a movie poster. And here's the last one. So I uploaded three photos of myself here. You can barely see them, but they're all from like the same photo shoot where I'm wearing the same jeans and shirt and I give the prompt to generate a four panel lookbook of this character. Full body shot, side view, back view and close up portrait, studio lighting, etc. And here's the four that it generated. Well, here's the original three images and here's the four shots that it generated. A side View, a back view and a close up view looks pretty good, I would say. Like, some of my facial details look slightly off, but overall, not bad. Here's the second example. This one actually looks a little bit better. Here's the third example it generated. And here's the fourth example that it generated. So not too bad. Now, I don't know if you've seen this meme circulating around that somebody generated with Nano Banana that has Elon, Sundar, Tim Jensen, Satya, Mark, Jeff, and Sam all in a parking lot hanging out together. I wanted to see if I can recreate something similar, so I uploaded images of all those same people. And it did okay. Nowhere near what Nano Banana did. You can see here's the eight images that I uploaded, and I gave it the prompt. Make these eight men standing around in a circle in a parking lot top and passing a cigar around. Well, this one, it only got five of them. You can make out Satya Sundar. I think that's supposed to be Jensen. And I have no idea who these two are supposed to be. Here's the second one where it got seven people, but none of them really look like who they're supposed to look like. Here's another one where it got six people. Jensen looks the closest. Maybe Sundar. But it feels like it's blending people. Like this kind of looks like the love child of Sam Altman and Elon Musk, maybe.

And here's the last one. But you can see it pretty much lost the plot on all of these. I didn't put these characters into the scene. Definitely got to give the win on this kind of thing to Nano Banana. Next up, I gave it another image of myself and I said, reimagine this photo. Retouch the skin slightly. Adjust lighting to dramatic cinematic lighting. Add subtle film look. Color grading teal orange. Keep original composition, preserve original subject identity. And you can see it. Relit these images. I'll open up one. You can see this was the original. It's definitely washed out and over done. I wanted to clean it up and it did a pretty good job with these images of sort of stylizing them and giving like a new color grade. But the thing that really impresses me about Nano Banana was how it sort of searched the web and then was able to create infographics based on what it found. So we had to test that. So I went and said, create an infographic that explains how a neural network works. And it did that. Let's see how the text came out. How a neural network works from Inputs to predictions. And yeah, I mean we have some of the same issues I've seen with Nano Banana where the text just kind of turns into gibberish. So we have our input layer raw data features enter as volition as ever. It enters hidden layers, weighted sums and activation or or conscient functions. Output layer, final outdiction or ang classification for activations. So it just kind of turned into gibberish. I mean the image kind of sort of gets the concept right. Honestly, the images aren't bad. Nano Banana definitely does it better. I also think Nano Banana gets the text a little bit better, but they both sort of have this issue where when there's a lot of text in the image it really goes haywire. Here's another version. I like the graphics, they look pretty good but. But we do lose some text here. Back CS lay learn patterns. Most of the text is not bad compared to true La bella spelled learned wrong graphically not bad. Text wise, little to be desired. I like the colors on this. Same of issues though. And here's the final one. I mean visually they're very appealing. Nano Banana I think is a little more visually appealing and yeah, definitely not perfect with text still. So let's test this even further. An infographic that breaks down a step by step process to bake chocolate chip cookies. And we can see it made some pretty decent looking infographics. Gather ingredients. Flour, sugar, eggs and what kind of eggs are those? These look more like eggs. And then make sure you grab your saw pickups. Mix dry ingredients together. Then you gotta make sure you cream the tin stickers. Then add eggs and vanilla. And then do step four again add eggs and vanilles. This time don't forget to cock a ni vanilla. Then you got your sod combat mixtures. And then add your drocolate chips. Cool, cool. And enjoy. How to bake chocolate chip cookies Flour, butter. This is some interesting looking salt. You get the idea. Visually not too bad. Text wise falls off a cliff pretty quickly. Here's another version where you can go bulk some cookies. And finally how to bake chocolate chip cookies. Don't forget to bake until golden and cool cod and definitely don't skip over step five. Fold mixtures. Cream butter and sugar. A waldo epide avish soyers. Don't forget that step. And then finally I want to test a collage image that has each letter cut out from different magazines like a serial killer or kidnapper would write. But this one invites people to a festive networking event on Friday night and it did pretty good. On this you're invited festive networking event this Friday night. Join us to connect, celebrate and grow. Yeah, it did fine with this festive networking night. That looks good. So definitely the more text you have, the more likely it is to sort of break down. You're invited to a festive networking event this Friday night. Join us, meet new people. Friday 8pm I mean they all look good. Infographics. The text breaks down but this seems to actually nail it pretty well. So that's Flux 2 from Black Forest Labs. A pretty big leap, especially in the open weight open source world. Nano Banana, I still think think does a little bit better job at most of these tasks. But again, this is one they're designing to be able to run locally on your own computer and open weight. So you'll be able to use it with ComfyUI and things like that if that's what you're into. And those are pretty much the two biggest stories of the week. Now let me get into some of the like medium sized stories before we get into our rapid fire. If you're a fan of Notebook LM like I am, they rolled out some pretty cool new features this week. You can now create slide decks and infographics. You've got custom style prompting for your visual overviews, longer custom Personas. I mean really the main thing is slide decks and infographics because they've incorporated Nano Banana. Now let's jump into our Birds aren't Real notebook that we've played with in past videos. And we can see over on the right we got all our normal stuff, but we also got two new buttons, an infographic and a slide deck. So let's generate an infographic. Let's generate a slide deck. Here's the slide deck it generated for us. Birds aren't real how a joke took flight to fight misinformation. And we can really see how Nano Banana holds up with with a lot of text compared to what we just saw with Flux 2. But I love the graphic that it made. Like this is an epic birdshard real graphic for the most part. The text is pretty good too until you start looking at some of these like little signs. It says Dirk Truther and something Toothes. But overall, I mean pretty epic graphic here. And here's our slide deck which also looks kind of epic. Let's go ahead and press play here. Here if it flies, it spies an investigation into the avian surveillance state. And I believe this is also using Nano Banana to create all of these images. So here's slide one. This is Epic. Now if there was a way to pull this into like Google Slides or into PowerPoint or something like that and actually edit and rearrange stuff, that would be killer as well. But like, if you're making a slide presentation, this could get you like 80% of the way there. Here's the next step. Big bird is watching you. Pigeons are liars. These are obviously AI generated. You can tell because this says YouTube. But I mean, dang, this is. This is impressive. But what if the birds just birds. This is. This is really, really good. Honestly. It's like pulling in images and everything. Awesome. This is so good. This is really, really impressive. Republicans are better at spotting the joke, apparently. Graphs on the Satire versus non satire. An unlikely ally for avian conservation. Such a good slide deck. Bird watching goes both ways. So good. So good. Oh man, Google's. Google's been cooking lately. Google's been cooking. I dig this so much. But we got other stuff to share. If you like to use ChatGPT's voice mode, well, they just rolled out a new feature where you can press the voice mode and it keeps it all on the same ChatGPT screen. Instead of opening up a new screen with like the circle that you talk to, it just keeps it in the same screen. They say it's rolling out now and all you have to do is update your app on your phone to see it. For whatever reason, I have the most up to date version on my phone. There's no update for me to download and I don't have this feature. So maybe it's not fully rolled out or I'm doing something wrong. But you can see on the video what it looks like. It just keeps the audio chat going down in this like bottom right corner instead of taking up the whole screen. OpenAI also rolled out a new feature just in time for play, Black Friday, which does shopping research for you. So their example, best TVs for a bright room searching for best TVs and then it does a bunch of research and finds TVs for you. So I believe if you just use ChatGPT normally and ask it the question, it'll do it automatically. Oh, there's a shopping research button. So let's turn that one on. What's the best microphone for the type of work I do? I'll ask it that because it should have the memory in the context of everything I've chatted with it about. Starting shopping research, spend a few minutes researching. So let's go ahead and tell it to get started. Primary recording context on Camera, videos, total budget. Let's go up to 500 bucks. I'm not actually buying a new microphone, but let's see. Xlr.

Sure. Desktop. Boom. It's probably going to tell me the mic I'm already using. No, it didn't. Let's see. Give quick feedback to help ChatGPT pick the best option for you. So it is. Is it asking me to do like reinforcement learning training for it? What's. What's going on here? So it told me the Rode Podmic is the best microphone. Let's see more like this HyperX Blue Yeti, which is weird because that's not an XLR microphone. Why did it ask me what I want to use if it's going to offer me a USB microphone? All right, there's the mic I actually have. I guess it didn't recommend it first because my budget was under 500, so that's what that does for you. It kind of makes it a real simple shopping research experience directly inside of Chat GPT Perplexity, not to be left out on Black Friday, also rolled out a new shopping feature based on their video. It looks pretty much exactly like the same thing. So let's go to Perplexity. I'm already in the comment browser, but let's go ahead and open up Perplexity anyway and say, wonder if it knows what I do. What's the best microphone for the work I do do? Short answer for your work? AI news. Okay, so the MV7 is pretty much the same mic I have. That's just the USB version instead of the XLR version. So it did that research for me, but I think it would have always done that research that doesn't seem new. So let me click up on the shopping tab up here and then we've got some comparisons and I'm not sure if this is what the shopping looked like before or not. I hadn't really used the shopping feature, so I can't tell if this is the updated version or the way it used to be. But this is what I'm seeing when I test Shopping in Perplexity. And since we're talking about Perplexity, they also rolled out better better memory and Perplexity, which is also kind of why I asked it what's the best mic for what I do? Because I wanted to see if this new memory feature knew what I did. But I don't think it really did yet because it just went looking for a mic for business purposes. So it didn't seem like it was using the memory. But we do have better memory instead of perplexity. So I guess maybe chats from here on out will be better remembered. It says with our upgraded personalization, Perplexity remembers key details across conversations, so future answers are more personalized and efficient. Structured preferences like favorite brands, dietary needs, or keywords you often ask about are securely stored for valuable context. As you ask more of perplexity, the system builds a dynamic understanding. So it works similar to how ChatGPT's memory works, it seems like. So that seems like that will be helpful over time. All right, so that was all the major news stories and middle news stories, but I do have more I want to show you. But this will all be quick. So let's get into a rapid fire.

Foreign.

This week Trump signed an executive order launching an AI initiative that's being compared to the Manhattan Project. And we all know how the people who worked on the Manhattan Project felt about the result of that. Now I am become death, the destroyer of worlds. So this is called the Genesis Mission and it's a push to mobilize the federal government's research and data to create artificial intelligence models. The Genesis Mission will build an integrated AI platform to harness federal scientific data sets, the world's largest collection of such data sets developed over decades of federal investments to train scientific foundation models and create AI agents to test new hypotheses, automate research workflows, and accelerate scientific breakthroughs, the order reads. So basically the government is pushing for and supporting new AI research. That's pretty much the main headline here is they're gonna pump resources towards trying to keep the US the number one in AI. I guess we got a new smaller model out of Microsoft called FARA 7B and this one's an agentic model designed specifically for computer use, but also designed to run locally on your computer. So a lot of the computer use models that we're seeing get built into like Perplexity, Comment, and inside of the chat, GPT Agent and Manus and things like that. This seems like a model that is solely designed for that, but small enough that you could run it locally on your own computer and will likely run probably like agentic features inside of Windows or something in the future. Unlike traditional chat models that just generate text based responses, computer use agents like this one leverage computer interfaces such as a mouse and keyboard to complete tasks on behalf of the user. It's smaller and can run local, which results in reduced latency and improved privacy because, well, it's local. This one is made available as well, so it's available on Microsoft Foundry and Hugging Face. Not one I've tested yet. But if it's something people are interested in, maybe I'll do a tutorial. And we try to run a local computer use agent. Well, we've seen World Labs marble. We've seen DeepMind's Genie and Sema. Well, now we've got Meta's World Gen, which is text to immersive 3D worlds. Now, I don't think this is something we actually have access to play with yet, but all these companies seem like they're trying to build the holodeck. Here's some examples of what these 3D worlds look like. Again, we don't have demo access yet, so they prompt a medieval village, it creates a few options, they select one of the options, it then generates like a point cloud or Gaussian splat or something. Cyberpunk slum. And we can see a bunch of different examples. But then I believe, yeah, they can zoom in on the world, actually move around inside of the creation that it made. Very similar to what we're seeing from World Labs Marble and from that Genie 3 model out of DeepMind. But this is Meta's version, and because it comes from Meta, you know, they're gonna stick it inside the Meta quest. So you can jump in there and say, generate a world of, you know, gumdrops and unicorns, and just like if you were in a holodeck, that world will just generate around you and you can walk around in it. That's what they're going to for, guaranteed. But that's also what Marvel is going for. That's also what Genie 3 is going for. But still, that'll be some fun technology that will feel like a really good use of VR. Honestly. There are more examples on Meta's website if you want to see more demos of what this is capable of. As always, we're going to link up everything we talk about in the description below. A while ago we got some teases that Sam Altman and Jony I've are working together on some sort of hardware. Well, I've got news for you. And that news is that we still don't really have much news other than the fact that. Fact that, well, I guess they have a prototype that we can't see. Johnny Ivan, Sam Altman say they finally have an AI hardware prototype and apparently it's less than two years away. That's all we know. There's no actual, like visuals of the prototype or anything. And Sam said it's supposed to be something more like calming than like a phone. But other than that, we don't really know much more. And since we're talking about OpenAI, well, they're having a tough time at the moment. They're getting cease and desisted by the company Cameo for their use of the word cameo inside of the Sora 2 app. So they're going to have to rebrand the Cameo feature to something else. Otherwise they're going to have to battle it out with the company Cameo. And apparently Sam Altman himself feels like they're falling behind Google right now. OpenAI CEO Sam Altman told colleagues last month that Google's recent progress in artificial intelligence could create some temporary economic headwinds for a company, though he added that OpenAI would emerge ahead. So he's seeing the progress Google is making and feeling like maybe they're falling behind a little bit. But he did go on to say ChatGPT is AI to most people, and I expect that to continue. So he's sort of betting on the brand name of ChatGPT to keep them ahead of the game. And, well, we'll see how that plays out. In AI Music news, Suno made a deal with Warner Music Group. And if this sounds like familiar news, that's because, like last week, the company Udio collaborated with Warner Music Group. Now, when Yu do collaborated with Warner Music Group, almost immediately Yudio took away the ability to download the songs that you generate. People flipped out, like, hey, you said I could make songs and download them. That's what I was sold. And then you took that feature away. Well, Yu Dio's response was, okay, we'll open up downloads for 72 hours. They let everybody download their stuff for 72 hours, and then they took the feature away. So I don't know how good of a sign that is for Suno that they just partnered with the same music group. But it sounds like Suno and Udo and Warner Music are all trying to to figure out some way that these AI models can use their music and train on them and allow users to create versions of songs from who they were trained on. But they're trying to do it while also minimizing the damage to the music industry and the artists. And it's gonna be muddy for a while. It's gonna be weird and muddy. And, you know, eventually we saw what happened with Napster. The big music companies fought it, fought it, fought it. And then eventually a solution was worked out with tools like Spotify and Apple. Like, that's going to happen with AI Music as well, where it's Going to be muddy for a little bit. There's going to be companies fighting, but eventually, over time, it's all going to sort of smooth out and they'll figure out a workable solution that everybody's happy with, most likely. And finally, we got an update out of LTX this week. I haven't played with this feature yet, but it looks really cool. So it's a feature called Retake, so you can actually direct your shot after it's rendered. You can rephrase dialogue, reshape emotion and reframe moments without redoing the full video. So it looks like what this is able to do is you can generate, generate a video and let's say it's like a. I don't know how long, let's say it's like an eight second video, but you don't like how the video plays out from five seconds on. Well, you could keep the first five seconds the same and regenerate those final three seconds. That's what I'm taking from it. So let's say you do like that first four seconds. Well, maybe you could generate like 10 alternate final three seconds and use the one that you like the best, that it had what you were looking for. I don't know. We'll play with it some more in a future video, but this one's already getting longer than I wanted so I'm going to go ahead and cut off the news there. Again, this was a slower news week, so I spent a little bit more time talking about things like Claude and like Flux. But even on slow weeks, we still get cool stuff and this does not seem to be decelerating. We're only getting more and more AI news every week. So if you want to stay looped in with all the latest AI news, get demo demos, see what they're capable of, get tutorials on how to use it, all that kind of stuff. You want to stay tapped in. Make sure you like this video. Subscribe to this channel and you will be looped in. I'm keeping up with this for you. I make videos twice a week like this one to make sure you stay informed. I really, really appreciate you hanging out with me, nerding out with me today. It's been another fun one and hopefully I'll see you in the next one. Bye bye. Mixtures A Waldo EPD Office. Thank you so much for nerding out with me today. If you like videos like this, make sure to give it a thumbs up and subscribe to this channel. I'll make sure more videos like this show up in your YouTube feed. And if you haven't already, check out FutureTools IO, where I share all the coolest AI tools and all the latest AI news. And there's an awesome free newsletter. Thanks again. Really appreciate you. See you in the next one.