AI News: OpenAI Made a Massive Move Against NVIDIA

So there's been an interesting narrative playing out in the world of AI right now, and I've just been absolutely fascinated by it. And that's the whole narrative of the open weight models versus the closed weight models. And that whole storyline sort of accelerated this week. Earlier this week, OpenAI showed off the first results from its Jalapeño chip. This is an AI inference chip that they're building to try to release reliance on NVIDIA. They announced that they were working on it Not too long ago, but they finally showed off some results of what it's capable of. We can see here on open weight models, it's getting upwards of 104 times the performance. Now they tested this on public models on purpose because it's much more verifiable. Other companies can go and use other chips and see what sort of performance they can get on GPT OSS or DeepSeek R1, etc. However, they can't really go and test GPT 5.6. on different chips since those all run in OpenAI's cloud. They do, however, claim that in their internal testing, Jalapeño's advantage widened further on their own Frontier OpenAI models. Now, this is interesting to me because it shows that OpenAI is actively taking steps to release reliance on NVIDIA chips. Now, they recently partnered up with Cerebras to do ultra-fast inference with their chips, and now it even seems like they're making like a competing chip to what they were getting from the Cerebras chip. One thing that is important to note though is that these are specifically for inference. So inference is essentially when you type a prompt and you get the output. That is the inference phase of AI. These chips don't appear to be designed for actually training brand new models, which they will likely still go to Nvidia to do, at least in the near term. However, this story gets even more interesting. According to the information, Nvidia agreed to buy Hugging Face. Now, I emphasized according to the information because neither Hugging Face or NVIDIA have publicly made an announcement about this as of the time of me recording it. Now, if you're not familiar with Hugging Face, Hugging Face is like the GitHub for AI models. When companies go and make open weight models and they go and make 'em publicly available to anybody, they typically go and store those weights on Hugging Face. Now, Hugging Face is also a GPU provider. You can go on their site and essentially rent GPUs from them. to play around with the various models that are available on their site. Now, here's why I find this super fascinating. You have OpenAI over here starting to build their own chips for inference, chips that they would have bought from Nvidia if they didn't start making them themselves. You have Meta starting to build their own chips in-house. You have Google, who's already building their own chips with their TPUs. And as these AI companies get bigger and bigger and start trying to own more of the stack themselves, it could start to shift some business away from Nvidia. But Nvidia is going all in on open source. All of the models Nvidia has been developing and releasing have been these open weight models that anybody can use. Now, most of them are too big to run on a personal computer, so you need to go and run them in a cloud. Hugging Face is where people go oftentimes to run these models in the cloud now. And if Nvidia owns all of that infrastructure to run these models in the cloud, and more and more people start switching to open weight Well, then Nvidia's no longer gonna be as reliant on the OpenAIs and the Metas of the world for this inference compute, because now everybody that's moving to open weight models is now going to directly go to an Nvidia company to get their inference provided. So they're betting that open weight models are gonna get bigger and bigger and bigger and take away more and more market share from the closed models. And with owning Hugging Face, they're gonna be the biggest winner. dominant compute provider of everybody running open models. And there is some data showing that we are seeing more and more people and companies move to open weight models. So check this out. This is a post from the CEO of Vercel, and he's showing that 2 months ago, most of the tokens on Vercel's AI Gateway were going to closed weight models, 71.6%, with only 28.4% going to open weight models. Fast forward to this week, and now open models account for 62% of the token usage versus 38% for the closed usage. Now, there is a caveat to this. When we go to their AI Gateway model leaderboard here, here's the graph that he was showing, showing the token volume. So back here, blue is the open weight, yellow is the closed weight. Closed weight models were getting most of the token usage. However, there is a dropdown here that shows the amount of requests. If we select requests, that tells us a little bit different of a story. It tells us that most people that are using this gateway are still making their requests to closed weight models. It's still 62% closed weight to 38% open weight on actual requests to the models. But the amount of tokens being used, well, it's pretty clear that the open models just use a lot more tokens. Saying that, we are seeing more and more requests recently going to more of the open weight models. So yeah, I think that NVIDIA acquiring Hugging Face, if true, is actually a pretty big deal because again, it is NVIDIA betting that open weight is more of the future. And to round this whole discussion out, there was another thing that happened this week that, you know, kind of adds to the story as well. And that's the fact that Apple just introduced some new chips this week. They announced the M6 and the M5 Ultra, and these are designed for AI compute. These are designed to be devices that live on your desktop that can run your compute completely locally. And this M5 Ultra chip is insane. I'm currently running an M3 Ultra, and this has 4.5 times the peak GPU compute for AI, and my computer right now is an insanely good AI computer. They're also releasing it in versions all the way up to 512 gigabytes of unified memory, which means pretty much most of that 512 gigabytes can be used as essentially VRAM for running AI models. Now, they're not going to be cheap. An M5 Ultra with only a 1 terabyte hard drive is still going to cost you almost 11 grand, and that's for the 256 gigabyte version. They haven't released the price of the 512 12GB version yet. But the bigger point being here is that we're starting to get more and more devices that are very capable of running really good models at home. Like, you're going to be able to run these really good models coming out of ZAI, like the GLM models. You're going to be able to run some of the DeepSeek models. You're going to be able to run some of these bigger QWinn models on this thing. And I mean, it's only a matter of time before we're going to get like fable-level GPT-5.6 SOL-level models that will run on a computer like this. But I've been finding this whole story just Fascinating watching the open source versus closed source battle. If you would have asked me 3 or 4 years ago, I would have told you I don't see open weight models catching up to the closed weight anytime soon. And here we are a few years later and like we're this close. Speaking of open weight models, we got 2 new models this week that I want to talk about real quick. First, we got GLM 5.3 Flash, and this model kind of flew under the radar and launched in stealth like a week or 2 ago under the name OX Alpha. And when we look at the benchmarks here, again, DeepSuite is kind of the one that I pay the most attention to because I think it's the most accurate to like actually what we'll feel when we're coding with these models. And this one scored a 63.4, which outperforms Opus 4.8 and falls just shy of Google's like most recent model, Gemini 3.7 Flash. I also think this chart right here is very, very telling. So on the left side, You've got the Artificial Analysis Intelligence Index. This is sort of a combined score across a whole bunch of different benchmarks to kind of pick the objectively smartest model. On the bottom, you've got your compute cost. This red dot here is where GLM 5.3 Flash sits. So as far as intelligence goes, it's right around QWinn 3.8 Max level, but way cheaper to run. And notice how some of these other models aren't even that big of a jump. Claude Opus 5, Claude Fable 5, GPT-5.7, Like not a massive intelligence leap over what you're getting over GLM 5.3 Flash. Everything that falls into this red box that you're seeing here is basically both dumber and more expensive than GLM 5.3 Flash. And that includes models like Gemini 3.7 Flash, Claude Sonnet 5, Gemini 3.1 Pro, DeepSeek V4. These models are all more expensive and less intelligent, which is why this model release is kind of a big deal. It just made a big chunk of models kind of obsolete. And if you're curious how it performed on BUCyBench, here's what we got. Now, it was during a period where I was able to use the model for free, so I'll probably run it again to get an actual accurate cost. And it also ran it when it was called OX Alpha, so the name needs to be updated here. But it took 12 minutes to run, used 46,000 tokens, and came in at a 5.4, which on my very, very subjective leaderboard here puts it at number 25 overall. But we got another open-source model this week as well. We got Now recently we got QWIN 3.827B, which was a really good model for running locally on high-end consumer GPU. This one is a little bit bigger. It's 125 billion parameters. Like that new M5 Ultra would run this model probably, but most likely you're going to want to run it in a cloud. It does decent on the DeepSWE benchmark and on artificial analysis. It falls at about a 56 compared to a 57 from the model we were just looking at, GLM 5.3 Flash. Quinn 3.8 Flash generated this on Bucybench. It took a little over a penny to generate, just under 5 minutes, and used 26,000 tokens. Interestingly, this one got ranked at number 19, which came in below what my AI ranking thought of Quinn 3.8 27B. But again, the rankings here don't really matter. You can just use your own eyes and determine which one you think did a better job. Here's another cool update that's been out for a little while now, but I just came up with a cool new way to use it. I'm going to use ChatGPT Sites inside of the Codex app to turn my spreadsheet of social media data into a live performance dashboard that my whole team can access and use. So inside the Codex app, I'm just going to type @sites and say, build me a social media performance dashboard that tracks all my platforms and that I can share with my team. And then Sites will create the website plus handle everything else like the hosting and the authentication and even a built-in database. Now the underlying data is coming from this spreadsheet that tracks my audience across YouTube, the Future Tools newsletter, Instagram, LinkedIn, TikTok, X, Threads, and even Facebook. So now my team can quickly see our audience weekly growth and which platforms are moving the fastest. And now we can even build and refine the site together and connect it to a custom domain if we want. With site analytics, I can see basic performance metrics like visitors and page views. And the really interesting part is this new WebMCP support. You can build predetermined actions into the site that compatible AI agents can use instead of the agent blindly clicking around the page. This is a really easy way to build tiny pieces of custom software for your team without dealing with all of the infrastructure that you'd normally need. So if you want to try Sites out for yourself, head to the link in the description box to download the Codex app. And thanks so much to OpenAI for supporting my channel and sponsoring this portion of today's video. We also got like a flurry of new announcements out of Google this week. They released an updated version of their video model called Gemini Omni 1.1 Flash. I'm guessing they're The older Veo model is pretty much done for because everything seems to be this new Omni model. We can see this new model is the same price as the previous model, 10 cents for a 720p output. You can also do a 360p output, 1080p, and a 4K output. The 360p is nice if you want like a quick generation just to like test to see how something might come out. And if you think it does a good job, I guess you can always do a new generation at a higher resolution. The new model can analyze up to 10 seconds of prior context. You can specify start and end frames so that can have them animate between or do like cool transitions and things. Do drafts in 360p, upscale to 4K, add video reference. Seems like a lot of the same stuff we were able to do with the original Omni. It's just we have better quality versions and a little bit more context. Here's some examples with the prompts on the screen. Start and end frame. A microscopic view of iridescent marine diatoms displaying intricate glass silica shells with breathtaking natural symmetry, etc. Interestingly, over on arena.ai, under the text-to-video arena, Gemini Omni 1.1 Flash is the new top model there. If you don't know what arena is, it's basically like a blind test. You give arena a prompt, it gives you 2 outputs. From 2 different video models, not telling you which model is which. It can be any model that's in their database. You pick which one is your favorite between the 2, and then after enough people have done their voting, that's how you get the ranking. So it's a blind test of user preference, and Omni 1.1 Flash is winning in text-to-video. And in image-to-video, it is second only behind Minimax, and not by a wide margin. But you can also see not a ton of votes yet. It's still early. Now, their announcement here says you can build in AI Studio using your API key Or if you're an AI Plus, Pro, or Ultra subscriber, you could use it inside of Google Flow. So first I went to AI Studio and have tried to prompt multiple times and I keep getting an internal server error. I don't know if I'm doing something wrong. I even went and pre-added money in, so I put $25 in here so I have credits and I still couldn't get it to work. So I took my prompt over to Flow because I am an Ultra subscriber and this was the prompt I gave it: a video of a robot wearing a hat that says Future Tools on it breakdances in front of a castle where a dragon flies overhead. My initial thought is it's okay. Again, I've talked about this a bit lately. I'm less and less impressed by video models these days, and this one I do see some wonkiness in it. The castle in the background looks fine, but my robot, for some reason, the head just pops off, falls on the ground, and then apparently it's got the same hat on both sides and the same face on both sides. Although When you look at it here, it clearly doesn't. So that's a little funky. And then there's some wonkiness when he starts doing his like ground moves here. Like yeah, it's just a little awkward. The consistency of the castle and the dragon are really good between scenes though. That's one thing that's gotten a lot better over time is consistency when it switches scenes and then comes back to the same scene. Things look really good. That's something that's really, really improved with AI video models. Google also released a new transcription model, Gemini 3.5 Transcribe. Now this is It's pretty much exactly what it sounds like. It's, you know, their version of Whisper, but apparently more accurate than everything else that's out there. And it's available in the Gemini app on macOS, Rambler on Android, and coming soon to Chrome. But you also have it in the various enterprise platforms and via the API if you'd want to use it. Google also added some new AI features for travel inside of the AI mode in Search. It can track flight prices for you now. It'll show you in both price and points, and you can apparently book hotels directly inside of AI mode in Search. Search now as well. There's also a new update to Gemini Notebook, previously known as Notebook LM. You can now incorporate books that you own directly into your notebook. So they say we've built the ability to add select ebooks you've purchased from Google Play Books directly to a Gemini Notebook. So you would've had to purchase it through, you know, Google's book platform. It's a pretty cool feature, but personally when I do buy ebooks, it's mostly on Kindle. And if you use Anti-Gravity to do your coding, they finally added a remote control feature This is a feature that pretty much every other sort of AI-enabled IDE has had for a while now, but now it is also inside of Antigravity. Remote control allows you to connect and drive your Antigravity sessions running across your different machines like laptops, servers, and desktops. All you need is a web browser. While you have your web browser handy, you're able to remotely connect to each of your different machines running Antigravity and retain full access to your files, workspaces, build tools, credentials, et cetera. Again, this is something we've had in Codex for a while now. We've had it in Claude Code for a bit now. It's been in Cursor for a while. Bit now. And well, you could do it in anti-gravity now too. Now it has been a really, really busy week and I've shared a lot already, and I still have quite a few more things I want to share, but I'm just going to run through them super quick in a rapid fire. We got a couple additional updates out of OpenAI. ChatGPT Work can now use its computer and browser to sign into websites on both web and mobile. Now if you're using the ChatGPT app, it has a built-in browser which can remember passwords and things like that, but this is now available on Here's an example they give to log into DMV and help me replace my driver's license. And you can see it logs into the site and then asks you to continue with your password. And it could log in with things like 1Password or your built-in password manager, and it never actually knows your passwords. Now, personally, I still like using the ChatGPT app better because it will just open the browser and then log in for you using whatever saved credentials you have inside of the ChatGPT browser. But if you like to use the web browser version, This is a pretty cool new quality of life update feature as well. There's another new feature inside of ChatGPT Work on web and mobile. If you're on a Plus or a Pro plan, you can now have tasks run based on things changing in places like Slack, Gmail, and GitHub. So in the past, you've got this scheduled tab here and you can have it schedule a task to run, you know, every day at 2 o'clock. Now you can say schedule a task every time I get this Slack message or schedule a task every time this Gmail message comes across. So more ways to have automations fire off on your behalf. OpenAI also dropped the price on their most powerful model. So their GPT-5.6 Sol model is actually 20% cheaper for the next 3 months. So that's pretty cool as well. Moving on to Anthropic news, Claude is getting its own browser inside of Cowork. You know, we saw this happen in Codex, and Anthropic likes to steal OpenAI's homework. And whatever a feature looks cool that OpenAI does, Anthropic seems to roll it into their products these days as well. Codex has its own browser A couple weeks later, Claude is like, yeah, us too. Although I don't believe it's available quite yet. Yeah, so we see the built-in browser is rolling out over the coming weeks to Pro, Max, and Teams plans on Claude desktop app on macOS, Windows, and Linux. They announced that it's coming, but it's not here yet. But they also have Claude in Chrome, which is now generally available, which is a Chrome extension that will basically, you know, summarize articles for you on the web. It will fill in details for you. It can actually click around, but it is pretty So I'm hoping when they actually build the browser into their app, that will be a much faster way to do a lot of those things. Anthropic also announced this week that Claude's memory works everywhere. So starting today, the memory you use in chat is the same as in Claude Cowork. Cowork now has memory and it's the same one you use in chat, leading to less re-explaining and more picking up where you left off. When Cowork runs a task in the cloud, what Claude remembers from your chats is there and vice versa. So you can sort of plan out something you wanna do in a regular chat, jump over to Cowork and have it do it. cowork execute it and it will remember what you chatted about in a previous chat. By default, Claude does not store topics related to personal or sensitive subject matter like health, race, ethnicity, religious beliefs, politics, etc. But you can go in and change what it does save about you and what you don't want it to remember about your conversations. Jumping back to the conversation about AI running locally, Perplexity's computer now has a portable computer which is local. So the model, harness, conversation, and trajectory all live on the user's machine. However, work that needs the outside world, such as web search connectors or escalation to a stronger advisor model on the cloud, is invoked only when necessary and always gated by the users. Local models carry no inference fee, so the system is private and cost-effective by construction. So they're using small, efficient models like NVIDIA's Neutron, Quin 3.6, Quin 3.8. And it does say systems such as the NVIDIA DGX Spark can now run these models locally. So you must need a pretty decent system to be able to run this local computer model. Basically, my understanding is that Perplexity built their own version of an OpenCLAW or Hermes agent that will just run on your computer and do tasks that you need it to do on your computer. It will run local models, it will go off and use cloud models when necessary, but it'll ask your permission first. But it's kind of designed for really decent computers that have a lot of compute capacity, seeing as they called out a DGX Spark, which is, you know, a roughly $4,000 box for 128 gigabytes of VRAM. So not quite accessible to everybody yet, but like, this is their local version of, uh, OpenCLAW Hermes kind of agent. Harnessed thing. Let's talk about AI music for a second. Dr. Dre admitted that he uses AI to produce songs, which is a very risky thing for an artist to do these days. He said, "I don't see it as a threat. I think the only people that see it as a threat are the people who have trouble creating." I had a discussion with a few people a few days ago. They were against AI, and I'm like, "Okay, you sound like the person that would have been against the drum machine when it came out, or synthesizers." Right now, I do have my own take on AI music, and. I agree with Dr. Dre. I think AI is a great tool to assist with music. Where I think the line should be drawn and where I'm not a fan of AI music is just like prompting a whole song into existence, right? Like going to Suno, creating a song that sounds like any other band, and then putting it up on Spotify and saying, I created this. No, you just prompted that song. However, I've played around with Suno a lot. I actually think it's a really, really fun tool. It's a blast to make silly songs for my own personal use and listen to 'em and play 'em for my family and we laugh about it. But I'm not a big fan of publishing 'em And trying to fool people into thinking that they're real. Something that is really, really fun though is I've grabbed my guitars before and started to play like a little riff, put that into Suno, and then had Suno play drums over the riff that I made. Now it's like a collaborative thing between me and AI, and that's really, really fun. So I definitely think there's like levels to AI music, and I think just prompting a song into existence and saying, look, I'm a musician now, I don't know about that. But using AI tools to assist in making cool music where it's like a collaborator with you, I'm actually pretty cool with that. That's just my take. And speaking of music, this is really interesting. The company Stability AI just raised a new round of funding, and they actually have music generators under their umbrella of tools they've built. And the investors that just invested are comprised of Electronic Arts, Sony Music Group, Universal Music Group, Warner Music Group, and more. So I think this kind of tells That the music industry, the big studios, aren't necessarily against AI; they're actively investing in it. They just, you know, don't want to be cut out of the loop of making money on AI music. Time Magazine released their list of 100 most influential people in AI, and it's an absolute joke. Now, don't get me wrong; there's a ton of people that I do think deserve to be on this list. I definitely think the OpenAI team, Elon Musk. Elon Musk, Dario, Peter Steinberger, Mikey Shulman, Fei-Fei Li, Ilya Sutskever, Mira Murati. These are all people that I think absolutely deserve to be on that list. However, Jensen Huang, the person who's making this happen through all of this compute that he's providing to all of these companies, one of the people that's probably shaped the, like, AI revolution the most. Demis Hassabis, the ex-CEO of DeepMind, who literally has a Nobel Prize for his AI with AlphaFold. Andrej Karpathy, who's literally the most followed voice on X around AI, and even like Mark Zuckerberg, who's trying to be like the voice of open weight models and push that forward. All of these people are like pretty influential and they didn't make the list. Yet somehow Paris Hilton is on the list. Joseph Gordon-Levitt is on the list. Bernie Sanders is on the list. Like, what? And the reason Paris Hilton is on the list is that she's an advocate against deep fakes. I mean, nothing against any of those people, they just don't deserve to be on a list of the most influential voices in AI. Take it from me, somebody who is literally in the weeds of AI every single day. I read every bit of news, I read all the articles, I read the research papers, I watch the YouTube videos, I scroll the threads, I try the tools. I am in the thick of it every single day. These are not influential people in the space. They're just not. And I know I'm getting worked up over something that's probably designed to work people up, right? Like Time Magazine is probably purposely rage-baiting us knowing that this is an incomplete list and like names like Paris Hilton are on it to get people to go, what the hell is Paris Hilton doing on this list? And get people like me and other like people on social media talking about how bad this list is just because it drives more eyeballs to Time. And I played right into it. But okay, I'm done ranting. I just think this list is absolutely ridiculous. Let's end with a cool robot. So Skilled AI introduced S1, which is a new foundation model for robots. It can be taught 10-minute-long tasks that it has never seen before from just one video prompt without any fine-tuning. So in this example, you can see them wearing like little wrist things with cameras on it doing the task. And then when they're done doing the task, the robots know how to operate their hands and recreate the task because they could watch the cameras, right? There's a camera on each wrist. And there's 2 robot arms. And now based on what they saw from the cameras, they can now do that. That's pretty cool. I love seeing advancements in robotics and, you know, making it easier and easier to train robots. So yeah, this one's pretty awesome. Oh, and by the way, if you haven't watched some of the videos from the World Humanoid Robot Games, you should definitely look them up. They're kind of awesome. And that's what I got for you today. Now, before I sign off, this week was a really big milestone for me personally as well. This was the week that I actually passed a million subscribers on YouTube, and I I can't thank you enough for being a part of that. Everybody who watches these videos has hit the subscribe button, who has liked the videos, which has helped me get more reach on these videos. I appreciate you more than you know. It has been an awesome ride and I'm excited to keep on going and seeing how big this thing could get and how much more helpful I could try to make these videos as I go. It's my goal every week to drink from the fire hose and keep up with all the AI news. I'm not gonna try to pretend that everything is the biggest deal that's ever happened. I'm just going to share my thoughts as they come out about the news that I think is the most important to the most amount of people. And if that's something that's interesting to you, maybe consider liking this video. And if you haven't already subscribed to this channel, a million of you have already done it. And again, thank you so much. I do record these videos on Thursdays and publish them on Friday. So if anything came out sort of late Thursday evening or on Friday, it'll usually make it into next Friday's video. But again, I'll do my best to keep you looped in and hopefully make you feel like you don't have to keep up with it on a daily basis. I got a lot of cool thank-yous and shoutouts from a lot of different people over the last few days as I crossed that million subscriber milestone. And so to wrap this video up, I thought it would only be appropriate to share a few of those messages that people sent to me just because I think it's pretty cool. And like, I'm very grateful that people took the time out of their day to send me some of these messages. So thank you so much for tuning in. I really, really appreciate you. It's been so much fun nerding out over the last 17 years, but really 5 and a half years about AI. And, uh, I'm gonna keep on going with this. Thank you again for tuning in. I really, really appreciate you, and hopefully I'll see you in the next one. Congratulations on a million subscribers! Congratulations on a million on YouTube! That is amazing! Congrats, buddy! Well earned! Congrats on 1 million subscribers! Congratulations on the million subscribers! Congratulations on the 1 million subscribers! Matt, congrats on 1 million subscribers! This is such a big milestone and so well deserved. Congrats on reaching 1 million! Very happy for you, and, uh, onwards and Upwards. Let's do it. I wanted to wish you a huge congratulations on reaching those seven digits. I wish you way more success than you could ever imagine, and I hope that this is just the beginning of your journey. Just wanted to wish you a heartfelt congratulations on reaching one million subscribers, big boy. Congrats, Matt! Congrats on one million subscribers! We love you.