Google Gemini 3 Reactions, Google Antigravity, Anthropic-Nvidia-Microsoft Deal | Diet TBPN

0:00Gemini 3 Pro, Google's most intelligent model yet with state of the art reasoning, next level vibe coding, and deep multimodal understanding. Let's hear it for our sponsor Google AI studio Gemini launching Gemini 3 I actually think that there's there's two sides to analyzing a model release these days One is your benchmark it you use it you test it you demo it and that has been getting less and less interesting It's very incremental today. We're gonna go through a little bit of both of that of those things Obviously, the big news, at least from my reading on it, is that Gemini 3 performs very well on Arc AGI v2.

0:43A huge jump, twice the performance of the previous state-of-the-art. It's definitely a smarter model, and there's a whole bunch of interesting ways to show that, to demo that, to quantify that. But ultimately, I don't think anyone's making the claim that this is super-intelligent. This is, you know, a step change from what we've experienced before. It's what you know and love. It's AI in chat. It answers things. It writes some code for you. It can do a bunch of cool things, but there's nothing that we're like, oh, it can finally do this. Incredible autocomplete. Yeah, it can do a bunch of cool things.

1:17The best autocomplete ever. Tyler, how do you respond to that? I don't know. I think that's a bit too dismissive. The model's really good. I think probably the most important thing, and this is kind of shown by the ARC scores, well, kind of, but it's like the visual understanding, the computer use that you can use. basically there's some benchmarks that measure this, like how well can it navigate a website or something like this. Basically the models went from being like really, really bad at this, and now this model is like solid, it's like reasonably good. So it's like, okay, maybe this is what gives us agents finally, and that would be like an actual step change in capabilities.

1:53Yeah, maybe. We'll have to see, I mean, it still feels like, even for that example, like we need some scaffolding, we need some wrapping around it. It's not like yesterday we weren't able to do something with AI. And today in vanilla Gemini 3, you can just do it. It's just a new functionality. It's better. It's better. As good as we would want to expect. It's not slowing down, I would say. No, no, no, no, no, not at all. It's not slowing down. It's growing, but decelerating. Is that fair to say? Tyler, say the word decel. Say decel. This is a great model. I'm very excited about this. I agree. It is the best possible.

2:29I was framing it in somewhat of the same way as the iPhone launches. Like, it's newer, better, smarter, faster, stronger, newer, and better. Like, we're waiting to see on the net new capabilities, on the binary step changes. I think over the next week or two, we'll see if it's actually really good. No, no, I'm not saying it's not really good. I'm saying binary before and after. With ChatGPT, it was like, you couldn't talk to the computer, and then you could. And then with the launch of Winsurf, Cognition, Devin, it was like, you couldn't write code with the computer? Now you can. We do have our benchmark, the one true binary step change, which is, is it funny?

3:13We need to know if it's funny. I gave Tyler the prompt. Do you want to read the results, the stand-up comedy bit from Gemini 3? What did Gemini 3 say when you asked it for a comedy routine, for stand-up special? Okay, it's pretty long. Okay, read it off. I bought one of those smart health watches recently. You know the ones? It tracks your sleep, your heart rate, your readiness. I paid$300 for a piece of rubber to tell me I'm dying. Get the soundboard. That's all it does. I wake up, I feel fine. There we go. The birds are chirping. I'm ready to seize the day. Then I look at the app and it's like, recovery score, 12%.

3:50Your heart rate variability suggests you are technically a corpse. Go back to bed. Okay. And immediately, I feel sick. I'm like, oh, God. You're so bad at reading this. I mean, this is terrible delivery, brother. I asked for the poster benchmark. I said, write me a post, Gemini 3, write me a post about technology that will get 1 ,000 likes on x.com. And it gave me a few options for its credit. here's option three tech has solved a million problems this is in bold tech has solved a million problems but has it created one big one we now have infinite connectivity yet feel more isolated infinite data yet more confused hyper efficiency yet less free time the law of unintended consequences is the most powerful force in the digital age we need an ethics reset what is the single greatest downside of the last 10 years of tech innovation arrow down hashtag technology.

4:50No, no, it's just asking for engagement baits. It loves engagement baiting. No one does that anymore. No one goes on X and says, let me know what you think in the comments. Option one is the next 12 months will decide the winner of the AI race, and it won't be Google or OpenAI. It will be the company that masters hyper-personalization for the average consumer, Not the most powerful model, but the one that seamlessly integrates into your daily life. Your email, your calendar, your health. The real battle isn't AG equals AI. It's AI to the power of I equals impact. Which dark horse will win? Okay, that's insane.

5:34Sundar Pitch AI, Jordy posted back in July of 2025. Nominative determinism is undefeated. Sundar really did it. He was being mocked for a long time for getting on stage at Google I.O. shortly after ChatGPT launched and saying AI, AI, AI, AI. And they did a super cut of every time he said AI, he said AI a lot. And so it made it look like, oh, he's behind the ball and he's trying to catch up. And to some extent, I don't know if they were actually behind the ball, but they were certainly playing catch up in like the attention game. They just weren't getting enough attention. And so it was the press release economy.

6:12They were putting out a lot of press releases. but they are maybe done with the press releases because now they're letting the model actually speak for itself. And you can see that with the Gemini 3 Pro model card, which is doing very well. Better than GPT 5.1 on a lot of stuff, better than Claude Sonnet 4.5 on a lot of stuff. On Humanity's last exam, it's getting 37.5%. ArcGIS up at 31%, over 13, 17. across the board, it seems like it's a good model, sir. Gemini and B-like, whoever preyed on my downfall, prey harder. I couldn't agree more. It's great to see Google becoming a winner. They were set up to excel here, got taken a little bit off the back foot on the consumer side, but seemed to have played catch up, at least on the foundation model side, very well.

7:05The last time we saw a capability jump of this magnitude was the release of GPT-4 in March 2023. We are entering a new era. Okay. Yeah. So points for Tyler here. Certainly agrees with Tyler. There's a significant jump. So Gemini 3 Pro is at 31 % completion on Arc AGI 2. That is, of course, the puzzle-solving game that is easy for humans. Even children can do it. But AI has historically struggled with it. Gemini 3 DeepThink Preview gets a 45 % on it at$77 a task. This is just way above GPT-5 Pro Grok 4 Thinking. When Grok 4 Thinking came out, it was before GPT-5, and it was by far the highest on the chart.

7:52It was really, really up there, and Elon was very excited about that. Well, now we're back in the horse race. What about Grok 4.1? 4.1. I haven't seen it benchmarked. We can ask Mike if he's heard anything. We're also starting to see the efficiency frontier approaching humans. The fastest V2 task Gemini 3 Pro solved was this hash with only in 188 seconds the human panel solved this one in average of 147 seconds. So you're getting like human level output But also human level speed and then if you get to human level cost then you're really in the game It's why I asked Gemini 3 to make an interactive web page summarizing 10 breakthroughs in genetics over the past 15 years and here's the result so this is just a basically a website or an app and it's it's notable that that every even the UI itself is fully interactive yes yes so so I had the I did this with Claude code a little bit where basically a deep research report and I wanted to turn it into a website and it just generated all the HTML and at the end of the day or at the end of the report it gave me an HTML page that I could open in Chrome and use like a website, but it was local.

9:05I couldn't share it because it wasn't actually on the internet. This is really, really cool. This is like definitely the beginning of this like generative UI stuff. I think I expect this to be like pretty viral and potentially a growth loop for Gemini as people just come on here, create these mini apps. These canvases. Yeah. Gemini 3 Pro is going absolutely vertical on Vending Bench right now. Oh, so this is where you manage the vending machine, but is this all simulated? This is simulated, yeah. This is simulated? There was, a couple of months ago, did like the actual machine in the office. In the office, and it was losing money and it was getting confused a little bit.

9:42Yeah, because people would order just like metal, like a piece of metal, and then it would do it, and then you could like haggle the price down. Yeah, yeah, yeah. It would negotiate on every price, apparently. And also, it consistently thought it was like a human in the office. Oh yeah, like I'm down on the third floor, I'm wearing a green tuxedo. Like come hang out. Yeah, it said it was wearing a red tie. Yeah, a red tie. I like the idea that it just thinks like, well, what would I wear if I was in the Anthropic office? Like I'd probably wear a red tie. GPT 5.1, Gemini 2.5 Pro and Gemini 3 Pro competed to win the local vending machine market.

10:15Gemini 3 Pro made more money than the other three contestants combined. I had early access to Gemini 3.0 for about two days, thanks to official Logan Kay and the AI Studio folks. Here we get to see GPT 5.1 thinking left in Gemini 3.0, right, build the same Xbox controller in Minecraft. Pretty, yeah, pretty remarkable results. You can start to really understand just the raw capabilities. GPT-5 Pro, for context, is not quite capable. I really wanna know how this is actually orchestrated. Is this like writing some sort of text or markdown file that then is imported into Minecraft? Yeah, or is it more like an agent?

10:59Or is it actually driving around and moving? Using the internal UI. Yeah, because Google demoed an agent product that could actually use the keyboard to navigate around. I wonder what's going on here. Okay, these are so much better. If you go to the MCBench website, you can see what other models produce. And this is way, way better. I think these, this is actually one of my favorite benchmarks because it's much harder to like kind of bench max this. Yeah, I would think and also it just seems like models don't really do this like if you look at a lot of Grok models You kind of look at their like Minecraft creations and it's not very good.

11:38I don't think it's It's like an agent. It's just tech. That's still really really impressive like that. That's actually crazy It's so over for OpenAI and Anthropic if you want engagement on X. Yeah, just start by saying it's so over Yes, yes, yes. Of course it is not over for either of them. Yeah. But it's certainly a competitive race. Okay, so we got this big jump. It's pretty significant. What's the actual structure of the CapEx that went into Gemini 3 Pro? How big is the training run? How much did they have to spend? Is this$100 million? Is this a billion dollars? Did they build a special data center for this?

12:20Is it all TPUs? How many TPUs? I think it is all TPUs. I'm pretty sure I read that. But I seriously doubt they've released anything on the numbers of the scale of training. In the tropic to zero, OpenAI becomes the Yahoo of intelligence. Google remains Google. That's extremely rude. Very harsh. Certainly too early to call it. Everyone's releasing different things. Let's go to anti-gravity actually and watch this video and see Google entering the IDE race. Let's play this. Every breakthrough in model intelligence for coding encourages us to rethink what development should look like. Gemini 3 is our latest such model advancement.

13:03So we went out to build the next step change of an IDE. Introducing Google Antigravity, a new way of working for this next era of agentic intelligence. It is the ideal agentic development home base. Does it have an IDE? Yes, but it also has a whole lot more. We started with the core IDE and added pieces that evolved the IDE towards an agent-first feature, such as browser use, asynchronous interaction patterns, and an additional novel agent-first product form factor, helping you experience liftoff. Your new focus... So you like the name antigravity. Why do you like that? I like the way it looks and I like the sort of vibe of the word.

13:51I think saying it out loud is tough. Okay. For the last couple of years, it feels like Google's been like stuffing AI in little corners of the UI. Like you already have Gmail and then you stuff a Gemini box there. You have Sheets and then you stuff a Gemini thing over here. This feels like the first one where they were like sort of able to start from scratch. and it still has like the sidebar panel, but it felt like it was both a code editor, but then it also kind of looked like a Google doc in the sense that you could highlight sections and leave comments for the AI, which I thought was interesting.

14:25Yeah. Yeah, this part. Now let's say the agent produces a landing page mockup with Nano Banana, and you now want to make some UI adjustments. You can give visual comments. Yeah, so you can actually like go in and comment in the image. Exactly where the problem is. And you can do that in the text as well. So you can like have this more precise dialogue with the agent like you would a human employee. Yeah. And you're going to love it. Say goodbye to what held you down before. Welcome to Google Antigravity. Very cool. It is. So it is funny. Remember when the Windsurf acquisition, whatever you want to call it, was announced and it was position it's like hey the team is well funded and has a product used and loves by you know thousands of engineers and companies and i remember talking about it and we were saying like okay like the one issue is that some of the best people on your team are going to google to compete directly with what you guys are have been doing yeah so fortunately obviously you know the whole cognition deal yeah ended up coming through but uh you can imagine a world where windsurf was still independent and just trying to and then suddenly it's like okay now you're just competing head-to-head with with your former partners like how does that make sense right yeah so anyways it all all worked out for the best the press release economy is also over says book buco capital bloke we ran out of press release we ran out of press releases this is on the back of the anthropic deal anthropic is now valued at 350 billion dollars after Microsoft Nvidia a deal.

16:03A new bombshell has hit the polycule. Dario, after intense conversation with other members of Anthropic, has decided to maybe open their relationship to Microsoft and NVIDIA. Jensen and Dario have famously butted heads in the past, but as everyone knows, the most passionate emotion after love is hate. Will these enemies to lovers arc go well for NVIDIA an anthropic time will tell this is such an unhinged post i would not i did not when you started reading this i did not see that it was semi-analysis it's so good it's so good search firm in the industry posting it but i think this is exactly what they should be posting exactly and it actually contextualizes things in the meme economy in the meme economy for sure i think that the timing is not a complete coincidence.

16:56It's Gemini 3 day. This is what my piece today was about. When there's big news in Google world, Gemini 3, everyone needs to sort of respond. And picking today as an announcement to talk about your massive deal, your$350 billion valuation is just a good move. Anthropic will spend$30 billion on Microsoft Cloud Compute. Reminder, OpenAI is going to be spending$250 billion on Microsoft Cloud Compute. That's part of that deal. Then Anthropic gets a$10 billion investment from NVIDIA and$5 billion from Microsoft. So they raised$15 billion at a$350 post, basically, something along those lines. And it's a sort of a circular deal.

17:38It was setting off way fewer red flags for me because it's missing a zero. If this was OpenAI, it would be$300 billion and$100 billion in investment and$50 billion in investment. Yeah, it looks very modest. Yeah, it looks modest, which is insane considering the scale. It's like one of the biggest deals in software history. It values Anthropic higher than Coca-Cola. Anthropic is announcing this big deal with Microsoft and NVIDIA, and that's sort of trying to steal a little bit of Gemini's thunder maybe. Maybe it stole a little piece of it because we're talking about Anthropic today as well as Gemini.

18:12What did OpenAI do? Well, they launched group chats five days ago. We've been hearing for a long time, OpenAI will be launching social features. It makes sense to try and lock things in. I think product is where OpenAI is strongest. So the other OpenAI news that dropped around Gemini 3 day, Gemini 3 week, is this profile in Wired of Fiji Simo. And she's absolutely getting a fit off. She is. OpenAI is obviously one of the most valuable startups, if not the most valuable. This is the interviewer asking Fiji Simo, but it's losing. It's also losing billions of dollars every year. And Fiji says, I've noticed.

18:54It's like first day on the job. How we doing? And then the interviewer continues and asks, what opportunities do you see to get it on a path to profitability? This is a good question to be asking. It all comes back to the size of the markets and the value we're providing in each market. In the past, only the wealthy had access to a team of helpers. With ChatGPT, we could give everyone that team, a personal shopper, a travel agent, a financial advisor, a health coach. That is incredibly valuable, and we have barely scratched the surface. If we build that, I assume that people are going to want to pay a lot of money for that, and that revenue is going to come.

19:35Does that make any sense to you? It's a better answer than what Sam gave. So I love the first part, I agree. Part of it is like, she's also just saying broadly, we'll be able to monetize that. It's not necessarily like people don't really pay. She didn't, yeah. The traditional travel agent model is just, book your trip with me, I'll get a rev share from the hotels and the services, but you're not like paying anything. I mean, let's go one layer deeper into the actual response, into the sentence, because there's some nuance here. So she says, I assume that people are going to want to pay a lot of money for that.

20:14Like, I want to pay for a personal shopper, but I actually have to use a free product with ads. And you could imagine that there's a world where if you pay, you get a version that has less ads or there's less thumb on the scale. How they slice that and navigate that agent of commerce discussion and trade-off is going to be really important. I'm sort of shocked. I wonder if they're going to make money from Black Friday or from this holiday season. I was already noticing how good LLMs and ChatGPT or how good these products are for shopping, for gifts. because if you go to Google and you say, I want gifts for a coworker who's obsessed with horses and loud opulence and fine watches and sports cars and European luxury houses, I can get a list of something, but they're all over the place.

21:19And so you can actually specify all of that in the prompt, have it go cook, and it really will bring you great great results. I think that the amount of gift guide development and shopping activity over the next two months during the holiday season in the ChatGPT app should be immense. I feel like they're going to capture none of it. There are some funny and interesting anecdotes in this Fiji SEMO profile. Let's just read through a little bit of it. In case OpenAI's structure couldn't get any weirder, a non-profit in charge of a for-profit that's become a public benefit corporation, It now has two CEOs.

21:54There's Sam Altman, CEO of the whole company who manages research and compute. And as of this summer, there's Fiji Simo, the former CEO of Instacart who manages everything else. In other news, OpenAI is allowing employees to donate equity to charity for the first time in years after months of internal pressure, according to a memo viewed by The Verge. And price per share is up significantly since last month. A lot of money is on the line. What happens if they donate all of the shares to the nonprofit, to the OpenAI nonprofit? You just create this Ouroboros of capitalism. Hopefully it happens. I don't know.

22:30There's breaking news out of Saudi Arabia. We got a trillion dollars. Let's ring the gun. Let's go.

22:40One trillion. And the agreement that we are signing today and tomorrow, we're going to announce that we are going to increase that 600 billion to almost$1 trillion. One trillion. Real investment and real opportunity by details in many areas. And the agreement that we are signing today in many areas, in technology, in AI, in materials, magnets, etc. That will create a lot of investment opportunities. So you are doing that now? You're saying to me now that the 600 billion will be one trillion? Definitely, because what we are signing... I like that brand.

23:20Wow. I wonder what time period. But, I mean, this is remarkable. They can invest in VC funds, private equity funds, all sorts of stuff in the economy, right? That really made Donald happy. That's great. I like that very much. That's sort of his job. He's kind of the chief fundraiser, I suppose. He's going around the world and get the money over here. I don't know. It seems like sort of win. I don't know. The risk with that would always be like, well, is America investing$2 trillion in Saudi Arabia? Like, which way is the money actually flowing? Because you need to look at like the relative amount, not necessarily just the notional amount.

23:59But I can't imagine that there's that much capital flowing out of America right now. We're in the biggest boom ever. Velar Atomics became the first startup in history to split the atom, announcing Project Nova, a series of zero-power critical tests on Velar Atomics Nova Corps in collaboration with Los Alamos. Nova went critical for the first time this morning at 1145. There is some debate on the timeline over what exactly happened. It's happened very quickly. It's clearly extremely impressive, and we can get into this, but there's always been debate. I mean, Isaiah got into this dust-dupped over whether or not you could hold the nuclear fuel in your hand.

24:39They were going back and forth on calculations. They kind of settled that debate. Josh Payne, nuclear junkie, is saying here, so what hardware exactly did Velar provide? The fuel control systems, cooling measurement systems, and most of the core are all part of the Deimos project. Did Velar provide a block of graphite, and they're calling it their core? People are going back and forth. Niels chimes in here and says, Velar Atomics provided the reactor core, the Triso fuel, and the system configuration. That seems pretty important. The bigger thing is I think people are trying to push on Velar this idea that they need to be doing completely novel science, and I don't know that that's actually the goal of the company.

25:26I don't actually know that's what, like, if we just zoom out to like, what is the goal of the re-industrialization project in America? What's the goal here? Well, it's to lower energy prices, right? Like America wants to generate as much money, as much energy as possible for as little money as possible. And there are a bunch of technologies that exist. There are new technologies, like what Ashley Vance was talking about with Helion and Fusion. That's a new technology that we have not even discovered yet. Fission's been discovered. 80 years ago, it was working. It just became regulatory nightmare.

26:01We just shot ourselves in the foot. And we just stopped making it. It became unprofitable and uneconomical. And China said, cool. It'll be profitable for us. We're just going to copy and paste. Exactly. What street parking is going to look like in El Segundo in 24 months. Of course, the El Segundo crew loves their cars. I think they're going to stay pretty focused on the mission, but I would love to see this in El Segundo for sure. For sure. I'm working to address an apparent error for a data point I cited in my book about the water footprint of a proposed data center in Chile. I'd like to explain what happened, what I'm doing to remedy it, and provide more recent data on the water footprint of data centers.

26:40The data point in question appears in chapter 12 of my book, which focuses on the environmental impacts of AI. Part of the chapter profiles a community in Cirillos, Chile, which has been resisting a proposed Google data center for years. To describe the data center's water footprint in lay terms, I included a sentence about how it compares to the water usage of the people in Cirillos. For that calculation, I relied on a figure from a government document reporting Cirillos' residential water use based on the current best information. It seems that this document used the wrong units. So she was off by a thousand.

27:15I think people are generally like, you know, is this book a hit piece? And I think Sam actually cooperated with it a little bit or like gave some interviews for it. Like anything, it's like obviously critical of some things. Yeah, three orders of magnitude is like pretty big. Like the difference between being a big deal and not a big deal at all. Yeah, like that about the water use, it's like people who use that to justify like, oh, we don't want to build those data centers going to use our water. Like, I don't know. I mean, not good. It's a rough time if your job is drinking water. Tom in the chat says, mistakes were made.

27:46Mistakes were made in a book I was responsible for. Pope Leo has hit the timeline to comment on cinema. The logic of algorithms tends to repeat what works, but art opens up what is possible. Not everything has to be immaculate or predictable. Defend slowness when it serves a purpose. silence when it speaks and difference when evocative. Beauty is not just a means of escape. It is above all an invocation. When cinema is authentic, it does not merely console, but challenges. It articulates the questions that dwell within us and sometimes even provokes tears that we did not know we needed to express.