AI 圈内人拆解 GPT-5 更新及其对 AI 竞赛的意义：与 Emad、AWG、Dave 和 Salim 对谈

Hey everybody, welcome to another episode of WTF Just Happening Technology. I'm here with my Moonshot mates, Salim and Dave, and two special guests—geniuses. Dave, would you introduce Alex Wissner-Gross? I think that would be important.

Alex—yes, I'd love to introduce Alex. Genius is probably a good word. He has degrees in math, physics, and computer science from MIT; he's a true polymath who understands everything, and we're going to talk about a lot of it today. He has a PhD in physics from Harvard, in addition to that, and reads literally every research document and every breakthrough in AI and many other fields. So, it's always incredibly informative to have him.

Welcome, Alex. Salim, would you do the honors with Emad?

Sure. Emad is one of those folks where, every time he says something, you have to take twice the time to parse what he just said and make sense of it. There's more intelligence per word density than in most people you've ever met. He's a founder of Stable Diffusion and Stability AI, a former hedge-fund quant, with a brain the size of several planets, and is building, I think, a systemic layer for the next version of the internet with crypto built in, which I think is really powerful. So, welcome, Emad.

First of all, I literally just landed from a week in Portugal, so my head is still spinning after a 12-hour flight. But hey, what could possibly go wrong?

Today, we're speaking about 2 or 3 special events from this past week—in particular, the announcement and launch of GPT-5 and the continuation of the AI wars. But before we get there, Salim, I think you've recently gone through surgery. Is that right?

I had shoulder arthroscopy, where they drill 3 holes in your shoulder and do kind of an oil, lube, and filter on it. I had a bone spur impinging on the tendon, et cetera. What's incredible with the advances in technology today is that I was in and out in about 2 hours. It's unbelievable that they can go that deep into your body and then you're just out again. It's amazing.

Don't forget the access holes.

I was going to ask about tennis this weekend. I guess we're not playing, huh?

No, not for a little bit. We'll leave that for another time.

This is a special episode because I'm filming in the new Moonshot podcast studio. Check out the background. I hope you like it. It's a real background, and we'll be doing a lot of episodes from here in the future.

Emad, you're in London and it's midnight or something like that.

Yeah, it's just time for the brain to get going.

You're amazing, buddy. And Alex, at that time, maybe his brain slows down a little bit so we can understand everything. That's my hope.

We'll find out.

Alex, you're in Boston. Where are you today?

That's right—Cambridge, Massachusetts.

The center of the known universe, at least for us MIT alums.

Certainly the center of Cambridge.

Let's dive into this episode. I'm going to start with this note. Sam Altman made the announcement of GPT-5 2 days ago, and in particular, this is the quote that stuck out: “GPT-5 is a significant step on our way to AGI, which also means it isn't AGI yet.”

I have a question for you guys. We also saw that the day before this announcement, Sam put up a tweet showing the Death Star. I have to ask: I don't get it. Why would you put this up? Is it to get views or to get people really worried?

A lot of this launch was pretty uncoordinated, but Kevin Weil also posted something with Elmo and a fire behind him, saying, “You know, it's coming.” So, there was a lot of pre-event tweeting and buzzing on X about something huge coming. I don't know why a Death Star, but a lot of people talked about it already.

It doesn't feel like a great look. You're trying to get people accepting and happy about the future, and you show that imagery. It's kind of like, “Okay.”

One of the Google people posted the Millennium Falcon and said, “No, we're meant to be the rebels.” He said, “This is meant to be from the point of view of the rebels.”

Oh, okay. There you go. That makes a lot more sense.

And everyone's like, “Nah, that's not the case.”

That's way too subtle.

A huge amount of expectation was placed on GPT-5. I'd love to ask each of you: What do you think of it? What do you think of the announcement? It was a little over an hour. Let's start with Emad. What do you think?

It was kind of in line with what I expected, because when you're doing an AI for 700 million people, it's very difficult to do a mega-AI. We'd been guided to expect a multi-routing type of thing, from mini up to pro, and that's kind of what we saw. It's basically o4 but with one front layer, so I thought the announcement was okay.

The expectations are so high now, particularly when you build it up, that you have to keep beating them every time by more than a little bit. I think we all thought it would beat expectations, but the question was: by how much? It was like, “Okay, wasn't it?”

Alex, how about you, buddy?

I tend to think the real net impact of a launch like this is more about lifting hundreds of millions of users up from a model like GPT-4o to a frontier model. I think the changing economics from a radical cost reduction in frontier models will be one of the long-term impacts.

To the extent that there were expectations of an ontologically shocking moment, when new qualitative capabilities would come online, I tend to think that ultimately lifting hundreds of millions of new users to frontier level and getting them to interact at scale with a frontier model over the long term will be just as impactful, and just as economically relevant, as introducing some jaw-dropping new qualitative capability.

I hear you, and that is true. That's what Sam's mission was: to deliver a single user interface that enabled you to do quick answers or long, detailed research and coding.

Salim, do you remember? You and I were together up in the Bay Area, with Dave in Boston, when Google I/O came out, and there were so many holy, holy, holy moments when Google I/O was showing its capabilities. What did you think about this one?

I had the same reaction as Emad, which was, “Eh, it's not 10× better than what was there before.” I think I'll concur with Alex, though, in terms of the real power coming from the cost drop, which will make it much more accessible to a lot of people.

Downstream, in a couple of months, as people start building applications, GPTs, and special agents on top of this, we're going to see some really big surprises, which I'm looking forward to.

Let's close it out with you, Dave. You've been thinking about this and watching all the telltale signs for a while. Were you excited, impressed, or depressed? What was it?

You called it right, Peter. Compared to Google I/O, which had incredible showbiz value and a ton of computer-generated video, OpenAI decided, for whatever reason, to go folksy, make it look like a high-school presentation, and feel startup-y. I don't know if they'll stick with that. Steve Jobs did the best showbiz in the history of the world, and the anticipation of this launch was up there with the top 3 product launches of all time.

You have an opportunity to really blow people's minds. Either they didn't have time to work on it, they don't have that staff built up yet, or they just don't care. Maybe—I don't think that's the case—but they really did not put a huge amount of effort into this event. It came through, and you'll see some data that supports that. It's pretty obvious that it came through.

Let's look at that. This is the view on Polymarket, and what we see here is the answer to the question: Which company has the best AI model by the end of August? Coming into this, OpenAI was riding high, with Google coming in second and Anthropic in third. Then we see the timestamp for when the release went live. Any commentary, Dave?

It's great that Polymarket exists, because I think we all watched it live here in the office. Alex actually suggested it. It was phenomenally cool. Watching the ticker in real time, there's a dip when they did their first coding demo, and then a huge plunge when they did their second coding demo.

And literally, the betting markets went from an 80% chance that they’d have the best AI in the world—not just at the end of this month, but also at the end of the year—to completely inverting and saying, “No, Google’s going to have the best AI at the end of the month and at the end of this year.” And I think they actually showed some incredible capabilities and rolled them out at a ridiculously great price point, but the market reaction to it was, “Wow, I think Google’s going to eat your lunch.”

So, yeah, you can’t deny it. It’s right there. People are putting money behind this prediction.

You know, one thing I just want to point out for folks listening—and I think it’s true—is that when you have this huge expectation of GPT-5’s launch or any of these new models, when Grok 4 came out, at the end of the day I sort of felt a sense of underwhelming. And I think it’s not because it’s not impressive. It’s because we’ve become so desensitized to extraordinary progress, right?

I think there’s something else here, though, that I’m really enjoying, which is that, given the closeness of the different models, it means it’s likely that we won’t have one runaway success. And that means you have a very competitive market, which is just good for consumers overall for the time being, and all the models will do incrementally better over time. So I’m excited by the fact that there’s not one breakout.

Sure. But I do think it’s important for folks to—let’s talk about the desensitization for a second—because folks who are listening to this have to realize that our expectations are getting so high. Every time there’s a new rollout that has additional capability, it’s like, “Oh, that’s not so impressive.” But compared to what existed a year ago or 2 years ago, it’s extraordinary. Emad, do you agree with that? What are your thoughts?

Yeah, I mean, it’s hedonic adaptation, right? When you get into a Waymo for the first time, it’s great. The second time, yeah. And now it’s just a whole experience around this.

I think that part of it was just the communication, though, because, as you’ve noted, Grok 4 was a good model, but we see people getting wireheaded, hallucinating, and all sorts of things. Lifting that up to a better base level should have been the communication, with practical examples, but they didn’t really show that. Again, I think the communication was a bit off in showing that lifting of the floor.

The other thing that I think is that, for the first time, what we saw was that there’s a big gap between what the consumer gets and what the lab has.

We actually saw a few OpenAI people say that, before this came out, we had Horizon and Zenith as the 2 models on LMArena, where you compare secret models against each other. They chose to release Horizon, but Zenith was better. And OpenAI has admitted they have better models internally as well, even before the next cluster build-out.

So, they’re pulling their punches.

Yeah. Because it makes more sense, as you head toward AGI, to actually not release the best model to everyone, particularly because it’s more expensive to inference. GPT-4.5 was so expensive, and that was their frontier model at the time, but it was too expensive for anyone to use for normal tasks for the 700 million people. For the genius tasks, you don’t want to give someone else that AI; you just use it for yourself to outcompete everyone else.

So I think we’ll see that bifurcation of decent models for everyone, for everyday tasks for 700 million people, and then you make $700 million using the other model, because it’s the only logical thing to do.

Yeah. One of the reasons I was so disappointed by the lack of really compelling demos and showmanship yesterday is because I’m constantly trying to make more people aware of how much change is coming, how insanely important and imminent it is, and how much they need to rethink what they’re doing tomorrow. And I was hoping to get some ammunition that I could actually just forward and use.

They managed to make one of the biggest turning points in history—the history of humanity—kind of boring.

I mean, maybe it was deliberate because they had the charts that were completely wrong as well. Maybe it’s just all deliberate in that, look, you don’t have to worry too much about this, right? No, that is a theory that is a viable theory, actually, because all the accelerationists, including me and Alex, know that a lot of this is being used internally for self-improvement—a lot of the compute, a lot of the capabilities—and it could be that it was intentional.

Don’t scare the world, don’t scare—

Don’t scare the world. Well, I mean, yesterday when GPT-5 came out, GPT-5 was a router model, so your thing goes in and it routes it to Thinking, Mini, or Nano, depending on something they said. Well, it was actually broken for about 24 hours, and you’re like, “Really? You released it and then you just left it broken?” The routing was off, but being broken is also a great way to actually gather data.

To do the model improvement. And they discussed this flywheel of data improvement. So again, I think we see this bifurcation now, where most of the announcements by OpenAI are likely to actually be very consumer-driven, very floor-raising, and I think we’ll see less and less of the big, massive stuff, apart from the outputs, like, “We’ve had a breakthrough in something or other,” but not generalizing that.

I’m still waiting to see what an AGI or ASI demo would look like or feel like. I don’t know, but we’re going to find out.

Don’t get me started. Move right along.

All right, let’s turn for a bit to benchmarks. When I was having the conversation before this podcast began about whether we should talk about the benchmarks and whether it would get old, Alex, what was your comment about the benchmarks?

Riveting. Some of these benchmarks, Peter, are absolutely riveting. We are so spoiled. We’re lifting hundreds of millions of people to the frontier level of these models. We’re collapsing costs. The economics are collapsing by an order of magnitude. And here we are complaining, “It didn’t demonstrate any ontologically shocking new capabilities.” How spoiled we all are.

We have gotten spoiled. Let’s jump into the riveting benchmark. So, Alex, since you’ve got the floor, let’s begin here. GPT-5 debuts at number 1 in LMArena. So first off, what is LMArena?

LMArena is—and I think we discussed this in the last episode a bit—a crowdsourced benchmark wherein the community, the internet at large, is able to interact with competing frontier models in a variety of ways. The ranking that we’re seeing here is focused on text-based interaction—conversations. There are other scores that deal with web development and other modalities.

What we’re seeing here is GPT-5 leapfrogging over the rest of the leaderboard to number 1 in text-based interaction. There’s another parallel benchmark with web development where you see an even larger margin, a larger difference in Elo scores, between GPT-5 and the next-largest, or the next-strongest, competitor.

And this is remarkable. Again, we’re so spoiled to see these leapfrogging capabilities every 3 months or so. It could get even faster. But this is going to be transformative in terms of everyday conversations that hundreds of millions of people have—

Software development and a number of other domains.

So, can I ask you, Alex, how do you reconcile this chart with the Polymarket chart? Does that mean Google will again leapfrog this before the end of the year?

I would say, to the extent that Polymarket is indicating a prediction—a rational prediction—about the market, and I think that was set for the end of August, I would interpret that market movement as a prediction that Google will launch a new frontier model by the end of this month.

Every expectation. And we’re going to see in a little bit how much Google has done. I mean, they’ve been extraordinary under Demis’s leadership. Here’s the next one, and I’m going to turn to you, Emad: ARC-AGI-1, and we’ll see ARC-AGI-2 in a moment. The leaderboard here—do you want to give us a dissection of what we’re seeing here?

Yeah, this is kind of very, very hard tasks that are meant to indicate progress toward AGI. Grok kind of led the way there, as you can see. Which one is that?

It’s Grok 4 Thinking, right?

And so this kind of Pareto frontier is about solving these very, very complicated tasks versus the cost. And o3 was actually really, really good, but it’s way out there in that it’s far more expensive.

GPT-5 has different levels: the high, the medium, and the low. It doesn't quite beat Grok, which is also the case for other benchmarks like Humanity's Last Exam. I think this was part of it: we see better performance on GPT-5 for everyday stuff, and it just has a lead on some of these, or is up there. I don't think they wanted to blow everyone's socks off, because remember, they also have models that scored gold medals at the IMO.

Mhm.

Gemini, for example, recently had Deep Think. That's a new test-time model of their version that scored a gold medal. I think, apart from xAI, who are trying to do the best they can on all these benchmarks and release the best they can, we're starting to see some punches being pulled at the top of these benchmarks on the AGI side, on the super-genius side. I think we'll see a bit more clustering up there. Alex, would you agree?

I think there are 2 ways to look at this chart. One, as Emad said, is which point in the scatter plot, which is plotting cost versus score, is at the top of the chart. That's one way to look at it. The other way is: what is the cost frontier? What's the Pareto-optimal frontier where you get the best score, or the best performance, at a given cost?

There, if you look just a bit to the left, you see the GPT-5 Mini series and, to the lower left of that, the GPT-5 Nano series have set—have defined—a new frontier for cost performance. I think the buried headline here is the hyperdeflation that we’re seeing in the cost of intelligence, which ultimately, I think, ends up being even more transformative than just narrow capabilities at ultra-high cost.

You could run the thought experiment: What would happen if we could build superintelligent computers so unaffordably that human civilization can't afford them? Compare that with what happens when intelligence is too cheap to meter, so that everyone can afford it. I think that's the central discussion.

I think you'll see Google and OpenAI compete on that left-hand curve, effectively.

All right, here we see the ARC-AGI-2 leaderboard. Emad, why don't you lead us off on this one?

Yeah, this is just a more complicated version of ARC-AGI-1, because they're worried that o3 might saturate it. Again, I think, as Alex said, you see the same thing with GPT-5 on the left-hand side, kind of keeping that as just a more complicated version of the previous one.

All right, moving along. Here's one that I think we discussed, Emad, on one of our previous episodes—or I'm not sure it was with Alex—that how we benchmark these frontier models is going to start to saturate, and understanding how these frontier models actually become economically useful and how they're able to solve grand challenges. So here we go. This is a look at economically important tasks. Emad, want to take a shot?

Yeah, I think this is the year where you break through that line, effectively, or reach that level of performance. There's another chart, I think, we have from METR, which shows the length of tasks this can do, and GPT-5 is right at the top of that. It can do tasks in law, logistics, and sales really well for a long time without supervision and with lower hallucinations, which is the other big news they had around this.

So they actually become genuinely useful. They released ChatGPT Agent, which you could just set off, and it will look up the internet and do all sorts of stuff. A little while ago, it wasn't quite good enough, but soon it will be. Once that happens, this is when you see real big things happening. Either a productivity boom or the inverse—people getting laid off—and we're not sure which of those 2 futures is going to happen.

Again, you can see you're just reaching that level now across just about everything.

Who wants to plug in on this one? Salim?

Well, this is what I mentioned. There are 2 or 3 really big things here, right? To Alex's point, the cost drops of running these models mean we can do a ton. To Emad's point, they're taking out the hallucinations and cleaning it up.

Even though the top line is not amazing, it's a lot more rock solid. Therefore, the agents and applications that build off these things will be very, very solid and stable going forward. I think that's where we'll see some amazing use cases coming out as we apply them in industry.

How should our listeners be thinking about this? Do they think of it from a point of view—

Well, if you're running a business, this is a time to really build, dig in, right? Before, you didn't know quite what you were going to get. What you're going to see going forward now is that it's pretty reliable, pretty solid. Go all in. If you haven't, you should be doing that anyway. Just go all in and start turning your business into an AI-native business.

Yeah, the problem I run into all the time is, as AI is getting better and better and better, the benchmarks get harder to interpret. In the early days, it's all just pretraining: This is 100 billion parameters, this is 500 billion, this is a trillion. It's getting bigger; as it gets bigger, it gets smarter. The benchmarks are nice and simple.

Now, post-training has become very important, but chain-of-thought reasoning is dominating.

It's just such a huge factor. It makes it much harder to track what's working and what's not working. The danger there is that people get paralyzed when they should be getting motivated, just like Salim said. And that's a challenge, actually.

A benchmark like this is vague, and it's a little bit difficult for people to take this benchmark and translate it into, "Should I start an AI law firm? Should I use it to work on discovering fundamental physical properties? Is it going to be good at materials science?" It's getting harder to make those predictions. But that's—

Of course, the answer to all of those is yes, you should.

Yeah. I like to think, in jobs, we teach people to be like machines. Obviously, the machines are going to do it better. If you look at the HealthBench scores, for example, on hallucinations—and hallucinations in general—I think something like 6% to 12% of all diagnoses are incorrect. AI has just dropped below that level now.

Close to 30% if you go to a primary care doctor.

Yeah, it kind of varies, but it's a lot. AI now makes fewer errors than humans, I think, just over the last month.

And again, that's going to be the most errors it ever makes.

Yeah. We'll go into this a little bit later, but there was an interesting study that said physicians by themselves do about 80%, physicians with AI models together do about 90%, but AI models by themselves were doing about 93%. That means the human pulls back and enters lots of bias into the answers.

When I was chatting with doctors about who was going to do my surgery, I came across a guy and I said, "How many of these shoulder arthroscopies have you done?" He said, "About 10,000." I said, "Okay, you're more like a robot than anybody. We'll go with you. We'll go with you, because I want that consistency."

By the way, that is the number one question you should ask a surgeon when you're interviewing them: How many times have you done this surgery this morning? Right? Because you're basically training the neural net of the surgeon by seeing every possible case. Of course, we're going to end up with robotic surgeons that can see every part of the spectrum and have had not just 10,000, but millions of cases.

You just don't want to be the 50th one that morning. That's all.

All right, here's our next benchmark: GPT-5 sets a new record in frontier math. Alex, I'm going to you on this one, buddy.

Yeah, I think this is perhaps the most exciting benchmark to come out of GPT-5 in the past 24 to 48 hours. So what's exciting here? If you look at the performance of GPT-5 High in the lower right-hand corner, FrontierMath Tier 4, FrontierMath Tier 4 is a benchmark that measures the ability of AIs to solve problems that would take professional mathematicians sometimes weeks to solve, but nonetheless problems for which there are known answers.

We're starting to see increments on FrontierMath Tier 4 that, if you extrapolate them—and I've gone through this exercise, and it's a running discussion between me and the folks at Epoch AI—if you project this forward, again by the law of straight lines, by the end of this year, we're seeing frontier AI starting to reach 15% to 20% of hard math problems being solvable by AI.

Project that forward another year, so by the end of 2026, you get to 35% to 40% of hard math being solved. Project it forward to 2027, end of year, and you get to 70%. So what I think we're staring at is a slow-motion solution to math.

That's one of the reasons why I think there's just a rewriting of all math, or at least all math as currently understood in the summer of 2025.

Isn't that amazing? I completely agree. It does play into Emad's theory that maybe they slow-played it intentionally, because if you were to ask me, "Hey, what happened yesterday?" they're crushing this benchmark relative to any other model.

They cut the cost of AI at least in half, if not more. And they caught up to everybody else in coding. If they had just said that in 2 minutes, that would have been the Death Star moment. Yeah, just do that.

Wait, can I drill into that just for a second? Alex, when you say it can solve math, can you give a specific example of what that looks like? I struggle with that.

Better than 800 on your SATs, I guess.

What’s a specific problem, class, or area where you could say it’s done something interesting?

Yeah. No, you can look at the Epoch AI website for FrontierMath Tier 4. It lists example problems that have been published. These are hard problems in number theory, analysis, and algebraic geometry that would take a professional mathematician weeks to solve, but are being solved over the course of a short benchmark by GPT-5.

Okay.

It also raises the question, “What does this look like in practice?” Say the dog catches the car and we actually get AI that achieves superhuman performance in math. I think it’s a profoundly different world.

It’s hard for people to grasp because not everybody’s a mathematician and not everybody’s an engineer. But the way a lot of things get designed, built, and created in the world is that you run into problems and immediately look them up in these massive books and tables: Has anyone ever solved this before? If the AI is continually solving and archiving all of these mathematical capabilities and making them available, then engineers and algorithms can just find them and use them—plug them in and go. It’s the same in coding: huge libraries of solved problems and solved modules that can be assembled to create things very, very quickly.

I want to close on Emad here before we move on, just because we have a lot to cover still. Emad, closing thoughts on this one?

Yeah, I mean, it’s an improvement over o4-mini. Again, we had the IMO gold medal from OpenAI, whereby they had a verifier on the other side of their model, and they said that just by extending the RLVR of GPT-5, they got a gold medal. So this model can get a gold medal. It can go even higher if you push it.

From the last few days of doing some pretty advanced math, I can say that GPT-5 High is probably the best math model out there. But the really crazy thing is, I think it’s getting to the point now where the solutions to math won’t be complicated; they’ll be really elegant. That’s how we typically see breakthroughs. People are thinking giant supercomputers and lots of work.

But most of the advances that we’ve had in science and math have actually been very elegant.

And if you can do a million different things at once, then you can maybe find some of that elegant theory under all of this. That’s what’s going to be a big leap. And if more and more people can do that now, because the mini, medium, and high are actually at the same level—which is crazy—then you might have a lot more mathematicians, and the humans and the AI can figure out what that elegant theory is.

All right, I’m going to dive into a bit of video here. This is labeled “Let the vibe coding begin.”

GPT-5 is clearly our best coding model yet. It will help everyone, even those who do not know how to write code, bring their ideas to life. So I will try to show you that. I will actually try to build something that I would find useful: a web app for my partner to learn how to speak French so that she can better communicate with my family.

Here I have a prompt. I will execute it. It asks exactly what I just said: “Please build a web app for my partner to learn French.” I can simply press “Run code.” So I’ll do that and cross my fingers.

Whoa. Oh, nice.

So we have a nice website. The name is “Midnight in Paris Together.”

Super romantic.

We also see a few tabs: flashcards, quiz, and “Mouse and Cheese,” exactly like I asked for. I will play that. So this says, “Luca.”

All right, I’m going to pause it there. Commentary. Dave, what do you think about this?

This is exactly when Polymarket plummeted. I’m so glad you captured that clip because the audience is looking at this and acting like, “Wow, didn’t this blow your mind?”

There are only 2 types of people in the world: people who don’t give a crap about this and people who already do it. They’ve been doing exactly this with Claude Opus or Claude Sonnet 4 on Max. They’ve been doing this for like 4 months. So it completely missed the mark, even though it was the best-presented part of the presentation, purely because it didn’t show off the new capability or the new abilities.

But the ability to do this in ChatGPT—in other words, a single model that allows you to do everything—is really big news. Yeah, I mean, if I’m an investor in the upcoming round, this is really big news because Anthropic generally claims to be the leader in coding. Most of the people who do heavy-duty coding lean on Anthropic, and they completely caught up in this release.

That’s a very, very big deal because not only are you good at everything else, but you’re actually as good as Anthropic in its wheelhouse.

Yeah. The question that was coming out was, “Is this an Anthropic killer?” Right?

Yeah.

Here’s my question about this. You generate this webpage, this web app, right? But if I’m a language startup and I want to launch that actual product, there’s a huge amount of backend work I have to do to make it system-integrated, integrated with Stripe, and so on. We’re finding that’s where all of the work is going. Therefore, if this generates a frontend that looks good, like a frontend prototype, is it actually doing that much behind the scenes, or is there still a lot of work to do? That’s the question I have for the folks on the panel here.

Here’s my question for you guys: How should someone listening to this who hasn’t played with ChatGPT-5—let’s call it that—play with their own vibe coding on this? What’s their first step? What do they do?

I would encourage everyone who has ChatGPT-5 Thinking access, in particular, to create a game. I think this is one of the simplest exercises. You’ve always wanted to create a long-tail application, a game, or an interactive app of some sort, but you don’t have coding experience. Go and ask ChatGPT-5 Thinking to implement a new app for you—a new game, a new something—and let it rip. Do it right in Canvas.

There’s a Canvas button right down on the little navbar search bar at the bottom. Click the Canvas button and do it right there locally; it’s much more convenient. They’ve added a lot of capability inside Canvas, so you can just build an entire game for yourself right there. Just go to ChatGPT.com and do it right there.

Amazing.

Yeah. Yeah. I think the performance isn’t quite there yet versus Replit, Lovable, or Bolt, which do everything, including all the other integrations. But again, these things all verticalize very quickly.

Let’s move on here. We saw the co-founder of Cursor come onstage and spend time with Greg Brockman, the OpenAI president. Dave, what do you think about this? How important was this?

Well, incredibly important. It was not just a little time; it was a huge amount of stage time in one of the biggest livestreams in history. So it was very important.

Of course, what happened is that OpenAI was going to buy Windsurf and essentially attack Cursor with an incredibly powerful competing product that’s virtually identical in functionality. Actually, here in the office, about half use Cursor and about half use Windsurf. They look virtually identical. And so that deal fell apart. Microsoft torpedoed it because of the intellectual property rights that Microsoft would have.

So they torpedoed the deal. Here we are just a couple of weeks later, and OpenAI is now saying, “You know what? We’re going to work very closely with Cursor. We’re going to give them a lot of stage time.” I think what we’re starting to see here is the alignment between the coding companies and their LLM partners.

Previously, everything connected to everything, so any LLM was available through any coding platform. I think going forward, it’s very likely that Cursor works closely with OpenAI. Windsurf is now part of Google—or sort of part of Google. Half in and half out. Microsoft wants VS Code, and they want to build their own thing.

You’re going to see this vertical alignment. Already, people all over Twitter, or X, are saying, “When I use it through its native platform, through the Canvas, it works much, much better than if I try and select it through something like Lovable or Replit.” Everyone is speculating that they’re doing what Microsoft always used to do: hampering the people that aren’t playing by their rules in very subtle ways. There’s no way to prove it, but it’s certainly all over the internet.

Yeah. I mean, I think if you look at this, OpenAI and Anthropic both have $3 billion in API revenue. $1.4 billion of Anthropic’s API revenue is from Cursor and Microsoft Copilot, so about half. They price GPT-5 about 40% lower than Sonnet, so they’re coming after Anthropic, basically. They will undercut them on price, and now the performance is roughly equivalent. They’re just basically trying to kill Anthropic’s revenue.

All right, the AI wars continue. Here’s another important part of the story from the GPT-5 announcement. We’re going to hear Sam Altman speaking about AI saving lives.

One of the top use cases of ChatGPT is health. People use it a lot. You’ve all seen examples of people getting day-to-day care advice or sometimes even a life-saving diagnosis. GPT-5 is the best model ever for health, and it empowers you to be more in control of your healthcare journey. We really prioritized improving this for GPT-5, and it scores higher than any previous model on HealthBench, an evaluation that we created with 250 physicians on real-world tasks.

I think a lot about this. The AI models are, at this point, better than most physicians, but they’re only as good as the data you feed them. That’s the biggest challenge: Can you get access to the data that truly tells your story?

Sam Altman did a very brief introduction to kick off the event yesterday, and then he did a much longer segment with a woman who was a cancer survivor. She had really done her own self-diagnosis and completely changed the course of her treatment by talking to ChatGPT and getting very good advice from it.

I think Sam chose to do that segment himself largely because, first, it’s a very emotional human segment, and I thought it was pretty well done. But also because it’s going to prevent regulators from ever saying, “Slow down” or “Stop.”

If you’re going to save lives that imminently would have been ended, you cannot slow down. You have to keep moving. I think that’s very important as a mission for OpenAI, to keep the throttle going. There are 2 drivers to keep the throttle going: the incredible healthcare benefits and the threat from China. Both of those are right front and center.

I think this felt more like PR to me than anything else, because I think you could do this with many of the models rather than this being incrementally better than the others. I think integrated broadly into somebody’s healthcare regimen is where we’ll see the real value of something like this, rather than this immediate thing. I do take Dave’s point. I think that’s exactly right. They’re pushing hard to show they’re trying to add a lot of value.

Let me throw in something on the personal front here. One thing that I do, and I’ve talked about it openly on this, is that I’m chairman of an organization called Fountain Life. When folks come in for what we call an upload, we fully digitize them. We get 200 gigabytes of data about you, including a full-body MRI.

Didn’t you just do yours?

I did. I did it a few weeks ago, and I just got my results back. I reduced my non-calcified soft plaque, which is the dangerous plaque that can give you a heart attack in the middle of the night, by 20%—the lowest it’s ever been. I got my liver fat from 6% down to 1%, which is fantastic.

What happens is that, in my Fountain Life app, we’re running this on Anthropic right now, but maybe we’ll go to GPT-5. We’re running much of the other programming on Gemini. Here’s the point: I can query all my data. The Fountain Life system pulls in all my wearables—Apple, Oura, and my glucose monitor—and I can ask a question.

I asked a question the other day. I said, “Listen, there’s a point at which my deep sleep increased significantly. What was I taking? What was the supplement or medicine that increased it?” Being able to explore things like that is amazing.

The best AI healthcare models in the world are great, but they’re directly a function of whether you have enough deep data about your physiology over time to understand what’s going on. Ultimately, that’s critical.

I think my liver fat went from 1% to 6% last week.

You're heading towards frailty.

Well, I mean, have you come through Fountain Life yet?

I haven't. I need to find the time to do it.

You're the godfather to my kids. You got to come.

I have done the heart test where they check if you have soft plaque in your arteries, and Lily was like, “With your diet, you must be on the verge of a thing.” We got it done and they said you're whistle-clean. We got nothing.

That's great. You know, but the challenge is—so then you had a steak and—you're just one—

I went to town.

One quick point: Your body is incredibly good at hiding disease, and you don’t feel a cancer until stage 3 or stage 4. Seventy percent of heart attacks have no significant precedent. You have to look. You need to get the data.

I have a schedule. First, I want to get this shoulder sorted out. Now that’s done, I can go and do other stuff.

All right. Well, Fountain Life for sure.

And Emad, you think you came through, didn’t you?

I haven’t been through yet. No, I nearly got there.

I will. I will. We will have to get healthy, and we all need the data. I think this will be really interesting, though, because the models themselves are getting good and, again, better than any doctor. They mentioned the HealthBench benchmark that they have. Doctors scored 20%, and the latest models score 60% to 70% on that, so they’re better than any doctor.

But the really exciting thing is that we built a healthcare model called AI-Medical, which we released open source, and we’ve got a much better version coming that outperformed every single model except GPT-5 and o3.

It works on a Raspberry Pi. It works on anything.

By next year, I think we’ll be at the point where the key thing is that you get the right data, especially with how much Fountain Life has. Then you just have AI running constantly, because what you want is for it to figure things out proactively as you feed it the data.

Yeah. Now we’re seeing these models being able to detect breast cancer 5 years in advance and other things like that. Wouldn’t it be nice if that happened? Now you have the capability of doing that, which I think will save so many lives.

Before the AI makes a diagnosis.

I love having a very deep bench of data for me over the course of 8 years. Right now, I go from an annual upload to quarterly updates. Ultimately, it is all about the data.

I think, just to say one quick thing, everyone on this should be trying to get as much data as possible, because the models are coming. The more data you give these models about yourself, the longer you will live and the better you will live. Before now, we didn’t have the right models. Now we have the right models, and they’ll be available via OpenAI and also open source.

Yeah. I mean, for all the folks building stuff around this, here’s my desired end state. I want to get it to a point where you’re about to drink a coffee and it says, “Hold on. Wait 10 minutes. I’m still metabolizing the donut. Give it time so I can optimize your digestion.” I think that’s when things get really fun.

Well, I want the AI to say, “Warning: Pull up, and don’t eat the donut.” Separate problem.

All right, let’s continue on here. You see another demo that came out of the GPT-5 announcement: an executive assistant for all of us. I use Outlook right now from Microsoft, and this got me thinking about moving to Google Calendar. Let’s play the demo.

We’re giving ChatGPT access to Gmail and Google Calendar. Let me show you how I’ve been using it. I’ve already given ChatGPT access to my Gmail and Google Calendar, so it just works, and it’s easy here. If you hadn’t, ChatGPT would be asking you to connect right now. Let’s see what ChatGPT is doing. Okay, that was pretty quick.

Okay, so ChatGPT has pulled in my schedule tomorrow and, oh, without even asking, ChatGPT found time for my run.

I don't think I was invited to the launch celebration.

We'll get you on there.

ChatGPT has found an email that I didn't respond to 2 days ago. I will get on that right after this. It even pulled together a packing list for my red-eye tomorrow night based on what it knows I like to have with me. It's been amazing to see that, as GPT-5 is getting more capable, ChatGPT is getting more useful and more personal.

Right. I found that impressive. I have an amazing chief of staff, Esther, that many of you know. She's incredible, but I think she could use this, and I could use this. Thoughts?

I'm really coming around to Emad's theory that they deliberately undersold it because this is coolish. Finding an email that you didn't open 2 days ago—you don't need AI for that. But we are using this stuff for business planning inside. I'm the chairman of a couple of companies that have hundreds of employees, and knowing what everybody's doing and why they're doing it is immensely challenging. We're having a field day with this in very high-level strategic planning, understanding performance, and understanding everything going on. It's an incredible unlock at the executive-management level.

Sure.

And again, for me as a watcher, it's kind of frustrating to see it planning out her run when I know it can actually plan entire business units. Still, the point is that it's very, very capable. I was frustrated, but I get it.

You know, the opportunity to now enable what Erik Brynjolfsson calls “white-collar drudgery”—there's a lot of cruft that we do just to get through. I think this solves a lot of that, and I think this will amplify the capability of a lot of people. I think chiefs of staff rise up a whole level because you could use this effectively and do a lot more.

Oh my God. The standard behavior in corporate environments is that individual people desperately want to help move the company forward. They want to contribute, they want to have maximum impact, and they want to know that the executive team knows they're doing that. It goes horribly wrong when either they don't know exactly what they should be doing, or they do something amazing and nobody notices. This completely unlocks and solves those problems.

Love that, Dave.

So when Donna and Nick ping me and go, “Here are the dates you're available for the next WTF episode, and here are 14 that intersect. Figure it out with your calendar,” I'll be able to get help with that.

You will. In fact, it'll get scheduled without your permission.

Well, we're dancing monkeys anyway, right? You're just being told, “Okay, be here at this time.”

Interesting, right? I do what's on my calendar. It's very funny: when I'd gotten to know Larry Page and Sergey Brin very well—Larry was on my board at XPRIZE in the early days—there was a point at which they said, “By the way, we fired our executive assistant.” I said, “What do you mean you fired your EA?” They said, “Well, we learned that if we don't have an EA, no one can put anything on our calendar without our permission.” Then, like a decade later, I was scheduling a podcast with Elon, and I said, “Elon, who should I schedule with?” He goes, “Me?” I said, “Don't you have an EA?” He goes, “Nope.” So maybe that's the mistake we're making.

All right, let's go on to the next topic here. This next slide reads, “AI revenue models.” GPT-5 is available now for free, including its most advanced models. At the same time, Gemini has its advanced models at $249 a month. Grok Heavy is at $300 a month. How do you think about the pricing situation here? ChatGPT has 700 million weekly active users, on its way to 1 billion probably within the next 6 months. Dave, thoughts?

Yeah, no, they really slashed the price. It shows up for the user, but also in the APIs, which I think we have on the next slide.

We do. Let me go ahead to that slide here. Yeah, here you go.

Yeah, this is what Emad was talking about earlier. They just absolutely slashed the cost per intelligence way, way down. I think you said 40%, and I had it at about half of where we were a week ago. That's a big, big deal, and it's more than you would expect on the curve. Again, they didn't really sell it yesterday in any big way, but it is a big step on that Pareto frontier.

GPT-4.5 was $75 input and $150 output.

Wow.

Yeah, as compared to $1.25 on input and $10 on output.

I would say per million tokens.

I would say GPT-4.5 was never quite on the cost frontier anyway. What I see in this, with this almost order-of-magnitude reduction in the cost frontier, is the unlocking of new use cases, and I would expect those to be qualitatively different. For example, if tokens for LLMs are suddenly an order of magnitude cheaper, that means that for scientific discoveries or mathematical discoveries that require searching lots of possible completions of sentences, theorems, et cetera, you can do 10 times more searching, and that makes a qualitative difference.

You can brute-force it in that sense.

Yes, exactly.

Exactly. I've got to give a shout-out to the thousands and thousands of engineers out there who listen to this podcast. If you've tried writing code through any of these really great models—either Anthropic's, the new GPT-5, or Gemini 2.5 Pro—and you tried a month ago, you have to try again today. It's just night-and-day different in terms of being able to build something without even looking at the code, in terms of getting exactly what you asked for. I'm using mostly Gemini 2.5 Pro Deep Think to do the planning, but then I'm putting it into either GPT-5 or Claude 4 Sonnet Max to do the coding. It's working like you would not believe, and it's night-and-day better than just a month ago. How many of the frontier models do you have open at a time, and are you trying the same thing on each of them, Dave?

Yeah, I keep them all open, actually. I've got—but, look, it's $250 a month. It's not going to kill you, and you can turn it off anytime. I keep them all open, and I don't usually try Grok for code. I do everything else. I'm not sure why. Maybe I should.

Alex or Emad, how are they getting these cost reductions?

I think a lot of it—this is based on public information shared by the frontier labs—comes from optimizing the inference stack. Moving to faster Blackwell GPUs, I think, is one factor. There are low-level optimizations in the tech stack at inference time, distillation of smaller models with fewer parameters based on higher-quality data, algorithmic innovations, and architectural innovations. These all compound. Some of them are 50% improvements, and some of them are 2- to 3-fold improvements, but collectively, as is now the norm in the industry, we're seeing order-of-magnitude-per-year cost reductions.

But how much money are they losing on this per transaction?

It's difficult to know from the outside, but I would also say that the matter is somewhat confounded by the enormous capital expenditures going into this space. It's not necessarily even a reasonable question to ask how much is being lost. You have to factor out the capital expenditures, as we've discussed previously. We're in the process of tiling the Earth's surface with data centers. This is an enormous capital expenditure, so it's a little bit difficult to separate out the amortization of CapEx from the OpEx of just day-to-day inference and electricity.

I can definitely tell you they're definitely not incinerating money. There's a lot of FUD on the internet about them: “Oh, they're incinerating money. They're losing huge amounts.” They're not. They're operating at about break-even or better. In the context of what Alex just said, the order-of-magnitude improvement in cost per compute that just came from the GB200s from NVIDIA would put this way over the top.

In fact, I was talking to Gemini earlier today about what it thought they spent training GPT-5, and it came back with this insane number: $1 billion on H100s. I said, “Well, I don't think they used H100s.” It said, “Oh, okay. Well, if they used GB200s, it would be more like $60 million.” Wow. But, yeah, it's about a factor-of-10 reduction in the cost of compute, and they're passing some of that through. The GB200s are just coming off the line and starting to get into production, so it'll be a little while.

When I saw the pricing, I thought they were doing this for competitive advantage and taking a huge loss. What I'm hearing you guys say is that's not the case. They're really running maybe about break-even, but passing on massive savings to the consumer.

Yeah, see, this is what hyperdeflation looks like. It's an interesting thought experiment to ask: assuming this is sustainable—and I have no reason to think that it isn't sustainable—what does hyperdeflation right now at inference time for frontier models look like once it starts to spread to the rest of the economy? This leads to Peter's abundance state, I think.

I just want to say something for those listening: I feel smarter during these episodes, getting a chance to speak with Dave, Salim, Alex, and Emad.

And I hope you do, too. That's the reason I do this. We put about 20 hours of deep research in every week, trying to find the most relevant content to share with you. Then we try to make it understandable, connect the dots, and deliver a distilled CliffsNotes to help you stay ahead. Selfishly, I do this because it's a blast.

I think the curation that goes into this, where we're looking across the spectrum and then picking out the most relevant things, is important. Dave talks about the actionability of it, but I think the fact that we can curate the very important bits for our viewers is the most important part and the most fun. We get to see that first.

Yeah. I always have my kids in the back of my mind when we're doing these podcasts because they're going to live their entire lives in the post-AGI world. One of my kids was talking to one of the guys here in the office and said, "All your dad ever talks about is AI." I said, "Yeah, but the whole time you were growing up, did I ever talk about AI once?"

I never mentioned it until suddenly it's going to change your life. You must get on top of this right now. You must have a plan. It's for their own good. So I'm always thinking about that in the back of my mind. How many listeners out there need this information in order to remap what they're doing?

And to be inspired, right? Our goal here is to inspire everyone to be in the thick of this, to find your own moonshots, to understand.

So, if we just connect the dots on one thing, the fact that GPT-5 is now free and has built into it the best doctor in the world that can diagnose anything on a much better basis, instantly, for you is a profound uplift. I think this is the point you were making earlier, Alex.

Exactly. When we talk about abundance in all of its many facets, taking 700 million people and suddenly giving them access to state-of-the-art AI becomes transformative.

It'd be interesting. I'm looking forward to this. Here's the thing I want to watch: How does OpenAI's user growth go from here, given that they've made it free?

I thought you were going to go in a different direction. I agree that that's interesting, but another is, in some sense, this is the greatest A/B experiment that macroeconomists should be all over. Prior to yesterday, most of the world didn't have access to frontier AI. Starting yesterday, a fraction of the world does—call it a tenth of the world. What does the before-and-after look like? Do we see dramatically different outcomes in different dimensions?

Sam didn't have a huge part in the event yesterday, but he did a lot of postgame interviews, which I watched. In one of them, one of the interviewers said, "Imagine college and education for me in, say, 2035," and he said, "2035? Is there college in 2035?" He said, "If it exists, I mean..."

We need to coin a term for it. Maybe this is an intelligence shock that's hitting the world.

Oh, I hope so. Intelligence inversion.

Just intelligence.

My son is 13. I'm hoping the university system implodes in the next 5 years, before he—

By the way, I was talking to my son. I said, "I'm going to go do WTF with my Moonshot mates," and he goes, "Have you reached a million subscribers yet?" I asked, "Why?"

It's not about view count. I think it's more about quality. If a smaller set of people gets much more value out of it, I think that's better.

All right, Emad, a huge fraction of the people I bump into have actually watched the pod. So, we've got a quality audience for sure.

Yeah. I think the closing thought is that there's a cap on human intelligence, but there isn't one on artificial intelligence. So everyone will have abundance, and you can expect that next year a 0 drops off here, and then the year after another 0 drops off.

And we're seeing that.

Insane. It'd be crazy.

That's insane. So I want to hit a couple of things. OpenAI is eyeing a $500 billion valuation, which is pretty extraordinary. It's one of the highest-valued private companies, along with ByteDance, SpaceX, and Ant Group. I wouldn't say much more here other than: How will they go public? When will they go public? And will this be the largest IPO ever?

We've seen OpenAI's GPT-5, but they also unveiled their open-weight models. I don't want to go into this in too much detail, but, Alex, do you want to lead us on this one? Actually, Emad, you're the open-model champion around the world.

Wait, wait, wait, wait. Hold on one second. Can we just go back to the previous slide for a second?

Okay.

OpenAI made $10 billion and is making about $10 billion a year. Microsoft is making about $300 billion a year in revenues. And so OpenAI is valued at half of Microsoft. I just want everybody to see that ratio: $3 trillion.

No, no, I mean revenues. It's $10 billion versus $300 billion in revenues.

Okay. Okay, so there's a very big difference. It's very lofty, but that feels overpriced to me. Anyway—

Sam's projection is $100 billion to $150 billion in revenue in—what is it?—2 years from today.

Which I don't doubt is entirely possible. There's only 2 versions of the world, actually. There's a version of the world where OpenAI easily hits that target, and there's a version of the world where Google destroys them and wipes them off the face of the earth. Those are the 2—

Possible outcomes.

I mean, talk about capitalism at its finest, right?

Look, SpaceX is $13 billion in revenue.

And what's its valuation like? Almost $1 trillion or something like that? Half a trillion?

$210 billion right there.

Oh, it's right there. Okay, okay. But when they own Mars, it'll go up a little bit. So, next slide. There was an interesting note that Elon pushed out on X: When is OpenAI going to buy Microsoft? Fascinating. All right, continue with the open models here.

I think it's pretty significant. A lot of people worried about Chinese open models going everywhere, and OpenAI has released a really solid model. It's a bit weird, it has to be said, but the main thing is this model costs $4 million to train, and it's better than any model that we had this time last year.

That's extraordinary.

Next year, it will cost $400,000 to train a model.

A model.

How did you know it was $4 million to train? Did they release that? They said 2 million H100 hours, and the 20-billion-parameter model that runs on your laptop was 10 times cheaper. It was 2 million H100 hours at $2 an hour.

And that's from scratch, or was there distillation from a big model?

No, it's from scratch. It's 80 trillion tokens—80 trillion words.

So, Dave, to your point, I think the footnote there is: Where do those tokens come from? I think it's reasonable to assume, in the style of, say, Microsoft's Phi models, that these are tokens generated through some synthetic process from a much larger, much more expensive, in terms of fixed costs, model.

In which case, whether you call the total pre-training cost just the marginal cost for training on the back of a much larger parent model or teacher model, I think that's the key distinction. We should do a whole podcast on just that topic because—

In the broader sense, AI that helps create the next AI is an incredible force multiplier for humanity. It's a good example because when you just distill the training data and create some synthetic data using the prior model, you knock 90% to 99% off the cost of creating the next iteration. It's crazy economics, how it feeds back, like no technology previously—other than maybe robots building robots someday. There's nothing that feeds back like that.

We need a new term that supersedes Moore's law here, because the speed of this is extraordinary. We're witnessing the evolution of something that I think we're going to look back on—

I can tell Alex is about to say something brilliant. I know that.

Look, let me point out a couple of things here. One, we do have this already. It's called education. Distillation is what humans use to take the years and years that researchers and teachers spend accumulating and then convey this in a concise lesson to a student.

So we as humans do distillation as well. It's very efficient, very economical. So it's perhaps not that surprising to see distillation give us radical economic efficiencies in these open-weight models. That's the first point.

The second point, just to go back, Peter, to your earlier comment: Having these supply-chain-safe, if you want to call them that, open-weight models is transformative for so many applications that are highly regulated and very sensitive to supply-chain risks—in finance, in healthcare, and in government.

Now we have American-trained models that can be embedded in all sorts of mission-critical, internet-disconnected systems, and that is going to be transformative.

Insane.

Yeah. I think just one final thing on this: This model only has 5 billion active parameters, and so it runs faster than you can read, even on a MacBook.

I think the big thing is that everyone's talking about billion-dollar training runs. I actually don't think that's true at all. I think you will have a GPT-5-level model in 2 years, max, that will cost under $1 million to train end to end.

And nobody's got that in their numbers. All right. Anything?

The expensive part was the journey to get there. I completely agree that at some point we're going to discover—I made this point previously—the perfect architecture, the perfect sort of microkernel version of a foundation model that's relatively small in parameter count and fully multimodal. If we knew what that were today, we could radically collapse training costs.

I think what this actually shows is that we don't even need that. We need to have 1 trillion good tokens. If we've got 1 trillion good tokens, then you can train a frontier model for less than $1 million next year.

And so, I think, do you then embed that into all sorts of devices and humanoid robots and moving cars and—

Anything? Yeah, everything, everywhere.

You embed everything. Everything becomes built in.

I think that's where this goes.

Yeah. Your question exactly defines the future entrepreneur: “What am I going to do with all that?” If, for $1 million—which is seed money—I can build a GPT-5-level model, what else can I build? This is going to be the age of abundance, where it's limited by people's imagination. If you can imagine something genuinely useful that people want, the cost of creating it is near zero.

Well, you don't have to create the model. One entity needs to create that model open source once, and the economies of scope mean it can be used anywhere.

Maybe I can stop arranging the room for the damn Roomba.

There you go.

That would be a great starting point.

Let's not go there.

All right. We have the back end of this WTF episode, which is to look at all the other companies in the AI wars. They include Grok, Gemini, Meta, NVIDIA, and Apple. I'm going to try and move us through this. There's some important data we need to share with everybody. This is what we're watching and what we're keeping in tune with. Hopefully, you are too. Let's jump in.

The first is, again, a quick look at the Humanity's Last Exam benchmarks. Alex—

Yeah, I think what we're really seeing here is that we see 2 models—Grok, with extensions and derivatives of various sorts, and GPT-5 and its derivatives—leading the pack. I think if you pull back that headline, what you're actually seeing here is the power of tool use and the power of parallelism, with GPT-5 leaning heavily on search and other tools, and Grok leaning heavily on the power of having multiple parallel agents collaborating and zooming out to a 10,000-meter perspective.

I think what this points to is a world in which it's not just the core foundation model, but arrangements—not even necessarily scaffolding—the ability to integrate these microkernel-type foundation models with each other in teams of agents, and the ability to integrate them with powerful tools in their environment. That's going to turn out to be one of the next big shocks in terms of how we're able to challenge the frontier for HLE and other hard benchmarks.

By the way, people listening—our subscribers listening—if you get a second and want to do something fun, just get onto ChatGPT-5 or Grok or wherever and ask it to give you 10 example questions from Humanity's Last Exam.

Yes. I'm going to just share a couple of them here that I asked for. Here's one in the classics category. Here's a representation of a Roman inscription originally found on a tombstone. Provide a translation for the Palmyrene script. A transliteration of the text is the following, and then you have to transcribe that.

Here's another one: What is the rarest noble gas on Earth as a percentage of all terrestrial matter in 2002?

All right, here's one I'm going to ask our geniuses here. In physics, a point mass is attached to a spring. The spring constant is K, and it oscillates on a frictionless surface. If its amplitude of motion is doubled, what happens to its total mechanical energy? A, it doubles; B, it quadruples; C, it triples; or D, it remains the same.

I'm not going to ask you to answer that.

It should quadruple. I would expect it to quadruple. Yes, correct.

All right, there we go. At least flashing back to my physics courses. Please, for God's sake, let's not do that.

All right, last one. Consider a balanced binary search tree, like a red-black tree, with N nodes. What is the worst-case time complexity for searching a given key?

O(log N).

There you go.

I'd like to point out that the open-weight models that OpenAI just released scored 19% and 17%. The 17% is the 20-billion-parameter model that will run on anyone's laptop. Crazy.

All right, we had Elon pipe up. He said, “Great work.” So, here was the tweet he's referring to: “Very proud of us at xAI after seeing the GPT-5 release with a much smaller team. We are ahead in many benchmarks, with Grok 4, the world's first unified model, crushing GPT-5 in benchmarks like ARC-AGI.”

So, we're going to have this continuous—I don't know if it's an ego battle, a financial battle, whatever it might be—where everybody's just trying to one-up each other. And, of course, his next tweet was, “Grok 5 will be out before the end of the year, and it will be crushingly good.” So, comments on Grok?

He just tweeted saying, “Grok 4.2 before the end of this month.” Number 1—

4.2.

I think, Peter, one of the takeaways here is that whoever is defining the benchmarks wins. It's like, you know, you create the evals and humanity wins. It's amazing how starved the research community is for compelling new evals, as discussed previously. To the extent that we can create more evals that—as I think your community has also chimed in historically—with some wonderful ideas for abundance-oriented benchmarks or evals, the frontier labs will, I think, race to achieve them.

All right. Most of the Polymarket predictions have Google winning by the end of the year.

And for good reason. What we've seen is extraordinary, and here's the title of the slide: Demis, in a word, relentless. In only 2 weeks, they've shipped or achieved—and I'll read the list here—Gemini 3. We'll see an example of that. Gemini 2.5 Pro Deep Think. Gemini Pro free for university students. AlphaEarth, amazing—we'll see a demo of that. Aeneas, deciphering ancient text. Gemini won the gold medal in the International Math Olympiad. Storybook, Kaggle Game Arena, Jules, NotebookLM video overviews, and Gemma, which passed 200 million downloads. This is Google's lightweight open-source, open-weight model. Really impressive work.

Yeah, well, Demis has 6,000 people in AI R&D. OpenAI is up to a little under 2,000 now, but these guys at Google have been working on it for years.

Mhm.

So they've got about a factor of 10 more person-hours put into it so far, and they're all operating on things in parallel. So, they're now unleashing it all. It was all just kind of sitting there in the lab until OpenAI put the competitive pressure on them.

Yeah. Now something has shifted in a big way at Google, and in a couple of fronts. One, they're unleashing all the things they've been working on. The other is that they proactively reached out to a bunch of our companies, including Blitzy. Blitzy is a particularly hot company, but I don't know how they found it—probably through all their big data. The Gemini people came over to our office proactively and said, “We need to meet with you.”

So they're really reaching out, trying to get the businesses to move over to using Gemini. That was also really evident in the GPT-5 rollout yesterday: the call to companies saying, “We're here, we're open, we want to partner with you,” and we're cutting the price point to make it easier to do. We're open for business. I think that's a new thing. I hadn't seen anyone proactively reach out to our companies until this week.

Amazing.

All right, let's take a look at a few of these examples coming out of Google. This is Google's Genie 3. It's world models for gaming. Let's play the video. I was blown away by this. I found this probably one of the most impressive things I've seen in the last week.

What you're seeing are not games or videos; they're worlds. Each one of these is an interactive environment generated by Genie 3, a new frontier for world models. With Genie 3, you can use natural language to generate a variety of worlds and explore them interactively, all with a single text prompt.

Let’s see what it’s like to spend some time in a world. Genie 3 has real-time interactivity, meaning that the environment reacts to your movements and actions. You’re not walking through a pre-built simulation. Everything you see here is being generated live as you explore it.

Genie 3 has world memory. That’s why environments like this one stay consistent. World memory even carries over into your actions. For example, when I’m painting on this wall, my actions persist. I can look away and generate other parts of the world.

But when I look back, the actions I took are still there. Genie 3 enables promptable events, so you can add new events into your world on the fly—something like another person or transportation, or even something totally unexpected. You can use Genie to explore real-world physics and movement and all kinds of unique environments.

You can generate worlds with distinct geographies, historical settings, fictional environments, and even other characters. We’re excited to see how Genie 3 can be used for next-generation gaming and entertainment. And that’s just the beginning.

World models could help with embodied AI research, training robotic agents before working in the real world, or simulating dangerous scenarios for disaster preparedness and emergency training.

All right, I’m going to pause there, but holy cow. I mean, first of all, the simulation theory just took a huge jump forward.

Boom.

This blew my mind. I actually showed this to a friend who spent the last 2 or 3 years building metaverses.

He literally had his jaw drop, and he said, “I don’t even know where to start.” The fact that you can have a responsive environment that tailors itself depending on where you look, and that all of this is generated on the fly in real time—he couldn’t cope. I’ve just never seen his mind broken like that.

Yeah, it’s a masked diffusion transformer, similar to a lot of the video models like Veo and others. Again, we’re seeing the breakthroughs coming in this, especially because Google has such an amazing data set. I think that you’ll see a video model like this from xAI as well. This is what Elon is going to be putting those 10,000 Blackwells toward with his video model.

But the fact that it’s real time now gives you a real idea about that. Similarly, we’ve seen real-time video generation from Wan and others now. Every pixel will be generated in a few years, which is going to be cool.

And what if Meta—you know, Zuck has wanted the metaverse forever—and of course, this is delivering the metaverse?

On the one hand, this is billions of dollars of capital expenditure that’s been allocated to video gaming, or to metaverse software, that suddenly is in danger of having been rendered irrelevant. On the other hand, if this can all just be the output of a single model, a thousand voices in the video gaming industry just cried out in anguish if this is all just a prompt away.

That’s the response I got.

Yeah. I mean, with Veo 3 potentially crushing Hollywood and this potentially crushing the video game industry—or reinventing it, accelerating it, making it possible for anybody to create magically compelling video games—

This is the Star Trek holodeck. This is the Matrix. This is the key node, potentially, in the tech tree of our civilization that unlocks general-purpose robotics and general-purpose autonomous vehicles.

Yeah, because they can train inside that.

That’s right. Extraordinary. Absolutely extraordinary. All right, here’s another extraordinary gift from Google. This is Google’s AlphaEarth Foundations, mapping the planet in real time. It turns massive satellite data into unified global maps, with 10-by-10-meter precision, tracking deforestation, crop health, water use, and urban growth. Take a quick look at this video.

This is how our new AI model, AlphaEarth Foundations, interprets the planet. Different colors in this map show how different parts of the world are similar in their surface conditions. So, similar colors mean similar things, like 2 deserts or 2 forests.

The model understands the unique patterns that distinguish any ecosystem, so it’s able to use those learned patterns and quickly find matching patterns in other places in the world. This allows it to tell the difference between, say, a sandy dune on a beach and the deserts of the Sahara.

It used to take months to years for scientists to accurately map the world. But with our data set, they can do it in minutes. Much like Google Search has indexed the web, with AlphaEarth Foundations, we’ve indexed the surface of the planet.

We’re making this available through Google Earth Engine for the years 2017 to 2024.

All I can say is, just in time. Thoughts?

What’s interesting here, I think, is that this is what’s called an encoder-only model. It takes 10-meter-by-10-meter patches of Earth’s surface and converts them to high-dimensional vector representations. Encoder-only models were very popular in natural-language processing prior to the advent of so-called decoder-only models like the GPT series.

I think the elephant in the room here is that once we have encoder-only models that cover the Earth’s surface, we’re about to get decoder-only models. What that’ll enable in practice is that, right now, with these encoder models, you can convert arbitrary land masses or ocean masses to vectors and do a bit of regression on them and maybe a bit of lightweight prediction.

With decoder-only models, you’ll be able to take a few square kilometers of land and extrapolate visually: What does the future of this land look like? You’ll be able to do searches of interventions. If I put a parking lot here, or I put a hospital here, what’s going to happen, in all likelihood, to development in the area?

You’ll be able to do urban planning as a matter of a tree search, in the same way in which AlphaGo, AlphaZero, or MuZero are able to play chess. That’s, I think, going to be the real amazing unlock.

Amazing.

See, this is the application usage where I think all these things start to really shine, where you can take all that capability and apply it to something like this. It’ll completely transform how we look at the world. My mind is kind of blown with this one.

Yeah. I’m really glad you said that, Alex, because I really did not get the implications of this until you explained it just now. I do appreciate that.

I had a meeting earlier today with Satya Nadella, the CEO of Microsoft. All these companies are thinking, “What’s my moat? What’s my moat? What’s defensible? What’s going to give me recurring revenue for the next 20 years?” I’m like, it’s just not a way to think anymore.

If you look at the rate of change, it’s all about small, nimble teams and great team dynamics. Overall, there’ll be far, far more company success than ever before, but you can’t expect to sit still. You have to reinvent yourself all the time.

Amen.

Amen. I think agility and passion-driven building, and understanding the root-cause problems that you want to go solve, those are the fuel for the future. It’s not setting up regulatory blockage.

All right. Next up is a video of Zuck on Meta Superintelligence Labs: “Superintelligence for everyone.” Let’s take a look.

I want to talk about our new effort, Meta Superintelligence Labs, and our vision to build personal superintelligence for everyone. I think an even more meaningful impact in our lives is going to come from everyone having a personal superintelligence that helps you achieve your goals, create what you want to see in the world, be a better friend, and grow to become the person that you aspire to be.

This vision is different from others in the industry who want to direct AI at automating all of the valuable work. This is going to be a new era in some ways, but in others, it’s just a continuation of historical trends.

About 200 years ago, 90% of people were farmers growing food to survive. Today, fewer than 2% grow all of our food. Advances in technology have freed much of humanity to focus less on subsistence and more on the pursuits that we choose. At each step along the way, most people have decided to use their newfound productivity to spend more time on—

All right, so he’s out pitching hard. He wants to get to superintelligence first. What could possibly go wrong? The poaching continues, and I love this. Zuck contacted over 100 OpenAI employees. 90% of them turned him down. Why? Because they think OpenAI is closer to AGI than Meta. That’s got to sting. Emad, what do you think about that?

I think he has a very different definition than Sam Altman does. I think one of the reports was that he was talking about how AI could make Reels a better product. So, I think it’s a very different view of the type of ASI that we’re talking about.

They should just call it Meta Intelligence. But I think it shows—you see there—million-dollar offers, and people still don’t move. I think everyone feels that we’re getting close to that AGI point, and you want to be where it’s going to happen, because what even is money after that? We’re going to find out soon.

That’s a very important point. In this post-abundance world, we’re living in a post-scarcity world as well. Money has very little meaning. Emad, you and I, and Alex, you and I, have spoken about that at length, right?

And I think, Peter, as the cost of talent is increasing—and it would appear that it certainly is—that’s going to force frontier labs to start competing based on algorithmic insights and ideas. I think that’s a net positive for the economy and the world.

Amazing. All right, I love this next one: the Zuck poaching effect. Sam just announced $1.5 million bonuses for every employee over 2 years. He's now officially made every employee at OpenAI a millionaire by giving them over $1 million. That compares with 78% of NVIDIA employees who are also millionaires.

Dave asked if that included the baristas. Do we answer that?

I don't know, but we'll be at OpenAI in a couple of weeks, and we'll ask.

Okay. It'll affect your tipping at the coffee counter, I guess.

Oh my God. Yeah, Peter, I would expect this to create a bloom of seed funding for startups in the next year or 2. It's just going to be absolutely enormous. I'm already starting with some of the startups I advise to see the beginnings of it.

That is such an important point, right? This is something that America does so well. We create these decabillion- and centibillion-dollar companies and trillion-dollar companies, and because of stock options and stock distribution, we make all the employees super wealthy. They turn around and invest in other individuals, and that doesn't exist in a lot of countries. Dave, you've spoken about this.

Well, there are 2 things that are different this time, but you're right: that is the engine of America, and it works really, really well. This time around, it's so fast, and the teams are so young. That's unprecedented, and I could see some things going wrong with that. But it's a field day right now, so we might as well savor it.

Also, it's much clearer now how you're supposed to work with either OpenAI, Anthropic, or Google.

How is that?

It's not clear—I mean, they've made it very, very clear that they want partners in all these categories, especially complicated, regulated categories or categories that have proprietary data. They've said, “Here's the API. Here's how we want to work with you. The pricing is going to be super low. We want you.”

For those 3 companies, it's really clear. It's not as clear with Grok yet, and I don't think anyone knows how to work with Meta, if there is any way. But for the 3 other big guys, it's just a field day: “Here's how we want to partner, and please just bring in the revenue, change the world, and we're all happy.”

I really am cheering for Sam in this battle, too. Mark Andreessen built Netscape, the coolest company ever, and got absolutely obliterated when Microsoft woke up. They just annihilated him and changed the course of his life. He did well in the end anyway, but it was a complete life change. So now Sam is that guy. He woke up Google, and he's got a little shaky relationship with Microsoft.

I think he pushed Google over the edge. I think they were awake already in that regard.

Yeah. Well, now he's got them all coming after him concurrently, and he's got to outrun them. It would be a great American success story if he can stay ahead of that and survive.

Can't wait for the Hollywood movies that are coming out on all these subjects.

Yeah. I think one of the really interesting things is that crypto has basically been legalized in America almost fully in the last week. I think next year, crypto × AI is going to be the most ridiculous thing you've ever seen, because these startups will start with a few smart people, get massive traction by leveraging these models, and then anyone will be able to buy them pretty much instantly. We're just at the start of the bubble, I think, versus what we're going to see. It's going to be the biggest bubble of all time.

Well, “bubble” has a negative connotation to it. Emad—

Of course, but we're just at the start now. This is the final hurrah of the current financial system—

—or societal system as well.

I really think, though, if you take a step back and try to visualize Sam's life for real, the biggest companies in the world are offering your direct reports $1 billion to walk out the door. You have to fight. At the same time, Mira Murati and Ilya Sutskever, 2 of your founders, are trying to raise $10–20 billion to compete directly with the thing they built at OpenAI.

They did raise—

They did. Does it get any harder for an entrepreneur than where Sam is right now? And he's bulletproof. He's just fighting his way through it. It's something the movie will be really cool.

I think there's so much to work on. This is a great testament to the fact that if you keep pushing product, keep launching new things, and keep innovating, you can stay ahead. Facebook showed us that, Yahoo showed us that, and Google showed us that. In their era, they just keep breaking boundaries. The only thing now is: can you break those boundaries, break the status quo, relentlessly keep doing that—

—and differentiate yourself from the competition.

Yeah, I think it's sort of an interesting economic experiment. In the past, I've compared the AI buildout that's happening in the US to 1939 and the prelude to the Manhattan Project. It's an interesting thought experiment to ask what would have happened if nuclearization and the Manhattan Project hadn't been a nationalized effort, but instead had been a private-sector effort where blue-chip companies were all competing with each other to see who could build the first atomic weapon. How much would they have been spending to poach the top scientists from each other to build that first atomic bomb, which had such strategic import for the future light cone? I think we're living, in some sense, a civilian version of that thought experiment.

Mhm.

Amazing.

The really interesting thing is that it's not hard to know how to build the models if you know how. The know-how is really, really rare, and that's why they're willing to splash these billions on top of that. It'll be interesting to see what they come up with now as these things get commoditized.

All right. On the OpenAI train, NVIDIA and OpenAI announced their first European data center in Norway. This is a $2 billion OpenAI data center with 100,000 NVIDIA GB300 superchips. It will host 230 megawatts of capacity, expandable to 520 megawatts—half a gigawatt—powered 100% by renewable energy from Norway. Let's take a quick look at this video.

The launch of Stargate Norway marks a new chapter for AI infrastructure in Europe. We're entering a new industrial era. Just as electricity and the internet became foundational to modern life, AI will become essential infrastructure. Every country will build it. Every industry will depend on it.

AI is no longer hand-coded. It is trained. It is refined with massive compute. It is deployed into factories, research labs, and digital services. Stargate Norway will be powered by GB300 superchips and connected with NVLink. It is designed to scale to hundreds of thousands of GPUs and support the most advanced models in training, reasoning, and real-time inference.

All right, there you got it. Emad, analysis, please.

Yeah, I mean, I think this is part of the big sovereign AI strategy, because your comparative advantage as a country will be how many chips you've got and how much intelligence you have when most of your workers are digital. We've seen OpenAI go very aggressively on this front. In fact, this week they announced that they're going to be rolling out ChatGPT to all federal workers in the US at a cost of $1 per agency per year. I think the land grab has really begun. They couldn't have said “free.”

I would add, Peter, that there's a less obvious angle here, which is pulling back the details on the announcement. This new data center is planned to be powered with hydropower, which is intrinsically scarce. You either have access to it or you don't. It's not that easy as a nation-state to create a lot more hydropower. So there is very literally a land grab here. Stargate is planting its flag in that hydropower. To the extent that Europe has a policy of bounding power to certain energy sources, there's only a finite amount that's available to be repurposed for AI. So, a real land grab.

And we'll see geothermal energy as a land grab, and we'll see other areas. I want to move this forward here. We saw a couple of interesting announcements coming out of the White House. Apple announced a $100 billion US investment, increasing its total investment to $600 billion. I don't know—this is finally Apple coming back to the US. How much of Apple's products are manufactured overseas right now? Anybody have an idea? It's got to be an overwhelming majority, like 90% plus.

Huge.

Comments, Dave.

Well, look, most of the countries that have a Samsung, like Korea, have very tight government-industrial integration. The US has never really had that before. This is the first time. It obviously works really, really well. It got Japan on the map, then Korea on the map, and now it's gotten China on the map.

Trump is the first president to really take this to its limit. He's a business guy, so he knows how to do it. It's obviously going to work really, really well. It's not super hard to figure out; you just need to do it.

I would also maybe add that, going back to this idea of a tech tree existing for civilization, it seems clear that there's an innermost loop to the tech tree at the intersection of fabs, electricity sources, drones, and rare earths.

To the extent that it's possible to collocate as high a density as possible of talent and infrastructure for building all of these, I think that has the potential to lead to an economic explosion for the US and for the world.

Amazing. One more article coming out of the White House here on AI: Trump demands Intel CEO resignation over China ties. Trump labeled the CEO highly conflicted over $200 million-plus in past investments in Chinese tech firms and a relationship with the Defense Department. For me, this has echoes of the J. Edgar Hoover anticommunist campaigns from the FBI.

Emad, do you have any opinion on this, being a non-American?

Yeah. Well, look, just posturing, right? I think that this whole US-versus-China AI thing is completely overblown because everything gets commoditized soon anyway. Actually, to be honest, we should have had the push for open source, given that China wants to get into all our systems. Then they would have actually put proper money behind it.

I think that it's completely wrong, though, because, again, the correct view is that this is abundant and it's going to come to everyone everywhere. You can't keep a lid on it at all. How do you keep a lid on math?

I threw this into the deck to spark the conversation around this. Right now, the chips that are driving this entire AI revolution have a 2/3, or 66%, market share through TSMC. A single manufacturer is utterly insane and not sustainable. So my guess is that the White House is thinking about this and talking about Intel every day. It's not a coincidence that Trump decided to tweet about one CEO. The China thing, I don't know what he's thinking about there, but Lip-Bu Tan is a 65-year-old guy. Intel must succeed; it's just an incredible national priority.

Incredible asset, right? I mean, it defined the last 50 years.

Mhm.

Yeah.

So, anyway, the point is the White House is talking about it. We need balance in chip manufacturing desperately, and we need a lot more volume of chip manufacturing.

If you're at Intel, what you should be thinking about is: How do you leapfrog?

Their 18A, or 1.8-nanometer, technology is absolutely fine. They need to get the yields up, and then they need to build more fabs, which means federal help. I just think that if someone running that company can get friendly with the current administration, it's all unlocked and it'll explode and succeed wildly, which is what America needs. I don't know. I hope they figure out the relationship between Lip-Bu Tan and Donald Trump quickly.

Yeah.

Actually, part of this was because Intel was trying to sell its fabs to TSMC.

Yeah.

So, again, it gets complicated.

I mean, that would be devastating for the world, really. There's no way that can get through. But I get it, right? All the losses at Intel come from the fabs. They would immediately monetize a huge asset. The remaining Intel would be hugely profitable the next day. So that's the allure of that transaction. But then you have one company controlling their entire destiny. There's no way that makes sense.

All right, I want to close out with this slide. I find it telling, especially on the backside of the Intel conversation, that we're still early in terms of buildout. Here we see infrastructure capex as a percentage of US GDP. Railroads were 6% of GDP back in the 1880s. Telecom was 1% back in 2020, and today AI data centers in 2025 are at 1.2% to 2%. We're still early. Alex, we've talked about this, and, Emad, we're about to turn the planet into computronium. We're building data centers everywhere.

And maybe the solar system. We'll see.

And maybe the solar system. You have an event coming up soon. Talk to me about it.

On August 20th, we have our next monthly EXO workshop. The last one, the last 2 or 3, have sold out. People are absolutely loving them. It's $100 to come, bring your company, and we'll teach you how to build an exo. We actually have a great ad which we'll get a link to and post in here. They created an ad where an AI reads out a real review by a real person, but it's an AI reading it out. Super funny.

And for those interested, I've got some comments interested in the Abundance Summit in March. Applications are closed at this moment. They'll be reopening in September, but you can get on the wait list by going to www.abundance360.com and let us know that you're interested. We'll have all of our Moonshot mates at the Abundance Summit as well.

Let's take a quick look around the horn. Dave, what's happening for you in the next few weeks?

The biggest thing by far is that we'll be together at OpenAI in, what, 11 or 12 days? I'll be there the whole week, actually. God, there's so much going on in that building.

Yeah.

So, really looking forward to that.

Emad, it's 2 a.m. Do you know where your children are? You're a nuclear power source, buddy. Thank you for sticking with us through the hours in the UK.

It's too much fun to sleep.

It is. What's on your plate over the next month?

We have some big releases coming up. In particular, I've been looking at the economics of the AI age. It's going to be wild. I'm going to be releasing a bunch of stuff around that.

I've seen what you're going to release, and it is stunning. Dare I say, just earth-shattering.

Oh, my goodness. I think we're in a time when, although we're on an exponential curve, every point looks like the knee in the curve or the inflection point. One has to be careful of such anthropic bias. I spend most of my time advising tech startups and making sure that the benefits of AI are evenly distributed throughout the economy. Every day is an adventure and an opportunity to smooth out the singularity, as it were.

All right. Well, everybody, thank you for joining us on this episode of WTF and the GPT-5 announcement. We'll be coming back to you with an episode again next week. Please tell your friends about what we do. Our mission here is to help you understand how fast the world is going, to inspire you, to give you the motivation to create your own moonshots, and to make this understandable. And actually, what was the word you used, Alex?

Riveting. I'm at the edge of my seat. An amazing time. The most amazing time ever to be alive.

All right, to all of you, thank you for a fantastic conversation.