中国崛起、GPT-5.2、Anthropic IPO 与 AI 信任之战：与 Emad、Salim、Dave 和 AWG 对谈 | EP #214

China is accelerating its push to become independent of NVIDIA, with Cambricon planning to triple output to 500,000 accelerators in 2026.

I fully expect, as I've mentioned on the pod in the past, that we're just going to see...

Having an open-source model that you can test as you're developing it really closes the feedback loop a lot more aggressively here. Most of the Chinese models are optimizing around this very sparse-type structure, with DeepSeek and similar architectures, with Muon kind of acceleration. So you're getting toward one architecture that they can just engineer and output industrially. And who's the best at industrial manufacturing?

The challenge with China is that people don't trust it. I think we're going to see a Cambrian explosion—no pun intended—of architectures coming out of China now that China has been effectively decoupled from the U.S. tech stack.

So, stepping up a level, is this good for humanity or not?

I think the big question is: What does the finish line look like?

Now, that's a moonshot, ladies and gentlemen.

Anyway, I'll get us going in a second, but what a fun day yesterday. We were all in Seattle. Salim, we missed you. You were in Brazil. Good morning. You arrived from Brazil at 6:00 a.m. this morning.

Yes.

Yeah.

I got an hour and a half of sleep, and so I'm foggy as fuck.

All right. Well, hey, that means we all have a shot at you.

I'm just back from Rome, and Emad's in London. Let's kick off with covering the world we've got going here.

Yeah. Fantastic.

And Alex was just back from San Diego.

Yeah. And God knows—

Yeah, I just got back from Vietnam and Japan yesterday. Look at us, traveling.

Globe-trotting, gentlemen. It's like no time for sleep.

It really is. I feel like we're going 24/7.

Well, look, if you want to transform the world, you have to go out into the world, right? I think that's what we're all doing.

Yeah.

I think that's a good opener, too, because, Peter, you just got back late last night from Seattle, and that'll come out in a couple of days. If we just mention the whole world that we've covered in the last week, that's pretty good. Globe-trotting. But a lot is going on. Shall we jump in?

Yeah. Is that enthusiastic? Yes, everybody.

Yeah. Let's get into game time.

Make it happen. Engage.

Trying to find my neocortex.

It's there someplace. Don't worry about it. It'll show up. All right, everybody. Welcome to Moonshots. This is the conversation that's changing the world. Hopefully we can help you get ready for the future. And this is the news that if you're not watching the Crisis News Network and you have time to watch Moonshots, we hope we'll deliver to you sort of what's going on, what's happened the last week in AI, robotics, data centers, energy. It's a lot here with all of my incredible moonshot mates. We have a five-fold increase in capabilities here today because not only do we have AWG, we have Emad as well, Salim, DB2. It's going to be amazing. All right. Going to jump into our first stories. We're going to start with where Alex was last week: NeurIPS 2025. So, Alex, this is you.

Yeah. NeurIPS this year was a bonanza. I've called it in the past the Woodstock for AI. A few observations: It had more than 29,000 registrants this year, which is almost a 50% increase over last year. It was enormous.

Alibaba, the Chinese lab, had 146 papers accepted, including a best paper award. Anecdotally, the language that I heard the most in the hallways was Mandarin. There was a sense that the frontier labs have all the resources, while the academic labs do not, and the frontier labs were at NeurIPS mostly to recruit academics.

Again, this is sort of the sense of the conference, if you will. The frontier labs and the frontier research coming out of the labs have largely gone dark. So what was being shown on the research end was largely coming from academic resources or academically resourced labs that are lacking the resources. There was, on the sidelines, a sense that even though frontier labs were there to recruit, the publications and the oral presentations at this point were largely not coming from frontier labs, except for Chinese frontier labs.

So what does NeurIPS stand for? First of all, let's start with that.

NeurIPS, formerly NIPS, stands for Neural Information Processing Systems. It is the largest AI conference in the world. It's held once per year, every December, and it is where some of the most striking AI research historically has been published.

It's also the one time per year when all of the frontier labs are under one roof, and you get a real sense of the pulse of the AI space just from being there. Some of the most interesting meetings happen in the hallways. Optimus Gen 3 humanoid robots were there in force. You also get a sense of the vibes—the zeitgeist—of AI right now.

These are a bunch of photos and videos I took from the conference. We've talked on the pod in the past about how the Chan Zuckerberg Initiative has now pivoted to solving all disease with AI. That “solve everything” mentality—that math, science, engineering, and medicine are just going to be solved imminently with AI—was very much on the show floor, to the point where it made banners. It was really such a wonderful way to tap the zeitgeist of the space.

Amazing. We should—

It's interesting to see that a lot of it is hardware, and much more than you'd expect.

There's a lot of hardware, and the feeling of the moment is that robotics in particular is the next big thing after agents. I should also mention that a lot of the attendees watch and are fans of Moonshots. I think there would probably be a lot of interest if we were to do a recording in the future from NeurIPS, ICML, or ICLR.

Well, it's a global thing. It's all over the world.

It's kind of like the Olympics. It's in a different city all over the world every session. It just happened to be in America this year. But where is it next year if we want to record?

I don't think they've announced that yet.

It's in San Diego again.

Is it really back-to-back?

That's easy. NeurIPS 2026, here we come. All right, let's move on.

There's something interesting that I think I should point out. At ICLR, which is another similar conference, the number of papers with Chinese-affiliated first authors went from 9% in 2021 to 30% this year, and the U.S. number went from 52% down to 36%. I think we saw something similar with NeurIPS, but it'd be good to crunch those numbers.

I think we can't emphasize that dynamic enough. Emad, I'd be curious to hear if you have a different take on this, but my take on the floor was that American frontier labs have basically gone dark and are largely no longer publishing all of their internal results. Chinese labs continue to publish, in the same way that they're continuing to push open-weight models. So the research publication gap is being filled in part by Chinese labs.

I have a question for you about that. I know exactly why the U.S. is going dark. Everybody working on these big frontier models is a former Googler—not everyone, but almost all of them. The billion-dollar signing offers are a real problem, and we saw that yesterday at Microsoft, too. It's just recruiting warfare.

Google has explicitly gone dark after being very open and publishing everything for years. I know why that's happening, but why are the Chinese still being super open?

Well, it's because the Chinese are backing open source now, aren't they? It's like: deploy, use open source to make it more efficient, and then that's how we'll win.

That's how the models propagate.

Yeah. Strategically, it's great differentiation. If you have really strong American frontier models that are largely hidden behind APIs, under the spirit of “commodify your complement,” release lots of open-weight models.

And it's a land grab at this point, right? If a lot of nations, entrepreneurs, and companies are beginning to use the Chinese models, they have a foothold.

It's also an integration angle. Again, under the banner of “commodify your complement”—a classic economic strategy—there is much more to AI than just the models. There's integration with society and the economy.

So if, strategically, you're at a disadvantage on the model front, open-weight, open-release all of the models, and focus your attention—focus your economy—on deeply integrating all of those open-weight models. That becomes the competitive advantage.

All right, let's jump into our first article. Is this one from NeurIPS? Google's Titans and MIRAS are helping AI have long-term memory. Google's Titans is a new architecture with deep neural long-term memory that updates itself in real time. Do you want to jump into this, Alex?

One of the many blockades to radical progress in terms of advancing AI models is obviously the context-window size. It would be miraculous if we could have, say, 1 billion tokens in context or 1 trillion tokens in context. If we could have the entire web in context or the entire human genome in context, imagine the reasoning powers we'd gain and all of the problems that we could solve.

The problem is, as folks in the space have long used to motivate just about every recent paper on the arXiv, the quadratic complexity associated with increasing the number of tokens in the context. It's really painful to increase a conventional vanilla transformer past a few million tokens in context. So we resort to techniques like RAG—retrieval-augmented generation—and other techniques to try to effectively increase the amount of working memory, if you will, that an AI model can keep.

So Titans and MIRAS, these are Google's latest attempts to get past that bottleneck. As we've talked about in the past on the pod, there are a variety of architectural techniques to try to break the context-window limit, like recurrent neural networks. A popular example is that they don't have any explicit context-window limitations, but they forget.

The approach that Titans and MIRAS propose is sort of biologically inspired, distinguishing between short-term memory and long-term memory. It's sort of ironic, given that the attention mechanism itself that's powering basically the whole economy at this point was originally designed for differentiable long-term memory. We found ourselves back there. The idea is to use surprisal—a numerical metric of surprise—to decide what to commit to long-term memory and what not to. It turns out it scales really well without catastrophic forgetting.

Just to give people a sense of this, the models historically—GPT-4 and GPT-4o—had about 128,000 tokens as a context window. 2 million tokens, which we're talking about here, is about 3,000 pages of text, or 16 novels, if you will. And you mentioned the human genome, which is 3.2 billion base pairs. That would fit—2 million tokens only hits about 0.06% of the entire human genome. Will we actually get to a near-infinite context window someday? Is there a strategy for getting there?

Absolutely. I think these papers and others—when Demis Hassabis talks about there being only 1 or 2 major AI advances at the level of transformers left before we achieve what he'd characterize as AGI—I think breaking the context-window ceiling is probably at least half of one of those great advances, and I think we will absolutely get there pretty soon.

Yeah, I agree with Alex, and this fits with Google's hardware architecture. The massive, kind of toroidal TPUs that they're doing can handle huge amounts of memory—gigabytes, probably terabytes, of memory in 1 instance. By making it more efficient, from the graphs that we see on Titans and MIRAS, everything else kind of just slopes straight down. This goes almost continuously as you scale, without the quadratic-complexity overhead, because you used to need almost exponential amounts of compute as well as memory to handle that increase.

Being able to just capture everything in one go means that you don't need to store stuff to files anymore. You could have an entire picture of every email the organization has ever sent just in the memory at once, so it can track everything and figure out the interconnections dynamically. Again, I think that's what helps you break through to that next level of performance. It's pretty exciting, and I think it's a very logical approach that they've taken here with the surprisal element as well. Right now, transformers are a bit brute force, to be honest.

This reminds me of Ben Goertzel. Back around, I think, 2010 or so, he launched the OpenCog project, which was an open-source effort at recreating a mind. So they had different modules of what constitutes a mind, like memory, pattern recognition, sensory adaptation, et cetera. They were trying to replicate each module in software and then improve the software—typically, like Ben, very, very early for the timescale. As we look at this and we're building memory, and we've got the processing power, it feels like this new generation will get to that point and, essentially, without realizing it, actually grow a mind.

All right. I can't wait because, frankly, I would love to have perfect memory. When we add augmented-reality glasses and an always-on version of Jarvis, having an assistant there that's able to constantly remind you of everything you've ever known and everyone you've ever met is going to be super handy.

All right, let's go on to OpenAI. We heard about this in the last pod—we talked about their code-red response to Google's growth in dominance. Some news is coming out, announced today, that looks like GPT-5.2 will be coming online next week. This chart we're showing, which shows the benchmark for GPT-5.2 and Gemini 3 Pro, is still hearsay. This was put up on X. We don't know if it's in fact true, but just for fun, we're going to see once again whether we're leapfrogging model on model on model. We'll see xAI come out again with something, I'm sure, very shortly thereafter. Alex—

Yeah, the catchphrase I heard over and over again at NeurIPS this week is that this is a rat race. There are so many employees at the frontier labs who are just grinding away, competing in what they view as a rat race to achieve the frontier max in this case. I fully expect, as I've mentioned on the pod in the past, that we're just going to see leapfrogging on a near-weekly basis at this point.

Absolutely.

Yeah.

Until we get to the finish line. I think the big question is, what does the finish line look like? It's interesting that Sam went out and made this code-red announcement, which got picked up by all the media everywhere, right? It's an interesting strategy because he's putting the organization on alert. He's letting the world know that he's put it on alert. He got a lot of negative press from that, but he doesn't care. He just wants to refocus the organization. It really is an interesting management strategy. People comment that a good crisis is a terrible thing to waste.

I love that. That is fantastic.

Leverage that, right? You're really leveraging that, and I think that sends a message throughout. Look, Google did exactly the same thing, right? Sergey Brin said, "I'm coming back in. We're going into founder mode, and we're going to solve this AI thing," and they've done it. This is now the leapfrogging rat race, which is good for the consumer in so many ways. That's amazing.

Yeah. We were going to be sharing the conversation we had last night on our next pod with Mustafa Suleyman, who's the CEO of Microsoft AI. Just foreshadowing that, one of the conversations we talked about is safety. If you're in a rat race and everyone's just trying to leapfrog everybody else, it feels like safety goes to the sideline and it's just an accelerationist point of view over and over again. Emad, any thoughts on this code red?

Yeah, I think the code red makes sense because the competition is intensifying, and there's only so much attention that consumers have, as it were, for these kinds of models. This benchmark chart, I'd be very surprised at because, for example, it has Video-MME, and GPT-5 can't understand video at the moment. So it would probably be a brand-new architecture underlying that. However, all these numbers will be hit in the next 6 months, probably within a year, and that's why we're running out of benchmarks.

Well, Peter asked me, "Should we put this out there?" because we have no idea if this is real or not. It's just an internal leak. We'll know next week. We could look really stupid if this is totally wrong. But I wanted to make sure we put it out there in case anyone wants to trade on Polymarket because, right now, at the end of the year, you can buy ChatGPT or OpenAI at, like, 6 cents on the dollar. So if that Humanity's Last Exam number is right, that's just mind-blowing.

We'll find out next week, but I wanted to at least give everybody a chance to see it and make their own guess on whether this is real or not. And, like Emad said, we'll hit these numbers within 6 months no matter what, somewhere. The other thing that was really interesting last night at Microsoft is that we'll start adding a column to all these charts that has Microsoft's numbers, independent of OpenAI. That's the mandate now.

I think Mustafa was really clear that, yeah, we're going to be another column on every one of these charts and another line on Polymarket.

Yeah. Well, I guess one other point just to note: Sam has got a lot of capital to raise in order to implement the buildout that he's announced.

And I think this kind of “I’m willing to do whatever it takes to stay out front” is part of the strategy: being able to raise capital and get ready for an opening.

That’s a great point, Peter. This is as much for investors as it is for the general public and for the employees.

Yeah. But I think the bottom line for all of our subscribers listening is: expect this on a week-by-week basis, which is what makes our conversations on this pod so interesting. It’s like watching—I don’t know what the equivalent race would be. I mean, it is a horse race, but it’s continuous, with $1 billion per day or more going in to fuel this crazy competition.

I think the closest analog I can think of is a world war with multiple campaigns, multiple fronts, and multiple thrusts and initiatives.

Yeah. If you only have time to look at 2 numbers on this chart, look at the first and the last, because the first one is the one that’s most correlated with self-improvement and accelerating AI. Well, just read them for those listening here.

Well, I mean, it’s speculation, but the high bar on Humanity’s Last Exam is, without tools, 37.5%, which is really good, but—

Which is Gemini 3 Pro.

And it’s Gemini 3 Pro. Then with tools, it’s higher than that. It’s closer to 50%. Here, the speculation is that they’ve leapt all the way up to 67.4%, which would be crazy.

I mean, again, only Alex can answer those questions, as far as we can tell, on this podcast. I think Emad would do a good job as well.

I mean, seriously, they’re damn near impossible.

What does “with tools” mean for those who don’t know?

Well, the AI doesn’t just answer in 1 pass. It’s allowed to actually use a whole variety of calculators, and really any kind of tool that doesn’t give it the answer, it’s allowed to use in its chain of thought. It can iterate many, many times. So it’s not just the standalone LLM; it’s the LLM accelerated or enhanced with other software, which is perfectly fair if you’re trying to solve a world problem, cure a disease, or whatever. That’s why they use that benchmark in addition to the raw benchmark.

Yeah. And then the last line is the one that Alex loves, for good reason. Self-improving is 1 thing, but then self-financing is another. And Alex, you talk about that better than anyone. I’ll hand it to you for that.

Yeah, Vending-Bench 2 and Vending-Bench Arena. To the extent that we have any sort of economic Turing test or economic benchmark for an agent’s ability to autonomously deliver a return on capital, that is what we have right now. As I’ve mentioned in the past, maybe just project this into the ether: I would love far better benchmarks for measuring economic autonomy than Vending-Bench, but Vending-Bench is what we have right now.

Because it’s coming.

Actually, there’s another thing that just crept in—it didn’t make it into the slide deck—which is a trading competition with AI bots, I think, on the crypto side of things. There was a mysterious model that actually made a profit reliably all the way through, and Elon Musk just announced that it was Grok 4.2.

Nice.

So the leapfrogging continues. He said that’s how you pay for all the GPUs: just don’t let Grok 4.2 go wild on the stock market.

So you’re going to compete against Elon and his million GPUs?

Dollars are the best benchmark.

All right. Anthropic is making news once again. There’s a lot of excitement and energy, and right now it’s still a bit of a rumor, but they could have an IPO as early as 2026. Anthropic is negotiating a new funding round that could value them at $300 billion, as revenue is projected to reach $26 billion next year. They’re aligned alongside OpenAI, which is also exploring a future IPO. So, if we’ve got OpenAI going public to access capital and Anthropic going public, I have to imagine xAI is also going to go public sometime in the near term. Thoughts on this one, gentlemen?

I’ll comment. Oh, go ahead, sir.

No, go ahead. I’m tired.

I think, if anything, xAI is relishing not being public, given its history. But I think more broadly, the worst-case economic scenario for superintelligence—or maybe not the worst, but the next-worst case—is that it remains decoupled from the human economy, and that an intelligence explosion happens not on the publicly traded markets. Insiders, early employees, and the machines themselves sort of skyrocket in terms of real wealth but are largely decoupled from retail investors and the rest of the human economy. That would be, I think, a highly suboptimal economic outcome.

Whereas if we get enough IPOs from Anthropic, from OpenAI, from SpaceX, and from some of these other firms that are achieving hyperscale on land and in space, I think that is probably the best case from a macroeconomic perspective for the economy. And once those happen—assuming they happen—I think it’s far clearer to see how the economy grows past the so-called debt crisis, how we achieve hypergrowth, real hypergrowth, over the next 3 years.

Dave?

Well, having founded and taken a company public, this is a big step. When Dario, the CEO of Anthropic, Dario Amodei, gets interviewed, he says, “I actually never visualized myself being a CEO at all. I’m surprised I’m in this position.” Then, when you become a public CEO, it’s a whole other level. So I’m guessing he’ll grow into the role, but it’s a big deal.

Why do it? And why would Elon be relishing not doing it? Once you’re in the public limelight, all your dirty laundry and your code reds become crystal clear in your stock price every day. It’s very hard to do what Sam does right now—roam the world selling the story—when your dirty laundry is right there on every stock ticker. It’s another level, but they have to do it because you can raise $10 billion or $20 billion privately, but you have to tap into the public markets. You’ve got to be talking about much bigger numbers, and the only way to do that is through the public markets.

It gives you currency for acquisitions, which is important. I think it also increases trust in the company if it’s a public company versus being a private company.

Yeah. Sam was at Davos last year, and I was following him around on the streets of Davos.

I mean, just the backing.

You couldn’t miss him. He had an entourage as big as a president of a nation. It was—he was just doing meeting after meeting after meeting, saying, “Give me another billion dollars. Give me another billion dollars.” And you can see how there’s no way that can scale to what’s happening next.

I mean, part of what’s coming out in the news right now is that a lot of the deals Sam and OpenAI have announced are options and not actual deals, which is fascinating.

Yeah, I think you’re at this fascinating time now, though, whereby the dollar benchmarks are starting to accelerate. The revenues ramped up, like Anthropic’s at $10 billion of revenue just a few years in. It’s amazing, and they’re actually catching up to OpenAI. But then the competition is going to get intense. Opus 4.5 is $25 per million tokens. Grok 4.1 is $50. Will you see substitution occurring, or will you see these models actually just being used to literally make money? I could easily see Elon, in particular, just say, “Grok 5 is going to pay for itself and all the GPUs by being the best hedge fund in the world,” and just let it loose on the stock market. And MCP is going to replace all the SaaS.

Yeah.

You know, the other thing that happened this week, concurrent with this—it’s not in the deck—is that the cost of HBM memory, the memory that drives AI, skyrocketed. It went way up. OpenAI was in the news saying, “We’ve reserved 40% of the world’s supply of memory for our own data center work in Abilene, Texas, at Stargate,” which is crazy. Normally, the cost of memory comes down at a rate of almost half every year, so to see it go the other direction is the first bellwether that there’s going to be a huge shortage of compute. If you don’t go public, raise the capital, and lock up the supply, you’re going to be left compute-starved.

The same thing is going on in fundamentals, right? I didn’t put it in the deck, but there’s a copper crisis right now. The price of copper is going through the roof because of wiring these data centers.

All right, let’s move on to our next story here.

Wait, quick point.

Yeah, please.

If their next-year revenue projections are $26 billion and the valuation is $300 billion, that’s only 10 times revenues. Palantir is trading at 111 times revenues, so that’s cheap in that sense.

No, it is reasonable. It’s crazy big numbers, but it’s perfectly reasonable. I think Salim is saying it’s overly reasonable. It should be at a higher valuation, and it will be. It’ll probably spike after an IPO.

OpenAI finds Confessions can keep language models honest. The Confessions method trains models to openly admit when they’re hallucinating or when they’re broken. The method encourages models to self-report mistakes instead of hiding them.

So, I am completely curious here. You use this methodology. Do you actually get 2 reports: “Here’s my answer” and “Here’s my confession”? Emad, what’s going on here?

This whole thing with next-token prediction is that the models kind of go along, and then they can’t have the self-reflection and things like that. I think that when you’ve actually got the right prompts, the right planning, and the right loops, you get very interesting things occurring.

The hallucination rates are dropping now as well. The models used to jump a lot and skip these behaviors because they didn’t have the self-reflection or some of these other things, so I’m not very surprised by this. Again, I think what we’ll eventually see is what we saw with the DeepSeek V3.2 math paper: a concept called a meta-verifier, where the models learn from their mistakes.

Rather than just checking against a very small baseline—“Are you being honest? Where have you made mistakes?”—having that as a verification loop is very similar to how humans learn. That’s what causes big leaps in some of these more frontier areas of thought as well.

What I found shocking is that I asked one of the models how often it was hallucinating or providing wrong answers on average. I found a couple of studies, and one of them said there were 25% wrong answers on everyday user questions. Another said GPT-4o and Claude 3.7 Sonnet hallucinate an average of 15% to 16% of the time.

I never expect that when I’m asking my questions. I’m assuming—and I think the majority of everyone, perhaps not you and Alexander, is assuming—that the answers are correct. If it’s really 10% to 25% hallucination, that’s scary.

It’s basically the same as a human doctor, right? The interesting thing, though, is that it has dropped. GPT-5 dropped it from 18% down to 3%.

So, the last generation of models—

Yes. Go ahead.

Can I give you a dramatic example of this? I think it’s worse than a human doctor because the models are actually trying to please you. Therefore, they’re saying whatever.

I had a TV where the power went bad, and ChatGPT said, “Oh, if it’s this model and it’s making a buzzing sound…” Then I asked it, “How do I know who can fix this?” It gave me the names of 3 local TV repair shops that were completely hallucinated, with phone numbers, websites, names, and addresses.

I started calling them. None of the phone numbers worked, and then I looked them up. None of them existed. This is a big problem.

If I lift up a level for this confession thing, I think it goes back to the earlier point that it’s great to have a feedback loop. It gave me somewhat of a chill because I went to Catholic high school, and the idea of confession is somewhat chilling. Who’s the priest, is my question, when you do this type of model? But I think the feedback loop is very powerful to have.

I think it’s also worth mentioning the 50-year-old notion from economics of Goodhart’s law, which is that when a measure becomes a target, it ceases to be a good measure.

The way that these models are trained certainly, and superficially, rewards various sorts of behaviors that might be construed as dishonest. Being able to avoid Goodhart’s law—clever ways to avoid Goodhart’s law—I think are admirable on the part of OpenAI.

Maybe the final solution looks a little bit less like naively optimizing just next-token-prediction objectives or reward-maxing on reinforcement-learning objectives to solve math problems or programming problems. It looks a little bit more like some sort of multi-objective optimization problem, where there’s some blend of a Goodhart-avoiding honesty reward, an accuracy reward, and an ethics reward.

That may almost start to look a little bit like separation of powers in the government systems we see. There’s an executive in some systems, a legislative branch, a judiciary, and so on.

I’m going to have to bookmark that comment, Alexander, because you lost me at “multi-objective optimization problem.”

But you have a good heart, see.

On that note, I’m going to move us forward here.

All right. The next release is Google’s Gemini 3 Deep Think, which uses parallel reasoning, testing multiple solution paths at once. That makes sense to me, right? It’s an upgrade from 2.5 Deep Think, and it excelled once again at Humanity’s Last Exam, GPQA Diamond, and ARC-AGI-2.

Going to our benchmark expert, Alex: this is a template for how revenue is going to scale to justify the trillions of dollars of capex. It won’t just be faster models, better models, or stronger models. It’s going to be lots of agents, fleets of agents, millions of agents, many millions or billions of agents, all running in parallel to solve problems.

That is, in my mind, and certainly based on the architecture of Gemini 3 Deep Think, which isn’t just a faster, better singular model. It’s also scaffolding to have fleets of agents—fleets of Gemini 3 agents—all running in parallel to solve a given problem. That’s the Deep Think part of Gemini.

It sounds like quantum computing to me. It’s like, “I’m going to run this problem in multiple parallel universes and bring back the answer.” I’m going to run this problem in a billion or a trillion agents and bring back the best answer.

Parallel, yes. But quantum computing wishes that it had the economic utility of Gemini 3 Deep Think.

I’m with Peter on this one.

Go ahead, Salim.

I’m with Peter. It feels like that kind of parallelism. But I think the broader point you’re making, Alexander, is that when you have millions of agents, each specialized—if you take something like the Manhattan Project, you have thousands of people, each with a deep specialty, connecting together, and the hive mind then solves the problem. We’re going to see the same with agents. Is that a metaphor that works?

Yes. When Dario speaks of countries of geniuses in a data center, this is what it looks like. We’re going to have billions of agents that are all going to be independently, probably pretty expensive, even though the cost of intelligence is going to zero.

Collectively, yes, this is going to generate trillions of dollars in revenue if we have so many agents. This is how we pay for all those data centers.

I think that’s a really important point, Alexander. Everybody talks about the cost of intelligence going to zero, but it’s not actually going to zero. It’s going down to a low number, but concurrently, the fleets of agents are getting so much bigger so quickly.

We saw earlier in this podcast the process of iteratively expanding the context window to hundreds of millions of tokens and running many iterations to get rid of the hallucinations. Those forces are going in the other direction, and it’s working far better than anyone thought it would.

It’s very unlikely that the cost of intelligence is going to go anywhere near zero. Everybody’s going to want more intelligence, and more intelligence, and more. There’s going to be an acute shortage.

I only mention that because many business leaders are out there saying, “Let me just wait and see what happens,” and they’re going to be starved of access. You can see this already when the new Gemini models come out and add another level of deep thinking. It works incredibly well, but you wait 3, 4, or 5 minutes to get your answer.

Very often it says, “We’re experiencing unexpected loads right now. Sorry, we’re offline.” How’s that possible if the cost of intelligence is going to zero? It’s not. The cost per token is going way, way down, but the use cases are expanding on at least 3 different dimensions on this ridiculous curve in the other direction, and everybody’s going to want it because it works so well.

So, what does this actually mean, other than that we have yet another faster model able to hit the benchmarks a little bit better, and next week we’ll be announcing the next better model? Emad, what do you think?

I think the models are getting to the point now where ensembles of these models, doing different things, are genuinely useful for real-world, advanced, complicated tasks. I’m really looking forward to using Gemini 3 Deep Think. 2.5 Deep Think was pretty good, but right now I think the only one usable for real frontier stuff is GPT-5.1 Pro, which again does something similar.

Ultimately, what you want is not for a task that’s really complicated to be done in seconds or minutes. You want to be iterating on a task for a period of time, giving it input and feedback, and having the model just not make mistakes.

Gemini 3 Pro still makes mistakes in math when I’m using it, whereas GPT-5.1 Pro doesn’t.

The model usage, again—I think the DeepSeekMath paper is fascinating for this. It’s the first open-source model that gets a gold on the IMO, and for each of the problems they use 2 billion tokens, so about 2 billion words.

That gives you an idea of how many more tokens you can use.

Yes, and it works. It works.

And it works. It got—

If you want to experience this firsthand, just go to GPT-5.1, soon to be 5.2, and ask it to do something complicated for you that it can’t quite do. Keep asking it to try harder about 1,000 times back to back, and it will eventually get it right.

You're like, why does that work?

What's an example of a hard thing to do?

I do this all day long. In fact, right after this podcast, I'm back. I have a fleet of Kimi K2 agents scouring the world right now, working on hard problems.

If I ask one to translate legacy C code into Python and it comes back slow, I say, "Make it faster. Improve the algorithm." Or, more down to earth, do you do that with your students, too?

Do you do that with your students, too? Work harder, try it again. [laughter]

Well, that's the difference. That's where everybody's analogy breaks, because there are a limited number of humans and students, but there are an unlimited number of AI agents. Deploying 1 billion in parallel to work on something is out of your normal range of intuition, but it just flat-out works. You need to expand your intuition to this new world we're moving into.

I think that's one of the most important things that we can say coming out of this: we're about to enter a new world where there is a near-infinite amount of intelligence to be thrown at things.

Yeah, anyone.

Yeah. Go ahead.

Interesting things. Sorry, please. Sorry.

Go ahead, Emad.

I think one of the super-interesting things in the context window, when you just realize it, is, again, the stuff that we get wrong when we're trying to solve problems. Usually, the only stuff that survives is the stuff that we get right. At NeurIPS, it's papers full of all the stuff people got right.

Being able to actually have scientific-method strategies and things like that is possible when the context window includes everything you got wrong for all the different models.

Fascinating.

We know that will do better because often it's what we got wrong that actually guides us.

Learn from your mistakes. Yeah, sure.

Yeah. I mean, if you had asked everybody in the community 3 years ago, "Is that going to work?" 90% of people probably would have said, "Yeah, I really doubt it." But it does. Now, everybody agrees. Everybody at NeurIPS, I'm sure, agrees that it just flat-out works.

That means you need massive numbers of parallel agents, and even any given agent needs many iterations. It's just a huge amount of compute, but it solves problems. It's incredible.

All right, our next story here is "Reasoning AI Became So Efficient." AI got more efficient, mostly for huge LLMs, with training becoming 22,000 times more efficient, while smaller LLMs only improved by 10× to 100×. So, what's the story here, Alex?

Yeah, this is sort of a finger in the eye to those armchair theorists who say that the small labs are going to benefit from algorithmic advances. This is a study out of MIT that found that 91% of algorithmic efficiency gains between 2012 and 2023 were the result of only 2 things: 1, the switch from LSTMs to transformers; and 2, the switch from Kaplan scaling—named after my former office mate in the Harvard physics department, Jared Kaplan, at Anthropic—to Chinchilla scaling.

And the switch from Kaplan scaling to Chinchilla scaling. Those 2 things. So, LSTMs to transformers.

What are LSTMs?

Long short-term memory. LSTMs were the favored language model prior to the transformer revolution. I remember the old days prior to transformers, prior to GPT, when Andrej Karpathy had his char-RNN language model that stunned people by being able to generate code. There was a life prior to GPT.

Those 2 algorithmic transitions—LSTMs to transformers and Kaplan scaling to Chinchilla scaling—were 91% of the efficiency gains. What that says is that this story that we're just stacking small wins on top of each other, and that eventually algorithmic efficiency gains are going to enable smaller labs to have some sort of advantage relative to larger labs, is probably not true. Most of the algorithmic efficiency gains are actually accruing to the large labs that are able to scale out the most.

There's one other thing in this paper that's noteworthy. Please don't read it. Alex summarized it perfectly. That's everything you need to know. [laughter] It's way longer than it needs to be. Classic MIT work, but it's a very, very good summary at the beginning of the document of why this work is so important.

We're putting an immense amount of societal energy into scaling the hardware. Elon Musk talks about Tennessee all the time, and Stargate, and there's a huge amount of thought, research, and discussion on our podcast about these massive data centers because they're so visual and so expensive. But the software side of it is very under-researched and very under-analyzed.

They're taking a first shot at trying to give us better insight into the future rate of improvement on the software side, because that's where it isn't as expensive, but the lift could be enormous. I think this is a really, really good focus area, and I'm really glad MIT is on top of it. But Alex's summary is all you need to know about the work so far.

Amazing. So the inner loop here is energy going to GPUs, to agents, to intelligence, and therefore all of that scales. The demand is so infinite in terms of adding intelligence to everything that it'll be a long time before we run out of it. That's why we're tiling the world with data centers.

I think there's some YouTube viewers somewhere with a drinking game or a bingo game for how many times we can say "tile the Earth with compute" or "disassemble the Moon," or whatever it is. Drink your whatever, or cross off your bingo game, for this one. [laughter]

Silly robot. Robots are part of the loop.

We're robots in the loop.

This next article is really important. This is about visual chain of thought. The notion is that visual chain-of-thought methods are now able to give us a better understanding of images, and this visual chain-of-thought delivers 3% to 6% gains in visual reasoning performance.

The image here is asking the question: Is the wall behind the bed empty, or is there a painting hanging on the wall? What you see is the analysis—the ability for an AI to understand what it's seeing.

At the same time that we're bringing about augmented-reality glasses and humanoid robots are coming online, this is going to be fundamental. I want my AI to understand what I'm seeing. I want it to be able to remember, during the course of the day, where I left my keys, who I ran into, or recognize a face and give me their name. How fast is this accelerating?

I think this is even more profound than just garden-variety acceleration. If I ask all of you—or I request, "Don't think about pink elephants"—what's the first thing that happens? You start thinking about pink elephants, not in terms of text tokens in your mind. You're probably not thinking in terms of language. You're using your visual cortex to create a mental image of pink elephants.

The ability to visually reason is something we've talked about on the pod in the past. We're finally, a few weeks after this was predicted to happen, starting to see major gains in reasoning performance by models that can include visual tokens in their chain of thought, not just text tokens. We're going to see a lot more of this.

How excited about this are you?

Yeah, I'm not surprised at all by this. It's very exciting. I think it's actually something fundamental to reality. Models are the things with the best math that approximates reality.

We've seen some interesting things before. Originally, we built Stable Diffusion, and then from Stable Diffusion we extended it to 3D using the same knowledge. We found out—actually, Harvard did a study—that an image model understood 3D. Then we extended that out to video. The same thing: it somehow actually had a concept of physics in there.

In fact, if you look at the latest image model that's at the top of the charts now, Flux by the Black Forest Labs team, my former colleagues, it started with a language model that then got a video model that then became the best image model in the world.

You see now, for example, Luma recently raised $900 million from Humane and others to build world models, where you input all this data—image, video, text, et cetera. Text is low-dimensionality. If you actually want to understand and reason, you need to have all the different types of data.

The latest models are actually very, very similar to one another in terms of understanding the universe. You can go from a text model to a video model simply by adding the right types of data, but the underlying structure doesn't change. I think that has big implications for the actual nature of reality itself, because each of those is modeling a different part of reality.

Can you imagine going back to AlexNet, when they were putting this together and showing them this capability, when all of it was trying to recognize the number 7? That was a conversation we had yesterday. I mean, truly extraordinary.

It is. To experience it firsthand, take a screenshot of something you're doing on your computer, dump it right into Gemini, and say, "Help. What's going on here?"

It's incredible that that works. It would have shocked anyone 10 years ago. Nobody would have believed you at all.

And go ahead. Go ahead, Emad.

Well, the crazy thing—I think it's always worth coming back to this—is that if you told someone 10 or 20 years ago, they would have thought it'd be this massive logic tree, right?

We have to remember: is it this, or is it that? Right?

They're just ones and zeros. They're literally like a movie file, and you push words in or images in one side, and it squeezes out the stuff on the other side. The reasoning isn't actually reasoning at all in the way that we think about it. And again, I think that says something profound about the way our brains work and the way the universe works. This static group of ones and zeros can do that.

I think what's important to realize here is where we're going. All of us are going to have an AI with visual capability that's always on, helping you and supporting you, right? And I think that's a vision of the future. People say, “Well, I don't want to lose my privacy,” and so forth, but it's going to be watching what you eat. If you want to turn on health mode, it'll tell you, “Eat more of that. Don't eat that.” Or, “There's a staircase over there. Go take the stairs instead of taking the elevator.”

I mean, the ability for an AI to be your always-on visual JARVIS assistant is going to be profound throughout our lives, increasing our efficiency in what we do and what our objectives are. Yeah. Salim, do you want to add on that?

Just to build on that point, Salim, I'm expecting in a year or 18 months, some sensor that's in your stomach saying, “Hey, you're about to eat that doughnut. Wait 10 minutes because I'm still metabolizing the coffee.” Right.

Creating radical efficiency in all these very little things that we never thought about much is going to be one of those areas where we're going to add a ton of compute against. So, I took a screenshot of our podcast as we're speaking and gave it to Gemini just to prove the point. And I said, “Hey, are these guys having fun?” It completely interprets the scene. It knows exactly what we're talking about.

And it says, “Yeah, it looks like it's fun if you're a tech enthusiast, you like futurism, or you enjoy brain food. It's probably not fun if you dislike technical jargon or you want casual entertainment.”

Okay, that's probably true.

The point is, it completely knows what we're doing from just that screenshot. And this is going in 2 different directions, too. It's making the AI more in touch with humans and the way we live. But the data is not specific to that. It can also go the other direction, where you feed in genetics data, you feed in satellite image data, and it can then get intuition in those domains where nobody that you know has intuition.

So, it's going in both directions at the same time. If you study what's going on with vision on this slide, you can develop some intuition about what it's very soon going to be capable of with medical imaging, satellite imaging, and other types of sensors that we're not familiar with.

I'm still reeling over last week's comment from Alex that we're taking brain scans and running them through AI. I mean, that's going to generate some unreal insights.

Well, Emad has thrown some cycles at that previously. Yeah, I know. I was catching up with some of your former colleagues from the MedARC days in Europe. fMRI wants to be its own modality.

It does indeed. I think all the modalities, again, we need to tell the world for some reason, right?

All right. We're going back to one of the stories we opened up with, which is the response China is having, or the leadership it's providing. China is accelerating its push to become independent of NVIDIA, with Cambricon planning to triple output to 500,000 accelerators in 2026.

So, this is a response to U.S. policy. It's always that way. As soon as we restrict a country from buying a product or service that we're providing, especially if it's fundamental to their lifeblood, they will develop competition. Without question, I think the competition will—there's a huge amount of intelligence resident in China. Don't forget, the Chinese educational system excels at math and compute, so I would not expect that they would deliver anything sub-NVIDIA. Emad, you want to kick us off here?

Yeah, I think necessity is the mother of invention, right? We also saw Moore Threads IPO this week in China. They raised just over $1 billion. They're, again, another GPU competitor. It was 4,000 times oversubscribed.

So, $4 trillion of demand.

Now, obviously, that's a bit much, but again, you're going to see more and more of this stuff ramping, particularly for the specific Chinese models, because having an open-source model that you can test as you're developing it really closes the feedback loop a lot more aggressively here.

So, you don't need just to build for one vendor; you can build for everyone. And most of the Chinese models are optimizing around this very sparse-type structure, with DeepSeek and similar architectures, with Muon kind of acceleration. So, you're getting toward 1 architecture that they can just engineer and output industrially. And who's the best at industrial manufacturing? Why, it's—

China has been. It's important to remember NVIDIA used to supply 95% of China's advanced AI chips.

And when that supply got cut, there was a red alert going on in China. I'm sure the government orchestrates and supports it and says, “Okay, we need our own NVIDIA, or multiple NVIDIA companies, in China.” Alex, your thoughts?

Yeah, we don't have a slide for this, but I would definitely encourage the audience to read the National Security Strategy that was just released in the past 48 hours. It is most certainly eye-opening, and I think it spells out a pathway for tech decoupling between the U.S. AI tech stack and the Chinese tech stack.

I'm reminded that during the Cold War, the Soviet Union had what, for those years, would have amounted to an independent tech stack and was experimenting with all sorts of crazy architectures, like ternary computing and other unconventional choices. I think we're going to see a Cambrian explosion—no pun intended—of architectures coming out of China now that China has been effectively decoupled from the U.S. tech stack.

Maybe many of those innovations will end up, one way or another, benefiting the overall world and benefiting the U.S. tech stack. I think we'll see a lot more experimentation coming out of China post-decoupling.

So, what's the implication of this? I'd like to spend an extra couple of minutes on this because China is going as rapidly as possible developing its models. It's developed its energy ecosystem fully, 10 times further than the U.S. has. We're going to talk a little bit about China's desire to put data centers in space.

I mean, any thoughts on the long-term implications of this complete parallel development between the U.S. and China?

I think we see an intelligence race, and that will lead to diversity. We're going to see so many different architectures that are all competing. In the U.S., in the West, we have a whole handful at this point of frontier labs that are all vertically integrating with their own chip architectures, many of them in partnership with Broadcom or other lower-level infrastructure providers.

Now we're starting to see the same happen in China. I think in the end, this heterogeneity that we're seeing in terms of tech stacks is only going to further accelerate the race that we're already in to the finish line. And again, I would pose the question: What is the finish line that we're racing toward? Because we're going to go much more quickly with this level of integration.

So, stepping up a level, is this good for humanity or not?

I think, all other things being equal, more experimentation can probably be better for humanity. I query whether it's good for the U.S. or not, and I query whether it's good for interoperability or not. But, all other things being equal, more experimentation is probably better.

Emad and Salim, I'd love to hear your thoughts here.

I think the good news is that more technology development is generally better for the world. If I think about the counterposition between the U.S. and China, I think a lot of the future will depend on where you end up with trust. The challenge with China is people don't trust it. Now people are losing trust in the U.S. on a week-by-week basis. So, there's that to be considered.

But I think over time, the concept of whether you trust Google or whether you trust ChatGPT, in terms of what the future of AI is going to be, a lot of it is going to come down to where we place our trust.

If you're an African nation over time, where will you put your trust? Right.

Really important. I saw a tweet today to entrepreneurs saying, “If you're building something that increases trust, double down. If you're not, then stop doing it.” I think trust as a scarce asset is a really important thing to—

And I got to shout out to Jerry Michalski here, who made that phenomenal comment that scarcity equals abundance minus trust.

It's just amazing.

Yeah.

Yeah. I think you'll see, just like China flooded Africa with smartphones from TCL and others, these chips will be very aggressively priced. So, to put Cambricon in context, they raised about $2 billion. Their market cap is $100 billion right now, and the 5090 is equivalent to an NVIDIA A100. The 6090 is about an H100, but it's about half the cost.

And it's much more power-efficient. Does this hit NVIDIA's bottom line?

Not for a while.

For a while, they'll all be used locally, but as they ramp from 500,000 accelerators to 5 million and more—and, again, China has the full end-to-end supply chain as well—then you'll see them flooding in a few years' time. To put the 500,000 in context, I think there were about 4 million Hoppers sold and about 10 million Blackwells coming. So, in a few years, you can expect even Cambricon—someone no one's really heard about—to be at 80% of a generation-back NVIDIA chip.

In a few years, you can probably expect them, just like Tesla and BYD, to actually be fully competitive. And now you're seeing BYD displacing Tesla over and over again.

Yeah. It also comes down to the engineering-versus-legalistic thing, right? The US is lawyers managing immigrants and engineers, and China is all engineers with an authoritarian state. Where will this play out?

Emad, you're kind of in Europe, or you're in a peripheral European nation. We've had this conversation on this pod over and over again about Europe being in an AI winter—or ice age, as the case might be. So here we see, “EU to open bidding for an AI gigafactory in early 2026.” Europe is finally making a serious move to close its compute gap with the US and China by greenlighting bidding for an AI gigafactory in early 2026. Emad, what does this mean?

Every nation needs sovereign compute because the intelligence of a nation will be dependent on the number of GPUs, right? The EU, with its regulatory acts, has kind of held AI behind. But recently we saw Yann LeCun hiring teams in Paris. We see teams in Germany, and we have Stability AI teams. There's a lot of talent there; they just have to cut the red tape.

We're seeing a change in that, and now the UK, Europe, and others are saying, “Well, this is the future. We have to cut the red tape.” But the US is still far, far ahead.

Can they move fast enough? I saw they're going to relax GDPR to finally give access to data. GDPR was just a chokehold on entrepreneurs. How are you feeling in the UK right now?

You've seen a step change just in the last few months. As the agents hit next year—proper agents, not these stochastic parrots—everybody has to change. No country has an option but to change and go all in on this, because if you don't, you're going to be left behind. You'll be outcompeted by your peers.

Well, I'll tell you, just observing without judging: there's a square mile in Palo Alto, a square mile in San Francisco, and a square mile in Cambridge—Cambridge, Massachusetts.

Yeah.

Not the other Cambridge, not the original Cambridge. The gap between those three square miles and the rest of the world is getting wider and wider at an incredible rate. I have the same observations that Emad has: amazing talent all over Europe and all over the world. Shouldn't this proliferate out to all that talent?

But when I observe that meeting we had with Richard Socher, Peter, I heard you go, “Holy,” about 12 times during that meeting.

There's nothing on the planet, and we'll talk about that soon.

Not that we can say anything about it, but the gap between what's going on in those three square miles of the Earth and the rest of the world is mind-blowingly big and accelerating very, very quickly. So I think building data centers around Europe is way too little, way too late, unless it's done in combination with some other force that I don't know about yet. Just observing it, that gap is accelerating very quickly.

Europe is best at public-private partnerships. The problem is that in this world, speed is the ultimate differentiator, and speed is not its strength. You and I have had so many meetings throughout many of the European nations, and the energy isn't there—the drive, the absolute urgency.

My favorite Joseph Campbell quote is, “A man whose hair is on fire seeks water.” That's what we're seeing right now in the hyperscalers here. It's code red; everybody is jumping in. It's 24/7. It's not 996. It's—I don't know what it is—6-12-7. It really is that way.

Not just that, though. I tried to make this point with Mustafa yesterday, and that'll be on our next podcast. You'll see it. He just name-dropped: “When I was talking to Sam the other day,” and “I saw Demis last night,” and “Sam and I were thinking about Dario.” It's all first names. All of these are just first names to him.

It's not corporate. It's not data-center investments made by the government. It's this very small group of people who are on a first-name basis and now have, no joke, $50 billion budgets to build this out. So that's what's really happening.

We saw the same thing in the space industry, right? In the United States, we gave birth to Blue Origin, SpaceX, Virgin Galactic, and a whole bunch of entrepreneurial space companies. In Europe, it's the industrial-military complex creating Ariane 5 and a few other smaller rockets.

They can't compete with the current entrepreneurial space industry. The only way to compete is because a government buys local, and that simply makes the entire space-based services launched out of Europe more expensive. It just can't be supported.

It's changing so quickly. It has to be on a first-name, informal basis at the rate of change. In Massachusetts—this will drive Alex crazy—but in Massachusetts, we have a very good relationship with our amazing governor, and she said, “You know what I need to do? Put together an AI task force.”

In Europe, they would say, “Sometime in Q3 of 2026, we'll start the discussion to create that task force.”

There's a very interesting initiative called Next Frontier AI to build frontier AI labs. Again, it's well-intentioned. They'll give 12 teams €2 million over the next few years to see if they can accelerate and get there.

€25 million per second.

Can I double down on this just for a second? We are working with one of the biggest companies.

I have to apologize to our European listeners. We don't want to make fun of the situation there. It's serious.

Just spend 2 years in the middle of it and then go back home. That's the way to solve the problem.

Look, I spent all the 1990s living all across Europe, in about 5 different countries, so I've got some personal experience here. In terms of the ability to live and have a great life, Europe is amazing. But in terms of technological progress, it's not really the place to be. You have to move to the West Coast and elsewhere.

We are working with one of the biggest European companies—one of the biggest global companies—on transforming their metabolism. We finished one of our major sprints with them, and they said, “We need to start another one right away.” This was back in February, and they said, “Let's have the first meeting about it in October.”

Yeah.

This is the problem. People aren't seeing that the metabolism of everything happening needs to accelerate by 100 times across Europe for them to jump and get there.

It's culturally blocked. There's an old adage in Europe: You work to live. In the US, you live to work.

Yeah.

Which may not be the very best thing.

Now we're going to live to compute.

Yes.

All right, let's jump into jobs and the economy. A few interesting articles this week. The first one is Michael Dell's $6.25 billion investment in America's kids. It's called Invest America, and it will give every child born after January 1, 2025, an investment account of $1,000, deposited to build financial security. Let's take a listen to Michael and Susan Dell describing what they're doing.

We're making a $6.25 billion investment in America's kids through our charitable funds.

Next year, every American child will be able to get an investment account powered by Invest America. We've seen what happens when a child gets even a small financial head start. Their world expands.

The real power of these accounts is that anyone can contribute. Parents, relatives, friends—everyone can help shape a child's future.

To philanthropists, companies, and community leaders: If you want to be part of something truly meaningful for our kids, for our communities, for our country, join us.

I celebrate them for that effort. There have been a number of players who have talked about a similar situation. Of course, being able to invest versus just save is part of what made America great. Now the question becomes: Is this too little, too late?

Or are we just flat-out irrelevant, given that all education is moving to AI? You know this better than anyone, Peter. I mean, one of the questions became: Maybe what's going to happen is every single kid will have an AI agent that's out there generating revenue for them, right? Where was the conversation we had about this? No, it was the conversation that Ilya had in his recent podcast, saying that one of the problems is, if you've got an AI agent that's doing all this work for you, generating revenue, supporting you, representing you, the question is: Does the human fall out of the loop, become irrelevant, and is it important, instead, to ultimately merge with AI? But that's a different conversation.

That's a different conversation.

But I think I'm 100% sure of this topic. You know, teaching at MIT, Harvard, and a little bit at Stanford, there was a huge push to move all educational materials online when the internet exploded, and it kind of worked and kind of didn't work. It made all the material available.

I am 100% sure that now that there's an AI face and voice that matches your personality on top of that, it's going to absolutely take off.

Sure.

And traditional education will be completely irrelevant imminently because it can match your accent, your favorite voice, your favorite star—

Well, I think this is more than just education, right? I think this is more about how you provide financial stability. We've talked about this on the pod before. This was back a few episodes ago, on the data from FII9, that the majority of the world is absolutely concerned about not being able to be employed and about the cost of living. If you have a seed kernel of capital when you're born that's growing by the time you're 18, does that give you some additional stability? Salim, you were going to say—

I have 2 thoughts about this. One is, I really, really love the fact that it includes every child born and is essentially universal. It's creating a wealth floor for every kid, which means every kid from day 1 will be thinking about investment and how to think about it. I didn't come across the concept of investment until I was 16 or 17. If I had had that way earlier, I'd be a way richer person today than anything else, and I think that by adding the employer add-ons and encouraging people to contribute, you're creating a community effort here. So I think it may be late to be doing this, but at least it's being done, and I've got to applaud them completely for doing it.

Well, I think also it's just money, so it'll be pivoted over to access to compute. You can say, look, education doesn't have to be tuition; you can also get your GPUs that you otherwise—

Ultimately, that's the currency that matters. Ultimately, it's just hard for folks to acknowledge that and understand it.

Emad.

I mean—

Well, yeah. It used to be that capital was what compounded, right? You give people money earlier, and that $1,000 becomes $20,000. Now it's compute and cognition that compounds.

As you move to self-learning systems, your capital almost becomes irrelevant, as the compute can get capital quicker than anyone. It's all about, again, how you build that whole cognition architecture around you. So I think this is great. And then the other thing is, we've got to give people access to frontier compute as young as possible as well, in a way that makes them able to compound the benefits from that.

I would just add that, to my eye, this looks like the beginning of universal basic equity. We've spoken on the pod about UBI, UBE, UBS. This looks like universal basic equity, where every person in the economy will have an equity stake in the economy. If we do see hypergrowth—macro hypergrowth—over the next few years, then $1,000 in a 530A account, which is again what this Invest America program is—thank you, Brad Gerstner, for helping to conceive this idea—what $1,000 in an account now in a 530A account a few years from now, if we experience hypergrowth, could be quite material to a person's living circumstances.

Yeah, Brad's brilliant, and he's agreed to come on the pod and talk about his moonshot. So we'll make that happen, probably in early 2026.

Along these lines, our next story here: College students flock to a new major, AI. Okay, not very new for us, but AI majors are exploding in popularity, with schools like MIT and UC San Diego launching AI-branded degrees. So, Dave, I'm going to go to you first. Everything is AI at MIT these days, right?

It sure is. Yeah. So, in MIT lingo, this is Course 6-4, which was only added just a minute ago, basically. It's already almost caught up to 6-3, which is core computer science, in terms of the people who are majoring in it. And I'll tell you, when you talk to the students, they say the curriculum sucks.

There are 2 or 3 great classes, because you're trying to build an entire major and there are only 2 or 3 classes so far. It takes the school too long to build the material because they're used to this much slower timescale. So I'm sure it'll fill in because the demand is so high. But as of right now, there's just a couple of classes and then a whole bunch of garbage, which is frustrating the heck out of the students, by the way.

Everybody wants to move to this, and for good reason. No matter what you're trying to achieve in life, whether it's biotechnology or space travel or whatever, the way to achieve it is via AI. So if you get a good grounding in AI, you're actually empowered to do virtually anything. It's the perfect thing to study anyway, especially when you're young and you have time on your hands and you can really grind through these complexities. So I'm really glad this is happening. I just want the curriculum to move much faster.

Catch up. The interesting note here to add is that AI-related job postings in the U.S. were up 50% year over year from 2024. So that's going to continue. Any other thoughts on this one? I mean, it feels kind of obvious, and I think the biggest challenge I've got is that it shouldn't just be in college. We should be seeing this in high school as well.

We're going to see a lot of people who are going to skip college. I think that's a debate that we've had. So if we can get you started in high school to think about AI, think about the world you're going to be inhabiting and inheriting, and how you use this technology to create your vision and your passion. Emad.

Well, I'll tell you, if there's one actionable thing: talking to every school administrator, every high school principal, every college administrator, approve the applications. When the students say, “I want to study this on my own. I don't want to study that,” just say yes. Just approve it. Let them carry themselves forward. Don't hold them back.

Intrinsic motivation. Intrinsic motivation.

Unleash them. Unleash them.

My favorite Joseph Campbell quote is, “Follow your bliss.”

Yeah, that's a good one, too.

Yeah.

Yeah. I think it's fascinating because the fundamentals of AI are not actually that hard. It's not easy math, but it's not that hard mathematics. My take is that if you have a semivocational course where you do fast.ai, which is a fantastic introduction to the math and the basics for programmers, Andrej Karpathy's video series on YouTube, and then you just vibe-code and build, and the entire class implements some of the latest research every month, that will put you way ahead of everyone. I think CVs and qualifications move to: Show me what you have built and done with AI.

Yeah. So if you're listening to what Emad just said and you're a student, take exactly what he said. Take the Karpathy material. Karpathy is the one guy from that original OpenAI crew who is not a self-made billionaire because he's building education for the world right now. He's given up. He could be a billionaire tomorrow if he just signed some documents. He probably is anyway, actually, from his own banana stand. But putting that aside, he's building out the best educational platform you could ever imagine. Just go find him online. Then tell your high school teacher or your college professor, “I want to study this instead.”

“Can you allow me to do that in replacement for this class I would have taken?” Brilliant.

That's the solution.

But what you just said is antithetical to the concept of a university structure—

Or a high school or a high school program. Yeah.

I think, just quickly, if you implement the stuff together and discuss it—

And again, that fits with it.

It's way better than doing it by yourself.

Mm-hmm.

For sure. But I'm just saying, you have to literally hijack a high school curriculum or university curriculum to do that because it's not offered today. To Dave's point earlier, Peter, Lily, my wife, has been pressured by all the local parents to have a day of just AI mind shift for all the teenagers. So we're going to do that and pilot that out and see how—

I heard that you've stolen Max Song, my strike-force member, to join.

Yeah. [laughter]

I should also point out, Peter, I mean, I just want to look at this for a minute from the perspective of economics. Right now, AI engineers are a complementary good or complementary service to AI compute. The cost of intelligence is going to zero, and so right now, pursuing careers and majors in AI is highly complementary.

But recursive self-improvement is also potentially imminent, and to the extent that recursive self-improvement gives us AI engineers soon, we might start to see AI itself become a substitute for AI engineering labor. In which case, maybe this rush to major in AI at MIT and UCSD may reverse itself, unwind itself, and everyone goes back to majoring in the humanities like they used to.

We had the conversation in the past about whether you should learn to code, and then along comes vibe coding. One of the conversations we had yesterday up in Redmond—we'll hear about it later this week—was the importance of studying philosophy.

All right, let's talk about the next job boom, which is in data centers, where the gold rush is for construction workers. The AI data center construction boom is making welders, electricians, and supervisors earn between $100,000 and $225,000. These AI companies need that kind of labor, and there's a national shortage of 450,000 skilled-trade workers.

It's significant. This is an alternate career path where you don't come out with hundreds of thousands of dollars in debt; you come out with the ability to earn immediately. How long will this opportunity last before Optimus 4 or 5, or Figure 6, comes in and does this work for us? I don't know. Maybe it's 5 years, 10 years—something in that realm. Thoughts?

I agree. I think that is the multitrillion-dollar elephant in the room. As with college majors flocking to AI, in this case, if you can pursue a career in the so-called skilled trades to facilitate tiling the Earth with compute, again, I think that's potentially a very promising local strategy.

But, of course, 5 to 10 years out—and I agree, Peter, with your timelines—we're going to see humanoid-robot substitution effects.

All right, let's jump in.

Yeah, go ahead.

The economics and dynamics of AI data centers are really similar to fracking, actually, when you think about it—even the financial structures and these booms in these industrial areas. So I think it'll last longer, but let's see.

I find this next article interesting: Amazon is expanding its network after talks with the USPS stalled. You may not know this, but the U.S. Postal Service is one of Amazon's main delivery carriers. The U.S. Postal Service is delivering Amazon packages the last mile in rural areas, and it's been a significant, roughly $6 billion-per-year contract between the two. That contract is breaking down.

My prediction is the U.S. Postal Service will be put out of its own misery, and Amazon will get a contract from the government. Today, the U.S. Postal Service has about an $80 billion-per-year operating budget, and it's losing $7–10 billion annually. Thoughts, comments, gents?

What happens when we have last-mile robotic delivery services? I think we have to prepare for that imminent future here, and that is probably best done by the private sector.

Drones. We saw an article probably about a month ago that Amazon is now giving its drivers augmented-reality glasses, right? It's saying to the drivers, "Okay, wear these glasses. They'll warn you if there's a dog in that apartment building or that house, show you where to drop the package, and so forth."

I think what's really going on here is that Amazon is collecting all of the last-mile, or last-100-meter, data and being able to train its future robots: autonomous trucks and autonomous robots doing that last 100 meters.

I keep wanting to say 100 feet. I hate the fact that we use feet and pounds in the United States. It really drives me up the wall, and it drives me up the wall that science-fiction writers are using that as well. Damn it. Did we switch over to metric back in the '60s? It's a pain in the ass.

Anyway, this is going to be an interesting battle. What's your over-under on how long the post office lasts? Anybody?

Well, this is an interesting bellwether because it should have been privatized probably 30 or 40 years ago. Everybody knows that. But it's written in the Constitution, and nobody wants to mess with the Constitution.

It's so obvious. Space travel—or space launches, at least—are getting privatized. It's not in the Constitution because space didn't exist when the Constitution was written, so it can just move over to SpaceX and Blue Origin.

By the way, look at the FedEx line. FedEx had such an amazing lead. Fred Smith was such an extraordinary entrepreneur, and it's been just slowly on a decline.

My guess is the U.S. Postal Service has, at most, 5 years left.

That's how I would have it, about the same.

Yeah, yeah.

It'll take an act of Congress.

Yeah, and two-thirds ratification by the states. It's a structural issue. This is a good case study. It's not that big a deal, but it's a great chance to learn: What are we going to do that's blatantly stupid because of legacy structure, and how is that going to get fixed?

All right, let's move on to space—a fun subject. There are 4 space stations under development in the U.S. today: VAST, which is going to be launched by SpaceX; Axiom Space; Starlab; and Blue Origin's Orbital Reef. I'm just throwing this out because it shows that, finally, we're going from government to truly commercial habitation.

Alex, do you want to add anything here?

Yeah, it's not a coincidence that there are 4 separate private space stations that are about to launch. These are actually all causally related to a NASA program. When we speak of privatizing the government, NASA in 2021 started the Commercial LEO Destinations program, with ultimately $1.5 billion in funding.

The SpaceX surge to space that we saw was in part the result of another analogous NASA program to try to commercialize—

Commercial Crew Program.

Yeah, that's correct.

In fact, the Commercial Crew Program was what saved SpaceX, right? SpaceX had 3 launch failures of its Falcon 1. It finally got the fourth one to orbit after Elon literally borrowed money to be able to put that together. At Christmas in 2008, he won a billion-dollar-plus contract from NASA to go forward with Falcon 9, which is today the most successful launch vehicle on the planet by an order of magnitude.

Yeah, that's right. The Commercial LEO Destinations program was spun up in part because the International Space Station is going to need to be de-orbited sometime soon, and there was a desire for a private U.S. space presence to succeed the ISS.

I'm very optimistic about all of these and other private space stations. I think we're going to see an exponential rise of humans in low Earth orbit sometime.

You know what I'm excited about as well, Alex? Jared Isaacman. I cannot wait.

Jared Isaacman is back on the docket to be our NASA administrator. I'm not sure when the congressional hearings are finalized. Do you know?

Several days ago.

Oh, is he in finally?

Well, there has to be a vote, but the hearing was several days ago.

Okay. I've been texting with Jared, and he's agreed to come on the pod as soon as the confirmation is done. I'm so excited about that. He is brilliant—absolutely brilliant. I've known him for a long time. I took him to Russia to watch the Soyuz flights from some of our commercial launches there.

Wait, I have a quick comment. This space-station thing reminds me of a date. I remember one of our NASA astronauts telling us that the most interesting date in the world for him was October 31, 2000. It was on that date that the first human being lifted off to the International Space Station.

Mm-hmm.

Since that date, we've always had at least 1 human being off-planet. The first molecules are kind of drifting off this thing.

So this next article is a bit of a surprise: SpaceX is considering a 2026 IPO. I've had this conversation with Elon. He was always resistant to taking SpaceX public for a number of reasons. When you're a public company, you have to disclose all the details. He doesn't want to disclose all of his details and how he operates to his competition.

But the other thing in particular was, if you're a public company and you're spending a whole bunch of money to build Mars vehicles to go and colonize the Martian surface, is that something your shareholders are going to support?

Listen, I'm a SpaceX investor. I've held it from the very beginning. I would love to see it public. I always thought that what Elon was going to do was spin out Starlink and take that public, and keep the launch capability.

That was the conventional wisdom.

Yeah.

I think—oh, go ahead. Sorry.

Yeah.

No, I mean, Elon’s a smart guy, and he’s got a million GPUs, so they’ll have more AI lawyers than anyone to attack the stupid people who come after them. But, seriously, this is like—

SpaceX, xAI, and Tesla will all basically have full AI teams, top to bottom. You can criticize their strategy, and they’ll all just clap back at you. You can sue them for the silly stuff, and the AI will just clap back. It’s a big difference in the way that you can actually run public companies.

Yeah. You take SpaceX public and you take X public, and Elon leaps over the trillion-dollar mark in terms of personal net worth.

I’m also, just to comment quickly, not sure about the historic story that Starlink would spin out and do its own IPO. I think with the rise of orbital data centers, that muddies the water somewhat in terms of Starlink as a pure communications service versus Starlink as a predecessor to orbital data centers, putting compute up there and not just communications capabilities. In that sense, I could imagine a scenario where orbital data centers are actually pulling all of SpaceX to go public, not just spinning off Starlink.

Yeah, and that’s a relatively new part of the conversation.

That’s right. It’s very recent.

I love this competition. It’s fun. My mission has always been to open up space, and I’ve built so many companies around the space theme. It makes the 9-year-old in me so proud and gives me such contentment that 2 of the wealthiest humans on the planet are battling it out to open the space frontier.

Blue Origin plans to start flying cargo to the moon in early 2026 using its New Glenn heavy-lift rocket, which recently did a launch and full recovery of its first stage. It’s backed by a multibillion-dollar NASA contract for targeted human landings in 2028. The government has always wanted dual suppliers. For most of the American spaceflight industry in the 1970s, 1980s, and 1990s, it was Boeing and Lockheed Martin competing for this.

Here comes SpaceX, which becomes the dominant player, and now the government wants the number 2. It looks like it’s going to be Blue Origin, which is super exciting. Alexander, thoughts on this?

This was basically the plot of Season 3 of the television show For All Mankind.

I love that show. Yeah.

It’s a wonderful, wonderful show. The 3-way race—in the case of Season 3, it was a 3-way race to Mars. In this case, it’s a 3-way race between SpaceX, Blue Origin, and China to land humans again on the surface of Mars by 2028 or earlier.

I think this resumption of a space race, which was dormant for 50-plus years and maybe also had collateral downsides for the rest of the economy in terms of overall innovation, is coming back to life. We’re back in the space race again, and one might hope we’ll see a lot of growth and innovation come out of it.

The 9-year-old in me is so happy. Just to put some numbers and size against this, NASA’s 2025 Artemis budget—Artemis is its lunar program, its human lunar program—is $7.8 billion. Let’s look at that compared to the Apollo program.

In 1966, NASA’s budget was about $5.9 billion, and the Apollo budget was about $3 billion of that. Apollo was half of NASA’s budget. If you adjusted the Apollo budget to today’s dollars, it would be about $35 billion to $40 billion. Compare that to the $7.8 billion we’re spending on Artemis.

We were spending about 0.5% of U.S. GDP on the Apollo program back in the 1960s. Pretty impressive. The reason we don’t have to do that anymore, of course, is commercialization and technology. We brought the price down orders of magnitude.

I find this article hilarious: “Sam Altman Enters Rocket Business to Compete with Elon and SpaceX.” What’s fascinating is that Elon goes for BCI and Sam goes for BCI. Elon goes for space and Sam goes for space. There are probably a few other areas. Any particular thoughts on this one, gentlemen?

Well, I'm really curious to see how the code red interacts with this. Sam has cut a deal in every single dimension related to AI, including Johnny Ive with the wearable device, Stargate with the data center, and Broadcom with the chips, the TPU.

Well, I'm really curious to see how the code red interacts with this. Sam has cut a deal in every single dimension related to AI, including Johnny Ive with the wearable device, Stargate with the data center, and Broadcom with the chips, the TPU.

That’s Sam personally versus OpenAI, right?

It’s a mix. It’s a mix. The bigger deals are with OpenAI, and then there are about 300 or 400 personal deals that are all the use cases and components. It’s a massive matrix in this Sam-verse, a massive network of connected parts.

But now you’ve got this Code Red: “Hey, wait. The thing that matters at the middle of it is these AI benchmarks, and we’re now off the chart on Polymarket.”

Code Red, Code Red. I’d be very curious to see what that means because he has a lot of talent, but there’s still a limited supply. It’s not infinite. All of that gets drawn back into the middle. That’s going to cut some of the things on the edges.

To give some more detail here, the company he’s in discussions with is a company called Stoke Space, founded by 2 former Blue Origin propulsion engineers. As with every launch company, it needs to be fully reusable. It’s a 2-stage, fully reusable rocket. It’s never flown, right? They’re using something called a ring-shaped aerospike engine, which also has never flown, so it’s a little bit of a risky bet.

If Sam wants to enter orbital data center capabilities, I think having space launch capability—and, of course, let’s not forget Eric Schmidt, who is also in the rocket business—this vertical integration by hyperscalers into space is probably an inevitability at this point. We’re certainly not going to get our Dyson swarms without that.

Yeah, I think it’s fascinating. Just like we’re getting hyperscaler fusion plants.

That’s right.

Yeah.

Facebook space station.

No, I’m not going there.

You’ve got to be full-stack to control the cone of humanity.

Exactly. Love that.

It’s a better place to put a name than a sports stadium.

Maybe Instagram space station. I’m not sure.

Oh, God. You guys are makers here.

All right. Orbital compute energy will be cheaper than on Earth by 2030. Again, I still find this challenging, just because we have so much solar flux on Earth and don’t have to worry about launch. But if we can really get the cost of launch down to $100 per kilogram, which is the projection with Starship, versus perhaps $500 to $1,000 per kilogram, who wants to jump in on this one?

I’ll just comment: It could also be even cheaper than this once we get compact fusion online. A lot of these orbital compute projections are assuming that solar is the primary power source for orbital compute. It doesn’t have to be. Once compact fusion is cheap enough—and there’s no reason to expect it won’t be—we can tile low Earth orbit, at minimum, with compute as well. It won’t require all of these expensive solar panels.

So we’re going to throw fusion reactors into orbit?

Yes.

Wait. So the theory there is that launching a fusion reactor is cheaper than just a solar panel? The solar panel in space is about 10 times more effective than it is here on Earth, but it’s still cheaper to launch the reactor?

It’s maintenance.

I think that—

I just won the bingo game. We said “launching a fusion reactor into space.”

No, solar in space is still fusion-based. It’s just using the fusion reactor at the center of our solar system. I think the question is, where do we want fusion power to be located for space-based compute?

There have been a number of deep-space probes that NASA and other organizations have launched that are using fission, for example, in ion-based propulsion. It’s not like nuclear energy is that foreign to space. Fission thermal has been used by deep-space probes for decades. It’s not like we don’t know how to do it.

What’s going to be new is compact fusion in particular. We’ve put fission-based energy in space for decades.

You know, Alexander, I’m looking at the numbers here, and it says the current terrestrial average is $12 per watt, and we’re talking about $6 to $9 per watt in space. That’s not enough of a difference for the level of complexity.

Yeah, you’re going to need at least a 10× drop in that.

Yeah. So maybe it is compact fusion. If we’re actually mining the moon—let alone disassembling it—

We already won the drinking game.

Yeah. Well, hey, maybe that occurs. But I love the fact that it’s now orbital data centers that are driving humanity’s expansion into space. That’s amazing. It was going to have to be something.

If you look at all the science-fiction plots, it was either going to be a discovery. Again, in For All Mankind, without spoiling too much, it was either going to be ice on the moon or the discovery of microbial life on Mars, or something like that, that had to motivate space exploration and development.

Who knew that it was going to be data centers? Well, it had to be something.

It had to be. I remember when I was at MIT and running SEDS, I put together this brochure on why we should open the space frontier. I used to have to rationalize things like better materials there. There was always some very soft rationalization, but this is real industry. Real people were trying to figure out what we could manufacture in space that would have value here on Earth.

This is the moment you've been waiting for in undergrad. You should be like a kid in a candy store. But it totally makes sense, and it always drove me nuts when Roman Septa[?], our buddy, graduated and went to Ford, where he was working on wire harnesses and rearview-mirror motors.

I'm thinking, why do we need another electric component in a Ford? Why don't you work on space or something foundational that changes humanity? It's because a new feature for this massive installed base is economically incredibly valuable, even though it's marginal for society. So it sucks up way too much great talent, and something really important like space data centers doesn't get worked on. But you need an economic crack that starts the whole process, and this is it. We finally have it, after all this time, and it's so much better of a storyline than For All Mankind.

Actually, it's a lot later. I'm still betting on asteroid mining. Everything we hold of value on Earth—metals, minerals, energy, real estate—exists in infinite quantities in space. Those nickel-iron asteroids are worth trillions of dollars in platinum-group metals, and those carbonaceous chondrites are what we're going to mine for oxygen and hydrogen for fuel.

Can I burst your bubble there, Peter? I think AI transformation of materials science will ride around all the scarcities around that. I don't know.

Or maybe it'll accelerate it. I don't know.

Peter, when you were founding SEDS, did you foresee the plot twist that the killer app for space would actually be getting enough compute available to do generative cat videos doing funny things?

I was not able to project that far ahead, to be honest.

Yeah. So who knows what the asteroid belt will actually end up being used for?

We'll assemble it for compute, obviously.

Lest we leave the Chinese out of this, a company called CosmoSpace is planning to build AI data centers in space. They're putting up a supercomputing cluster with 3 modules. One has 100-megawatt-level power, the other has 10 terabits per second of communications, and the third is a 10-exaflops, 10^18-operations-per-second compute module. Alex, what do you think about this?

I think we're seeing a race to build Dyson swarms. It's as simple as that. It's not just a race to the Moon, and it's not just a race to tile the Earth. It is a race to put as much AI-accelerated compute into low Earth orbit as possible.

CosmoSpace emerged from nowhere. I had never heard of them several months ago, and I don't think an otherwise obscure Chinese provider is going to be the last story we hear about Chinese orbital AI compute. We're probably going to see a dozen different vendors from China. We'll see, as just discussed, a dozen hyperscalers in the West, and the concern maybe becomes overpopulation on Mars, making sure that all of these Dyson swarms remain interoperable.

Can we just point out to all of our listeners that, if you've been following us on Moonshots, this conversation about orbital data centers did not exist 4 months ago in any way, shape, or form? It was there, and I'm sure people were speaking about it, but it has now become a weekly conversation over the last 3 months. It literally came onto the scene with a vengeance. It's extraordinary.

Well, that technology clearly works. It's proven technology now. So it's just a question of launch costs. That's the only missing link, and it looks promising.

Even if they figure out the dissipation side of things, I think that was still outstanding.

Yeah. No, no, they've got it. I mean, there is still an issue of how you can get efficient energy dissipation on the back end. In fact, that was proposed as an XPRIZE this year at Visioneering.

They'll get efficient.

It'll make more sense.

The thing that always gets me with this is that it's horribly insecure.

You had the proposal by Eric[?] and others about how things go in space and what that means for data centers. Space data centers are saying they'll go up, and they'll just start disappearing, honestly.

Yeah. I'll give you the counterargument, and I totally agree, by the way. The hardware depreciates in 3 years anyway, so you only need it to be secure for 3 years. I think the counterargument is that the US Space Force will basically guarantee enough safety that you can get 3 years of hardware out of it before something bad happens, and that's all you need to pay it off.

So what about solar flares? One of our subscribers asked that question. What happens when you have solar flares hitting these, and what happens when there's an EMP that hits them as well? All of this gets knocked out, right?

Actually, I've got a very interesting thing. We were training on thousands of A100s a few years back, and we kept getting errors exactly at the time of solar activity because it was basically messing with the ECC memory.

Wow.

Yeah.

We'll find out how these things sustain in the back end.

Yeah, you know this better than anyone, but a little inside scoop on the million-GPU clusters that are being built now: they have errors here on Earth, too. You have to solve that problem in order to have coherent training anyway.

The error rate goes up a lot in space, but you have to have a process for backing off. Right now, everybody does checkpoints and rollbacks, but you can't invest an hour of a million GPUs at millions and millions of dollars and say, "Oh, wait. We have an error. We're going to roll back all of those GPUs for an hour."

You have to do it unit by unit, and that work is well underway. Presumably, that'll work fine in space, too, and you can tolerate the error rate.

The other comment, of course, is that those models just want to learn. For AI compute workloads, to the extent we're doing training or inference, they can be structured to be fault-tolerant.

Yeah, Microsoft was actually the leader in that, and now Google's the leader in that. It's seamless, the way that it flips. It'll be interesting once we get out in space.

Our last topic here: we're going to dive into robotics for just a few stories. Here we are. After an AI push, the Trump administration is now looking to robots. Robotics are the focus of a 2026 executive order to accelerate US robotic development.

We're seeing this in China, right? China is crowning its winners. It's got huge investments in the robotics industry. In fact, in our last pod, we talked about the fact that there is, at least according to what the Chinese are calling it, a robotics bubble going on today, with over 150 Chinese robot companies.

I find this one fascinating. As the acceleration continues, major national robot strategies are coming online. Robotic firms are likely to have tax credits, subsidies, and protection against trade measures, something along the lines of the CHIPS Act. Once again, thoughts, gentlemen?

I'll comment that the zeitgeist at NeurIPS this year was that humanoid robotics is the next big thing for AI after agents. I think this is how reindustrialization of the US and the West happens. I think this is probably the best path for radically increasing economic growth and automating the two-thirds of the services sector that relies on physical intervention. This is instrumentally convergent for the future that we want.

Just to remind folks from our last pod, we talked about the fact that China installed 54% of the world's total robots last year. Again, massive, massive push.

I want to show a couple of quick videos here just for fun to close us out. We saw, in the last week, a little bit of an Optimus-versus-Figure competition. Elon posted this image of Optimus walking. Let's take a look. Here it comes, running along, jogging. Then we had Brett posting this one of a Figure running across.

I have to say, they look pretty natural compared to where they were 6 months ago. Which one did you like better? Let me play this again. Here comes Optimus. Optimus coming along. I don't know. It kind of looks like Figure is doing a better job running to me. What do you guys think?

I think they're both incredible. I'd also throw out a request to the audience for the show: if you're interested in supporting robot athletics in the United States, either as a host, as a vendor, or in some other capacity, please reach out to me. I'd like to do what I can to ensure US dominance, and Western dominance in general, with humanoid robots via robot athletics.

Yeah. Well, let's take a look at Chinese dominance with this video, and then we can talk about it. We saw last week the T800 humanoid robot from EngineAI in China. This is 5 feet 8 inches tall, with incredible capabilities.

They put out a new video that I wanted us to take a look at here because it's a little bit shocking.

Are we 100% sure this is real, by the way?

Yeah, they claim it's real, and this is a follow-on. So here we are with this T800 robot basically kickboxing. But check this out when it goes up against a human opponent. Wow, kind of scary. I'll go back to my standard comment that having a robot doing kickboxing is not a great marketing message.

Yeah, I think calling it a T800 is also not a great marketing message.

What about all the skulls that they put in there?

Can I do a little headline?

Yeah, we love your rant. Go for it.

Look, a human being has evolved for 4 billion years, which is an optimized strategy. 200,000 years as a human being—for survival, we have the human structure to survive and be able to quickly pick fruits off trees, so we have opposable thumbs and whatever. I mean, a wheel is so much more energy-efficient than walking; it's ridiculous.

For God's sake, at least put little wheels in the bottom of the robot, like the kids with wheels in their sneakers, so they can be more efficient, because battery power seems to be a huge limiting factor in this. So why don't we have a wheel along with the leg so they can do both when it's needed? Just having robots copy human beings seems to be the stupidest thing in the world.

It feels to me like when we first had TV, with radio announcers and television reading the same scripts. Podcast: we're going to be doing a podcast from Figure's headquarters in Palo Alto in January. You're not invited.

Yeah, I don't know. I think this is really interesting. Box the robot.

Yeah.

This is really interesting. The entire chest cavity of the robot is actually a battery here. But the 450 N·m maximum joint torque—that's the really interesting part. Basically, this thing can punch harder than a gorilla, like 4 times Mike Tyson. Do you really want those walking around in the streets? Are they going to have regulations on the maximum joint torque?

Also, if you actually look at the full video, which is real, they even show a behind-the-scenes one. And if you look at the previous video, there's something called sim-to-real, which can basically model human actions in a robot. So we have full—almost Real Steel, the movie with Hugh Jackman—teleoperation capabilities now in robots, and soon it'll be policy learning.

This robot can do freaking kung fu.

With UFC. Punch through.

UFC is coming. We're going to see Tesla Bot—basically, Optimus versus T800. I mean, it's going to be Olympic-level sports. It's going to be amazing.

Olympics versus UFC. I think that'll be exciting. But we have to actually ask: do we want to have regulations around the maximum joint torque of humanoids in the street? I'm fine with these being in the fighting arena. I think it's fantastic, or in an industry, but I don't really feel comfortable with them walking around.

The challenge comes—

No, no, the kitchen. You're going to be in the kitchen.

The challenge comes when they enter warfare, right? I mean, this is the Terminator in the pure sense: robots on the battlefield. It's a scary direction for us to take humanity.

Yeah, I think it's all of the above. It'll be warfare. It'll be in the kitchen. It'll be on the street. And you'll see governance and governments at different levels, whether it's municipal, national, or international, regulating the parameters of what the rules of engagement are, from making an omelet to fighting a war.

All right. I want to thank CJ Truheart, who gave us our first song on the Moonshot Mates. This is an outro piece called “The Exponential.” But before I play our outro piece, gentlemen, it's been a blast to spend time with you guys again. I love this. Emad, it was wonderful to have you as a fifth here today. I'm grateful for you. What's your week ahead look like, Emad?

Lots of policy work and more agent stuff. We have lots of releases coming. Exciting times.

And how was Japan? You were there for FII Japan.

Fantastic. Huge amounts of corporate and government interest in using AI to help accelerate the way forward. Hopefully some announcements about that soon.

And Dave and AWG, you're about to hop your flights back to Boston, I gather.

I think collectively we covered 12 countries in the last week and a half in this group.

So it'll be nice to be home for at least a week.

Nice. And same for you, Salim—a chance to stay home?

Yes, I'm here for a bit. I just got back. So I'm going to be preparing for the big online Meaning of Life session, where, if people are interested, come armed with any question you have about life, and let's frame any question.

When is it, Salim?

It's December 17 at 11:00 a.m. Eastern. We'll go for several hours on metaphysics and philosophy.

And when Salim says “go for several hours,” he means 6 to 8 hours.

Well, it's a big topic. There's lots to cover.

This is an example. We start off with a conversation about what truth is and have it broken down to a 2-by-2 framework. This is sensemaking. It allows us to have a decent conversation: What do we mean by that?

Well, on Monday, I'm heading up to the Buck Institute in the Bay Area to talk about longevity and AI—my favorite one-two punch. And with that, let's listen to the music of CJ Truheart as we wrap this episode. And gentlemen, see you on the next episode of Moonshots. Thank you to all our subscribers. If you haven't subscribed yet, please do. We're now putting out more than one episode a week just because the speed is moving so rapidly. So, if you want to know when the episodes drop, a quick hit subscribe and let's listen to CJ. All right. If you're a music producer using AI and you want to give us an outro, just go ahead and let us know. And please, next time you're watching this and you have questions, please post them in the comments. We are going to do more AMA in the next couple of sessions. Gentlemen, moonshot mates Dave, AWG, Mr. Exo, Emad. Thank you guys. Having a fantastic week.