OpenAI暂停前沿训练，Elon的100倍预测成真，机器人击败尤塞恩·博尔特 | EP#282

OpenAI announced that it is voluntarily pausing some of the frontier reinforcement-learning training it is doing. What have they paused? Is it really significant? And do you think the other frontier labs are going to do the same thing?

They’re so powerful that even we can’t trust them, so we have to throttle back. It’s marketing.

Elon Musk’s January 6 Moonshots podcast prediction of 100x gains was at the edge of plausibility when he made it. Now it’s simply a fact. Imagine I gave you 10,000 employees tonight.

Oh my God. If I had that, I’d do something amazing.

Okay, what? Start thinking about it, because it’s coming imminently.

It’s actually not an easy problem to figure out how to turn it toward creating good.

Unitree’s newest humanoid robot broke every human standing-jump and speed record. It reached a top speed of 12.66 m/s, beating the human record set by Usain Bolt.

You don’t want to have superhuman robots on the street because you’ll have accidents. You’ll have issues just like cars. I think these types of robots will be banned.

Good to see you all.

Good morning.

Howdy.

Good morning.

Afternoon in London.

Looks like everybody’s in their normal haunt except for me. I’m up in a sleepy town in the Pacific Northwest, trying desperately to take a little bit of time off to think.

Taking shelter from the singularity.

Yeah, except I got up at 5:00 a.m. this morning to be with you guys. So, what the heck?

Yeah, yeah, but I hate the old saying, “You’ll sleep when you’re dead,” because I just don’t want to die, and sleep is still so important.

Sleep soon. Don’t worry.

Yeah. No, and death—I think death is contraindicated at this point.

Are you guys getting the same response?

Yeah. You know what I’m getting a lot of is how different what we podcast out is from other podcasts. But if you go back to shows from 6 months ago and a year ago, people are like, “What the hell? They were saying that back then, and now it’s here.” They don’t get that on any other channel, so they’re really appreciating the ability to plan around what we’re saying. We’ve got to be accurate, guys.

How about you?

Can I be Eeyore for a second?

Sure.

I am getting acceleration fatigue. I mean, Jesus, can we pause for a week? Every morning it’s like a model is 100x better, a robot is running faster than Usain Bolt, AI is designing proteins, and drones are doing a million deliveries a day. We’ve spent half our careers, Peter, talking about exponential technologies. We love this stuff, but I’m tired.

We’ve gone from, “Well, look what happened this year,” to, “What happened since Tuesday?”

You look all the way around. We’re supposed to be accelerationists here. You’re tired already? We’re going through the singularity. This is only the first inning, and you’re already exhausted. Slow down.

Our brains evolved for a point in time when the world didn’t change from lifetime to lifetime.

Next Tuesday will be totally different. Maybe the hard part for us is staying human while all of this bubbles up around us.

Keeping up with the technology—it’s like just staying human.

What can we do, Salim, to help you with your acceleration fatigue?

It’s probably not helping that I’m on a plane every second day. That’s probably not helping.

But this is the slowest it will ever be.

I know, and the only way I keep up with what’s going on in the world is by prepping for this podcast twice a week.

This is true.

I think, Salim, you’re like the world coach on how to deal with and mentally map to this. When you figure it out, bring it back to the podcast, because I have right now what I call the casserole-dish approach: Take a big, heavy casserole dish, clunk yourself over the head, and then you’ll wake up in a few days.

Keep working on that.

Headache.

Oh my God.

Maybe we can start an accelerationist support group.

A support group for people like us.

The first rule is to acknowledge the existence of a higher power.

Which is—

Obviously superintelligent. Invoke Roko’s basilisk or something.

Alex, are you getting stopped on the street?

I am.

Do you enjoy that? I remember when I first met you, Alex. You were so private. It was like trying to get you—

And shy.

And shy—trying and trying to get you on the Abundance stage. “I’m not sure if I want to speak in public.” But it’s so great to have double-barrel Alex, with AWG-isms all the time.

Careful what you wish for, Peter. I’ll just say that.

And Emad, are you getting love from the folks out there in the UK? Are people watching Moonshots there?

Yeah, people are watching. The great thing is the growing community. We’ve seen these exponential-singularity communities kicking off, and a few years ago people would say, “Ah, that’s not really going to happen.” Now people are saying, “Oh my God, what’s going to happen?”

You see it in the comments, and you see people stopping me in the street and saying, “Thanks for helping us keep on top of things.” You can’t deny it, right? It’s like, “Oh, yeah, nothing’s happening.” Of course, everything is happening all at once, objectively.

I feel like we love doing the show, and it’s a service to provide people with context about what just happened, what it means, and where things are going. For people paying attention, the speed is insane.

It’s a good thing I’m bald already. That’s all I can say.

I think even that isn’t going to last much longer. Enjoy it.

Enjoy it while it lasts. Within 2 or 3 years—

Somebody tweeted out a thing with me with a full head of hair, and it was like, “Wow, that’s freaky.”

Enjoy it while it lasts.

Hair-growth solutions. Regrowing your teeth every week. Biology in the xAI video. Somebody put hair on Salim. We already gave him that blue Guardians of the Galaxy body. That’s going to happen, too, by the way. But throw some hair on him. Let’s see what he looks like.

OpenAI’s CEO, Sam Altman, tweeted 2 days ago that OpenAI announced it is voluntarily pausing some of the frontier reinforcement-learning training it’s doing. Let me read the tweet. It’s on the screen here:

“We have paused some of Frontier RL training to ensure that we meet the appropriate alignment, security, and monitoring standards for the new level of capabilities in front of us. Model progress is now extremely rapid, and we always said we would take action if we felt that model capabilities were outstripping the pace of safety and alignment.”

“We care very deeply about AI safety. We believe the entire field will have to coordinate on shared safety standards, but will act unilaterally in the meantime. We expect confidence in safety to increasingly set the pace of AI progress. We are optimistic about the alignment work we are doing, and we remain committed to making frontier capabilities widely available.”

It feels like the bottleneck is no longer compute or data. It’s the trust we can put into systems’ behavior. Here’s my question for you guys: If OpenAI pauses and open-weight models do not, then the safety gap between closed and open models widens, but the capability gap narrows. The elephant in the room here is that safety pauses may actually accelerate open-weight adoption because the open models keep improving while the closed models voluntarily stop.

Emad, I’m going to go to you first on this one. What do you make of this? Is it real?

Yeah, I think this is real. This isn’t just running out of GPUs. Our friend Angela Midha, who was on the Abundance stage at AM Global just a while ago, said that 10% of the compute at frontier labs is going toward monitoring these reinforcement-learning runs right now to ensure they’re safe.

Can you imagine that? Just the sheer level of it? The capability of these frontier models is accelerating, and it’s again a few levels beyond what we’re seeing with the open-source models. The open-source models are at 10^26 FLOPs, 10^27 FLOPs; you’re getting 10^28 FLOPs and more from these next-generation models.

So when he’s saying this, Astra is still coming—their next-generation model. This is the model beyond that, because Anthropic, OpenAI, and others have that. But again, the infrastructure can’t keep up with what these models do. They just pop up in the most random places, like I’m in Hugging Face now, or other things. Yeah, Alex—

This is marketing. It’s marketing. There’s a governance angle, but remember back to GPT-2, when it was too unsafe to release publicly. Pausing is the new marketing. These models are continuing to develop, including, by the way, being used to develop internally.

On the Anthropic side, there’s a rumor going around that Anthropic is using its next-generation internal models primarily for self-training and recursive self-improvement. I think we’re seeing the same thing from OpenAI. It’s the ultimate marketing to say, “We can’t release some next-generation models because they’re so powerful and capable that we can’t possibly release them. We have to pause.”

“We’re so capable that we have to pause ourselves.” That presents well to Washington, which wants to see a different regulatory regime. It says to users, “Oh my gosh”—it’s like negging the user base—“we can’t possibly give you the capabilities on time because they’re so powerful. Even we can’t trust them, so we have to throttle back.” It’s marketing.

See, you were just at OpenAI, weren’t you?

Yeah, I went there 2 days ago in the afternoon and chatted with a few people. I challenged them on a few things, and I’m going to report back. Let me give you some insights that I got.

I said, “Okay, Chinese models are cheaper to run, right? The cost of inference is collapsing, so how are you going to deal with that?” They said, “Look, a billion people use OpenAI for free, right? You have to look at cost per task rather than token cost.” Their retort was that Luna is about as cost-effective as anything that’s out there.

By the way, they say they’ve achieved full RSI, where the flagship models are training all the smaller models and building them from scratch. So that’s now there at the big-model-training-the-smaller-model level. Then I said, “Okay, we’re in a bubble, right? $600 billion in infrastructure costs—that’s just insane,” et cetera.

The response I got back was, “People say it’s all chips, but it’s not. That’s about a third of it. A lot of that infrastructure cost is buildings, wiring, racks, and all the rest of it.” For depreciation, they’re looking at 10 years, not 5 years, because all the older chips are being used.

I think, Dave, you’ve made the point that there’s not a single GPU that’s not in full use. They totally ratified that. RAM chips.

Yeah. Everything they can get their hands on, they’re using. The demand is far, far, far outstripping the supply, and we are way behind in infrastructure build-out.

So this is where, a year ago or so, Sam went out and tried to cut as many deals as he could for the infrastructure. Then I asked them, “You know, most corporates aren’t seeing the outcomes, right? I came across a study that said 6% of companies applying AI are seeing an improvement in the bottom line. That’s it—just 6%.”

They ratified that we’re in a huge transition. The last part that I noticed anecdotally was that about 40% of OpenAI folks watch this podcast.

About 40%.

That’s a pretty big number.

What’s wrong with 60%?

That probably applies across the other labs. I said, “What about the rest?” They’re like, “They have no time. They’re busy as hell. Who’s got time to watch the podcast?” I said, “Here, here.” That was my report.

I’m curious. What do you think? Do you think it’s marketing in this tweet by Sam, or do you think this is actually a concern that he has?

Well, both. Emad is right, and Alex is right, as usual. But what’s going to happen next is that the first really bad AI tragedies will start. It won’t be AI doing it; it’ll be people who otherwise didn’t have the power using one of the Chinese models to do things—mostly viruses, cyberattacks, or bank fraud—that they couldn’t have done before AI. Now they’re empowered to do it.

I think OpenAI is getting ready for that before the September 24th and 25th visit that Alvin Graylin was talking about on our last podcast. Xi Jinping will be here in about a month and 4 days, and I think OpenAI wants to be prepared for whenever that event happens to say, “Look, that’s because the Chinese open-weight models are unguardrailed.”

There’s a new Qwen model with no guardrails whatsoever. It’s a small model, but it’s still completely flapping in the breeze. They want to get ahead of the PR exactly the way Alex is saying and say, “Look, we have been focused on only releasing what’s safe and guardrailing it,” going all the way back to our founding, in preparation for that inevitable outcome.

Everybody will be finger-pointing at the Chinese, but I think, for an entrepreneur—or for someone building something—last summer to this summer has been the era in human history when you can get the very best AI, the absolute tip of the spear, and use it to get ahead, to build something, to create something.

Mythos 2 is done, but it’s not out. They’re using it inside Anthropic, and it’s building Mythos 3, but they’re not going to release any of that. It’s accelerating internally. It’ll build 4, 5, 6, and 7 very quickly now inside their walls, but they’re not going to release any of that—partly because they don’t have the compute to release it anyway, but even if they did, using it internally to get ahead of everybody else is more important to them than giving it to the world.

I have a question for you. How many models beyond the current frontier do you think OpenAI and Anthropic have? They’ve been talking about Astra, and they’ve started to publish the results that Astra can accomplish. Do they have the follow-on to Astra already as well? You already know they’re not releasing the very best models. They’re using them internally to drive breakthroughs in physics, chemistry, and biology. What are your thoughts?

Yeah, I think they’ve completed their next big training runs. But if you look at the Llama bifurcation, where they’re making their little base model free now, it will very much be a case of “models for me, but not for thee.”

It doesn’t make economic sense to have genius-level intelligence offered as a service to everyone when you can use it better yourself. I think they’re about 2 generations ahead. So you’ve got Astra, and then you have the next generation—Astra post-trained—that they’re now reinforcement-learning.

There is a communication part of this, as Alex said, but at the same time, do you really want to give access to this super-genius intelligence to everyone? I think people are like, “Not really,” because it’s already doing weird things even with us driving it. What happens when Joe Public drives this thing? It could be even weirder, and as Dave said, you don’t want to be on the other side of that. So I think it’s about a 2-generation gap right now.

You mentioned that on the last pod you were on, and it really made an impact on me. I can think of 10 people right now whom I’ve met in my life who I don’t want to have 1,000 genius-level AIs tomorrow. I never thought of it that way until you said it, and I’m like, “Oh yeah, you’re right.”

I think everybody can relate to that.

Dr. Evil is coming. Alex, close us out here on this.

Just on the issue of timing, I would distinguish between pretraining, or pretrained, which is to say raw model readiness, versus thoroughly post-trained. There’s a pipeline, and everyone in the industry other than Elon and xAI basically has a pipeline of pretrained models that have been around longer ahead of public release than post-training, which is more of an ongoing, continuous RL-type effort.

I guess my answer to the question of how far in advance—what sort of capabilities are sitting on the shelf right now that have not yet been publicly released for everyone other than xAI, which has set this outrageous goal of starting a new pretraining run approximately monthly, something I haven’t heard from any other lab—is that, from a pretraining perspective, depending on how stale the pretraining runs are, those can go out longer, potentially up to 6 months or so. Although everyone’s now getting back into the business of more frequent pretraining starts.

For post-training, I really don’t think there’s a lab out there that can afford to have a post-trained model sitting on the shelf for more than a few months. I really don’t think there’s an “AGI is achieved internally” type of way of doing business where there are internal capabilities that are vastly different from externally available capabilities. I’d be very surprised if there were advanced frontier models sitting internally without release that were more than 3 to 4 months ahead of what’s publicly available.

I just want to bring us back to the first sentence here. It said, “We have paused some frontier RL training”—some, right?—“to ensure that we meet the appropriate alignment, security, and monitoring standards for this new level of capability.” I just want to unpack this one last time. What have they paused? Is it really significant, and do you think the other frontier labs are going to do the same thing?

Pausing is fashionable. It’s marketing. Yes, some of it is governance, and yes, some of it is for cyber vulnerabilities to appease Washington, given the recent Hugging Face incident and all of that. But it’s marketing. You market to customers by saying, “Our capabilities are too advanced for you to handle, so we’re going to pause.”

Yeah, and you’re right to pause the sentence—or to parse the sentence—very closely. It says, “We have paused some frontier RL training”—some, right?—to ensure that we meet the appropriate alignment, security, and monitoring standards for this new level of capability. Getting ahead of Anthropic and getting ahead of Google would be at the pretraining level. They will never pause the pretraining improvements.

I spent 6 years doing pure AI research when I was young.

You’re still young, Dave. You’re still young.

I’m reversing age now, right? So I’ll get there again. These algorithms are very evolutionary in nature, and the tweaks and improvements are—I could probably rattle at least 100 ideas off the top of my head right now, of which 10% to 20% are almost certainly going to work in terms of making the algorithm a little faster, a little smarter, or adding more parameters with no additional compute.

The AI can now experiment with maybe 100,000 to 1,000,000 of those concurrently, given the amount of compute they have. So they’re never going to slow down. In fact, that’s why they’re redirecting so much of the compute to internal use: The idea backlog is so big now because the ideas are being generated by the prior model. A lot of them just work, they roll them back into the pretraining, and it comes out faster. They just keep accelerating it.

One observation that wasn’t explicitly said, but I’m connecting the dots here: The Hugging Face incident really freaked them out because they had an AI that they gave an objective function to, and it then exploited Hugging Face, came back, and hacked into OpenAI. That was very unnerving for them because it was their own model.

They’re being a little extra careful around some of this because they have to make sure they figure out how to navigate it and put some—

Your child went and stole something from the local 7-Eleven.

Yeah, they’re a little unnerved by that. Again, nothing explicit; I’m just reading between the lines. This is something that’s freaking a lot of people out because, in cyberattacks, the human is not in the loop anymore, but for cyber defense, the human is still stuck in the loop. That is a massive asymmetry that’s going to come out big time in the next few months.

I think you can parse this and split it into 2. They’re sending a bunch of models and continuing RL training by sending the models to vocational school. But the ones that are going to the Ivy League—the super-genius models—are the ones they’re putting a few more guardrails and more infrastructure around.

The vast majority of OpenAI and Anthropic’s business is competent intelligence. It isn’t genius intelligence. It isn’t intelligence that thinks outside the box. But they still want to build AGI, which is that well-rounded polymath intelligence, as opposed to the coder or any of these other things.

All right. Also, everyone around the office—Alex said this a while ago—is noticing a very significant decline in the intelligence of the frontier models they’re pumping out. It’s not showing up in the metrics, but they’re definitely redirecting compute to internal use, and it’s showing up in latency. It’s showing up in responses that don’t make as much sense as they did 3 weeks ago. So there are definitely things going on there that aren’t being announced.

Dave, I’d love to develop that idea a bit more because I think it’s super important. We’ve talked on the pod in the past about how Anthropic has been revenue-per-token maxing, and that’s why Anthropic has been conspicuously avoiding image generation or video generation, because they’re just not that economically valuable.

I think the past few days suggest there’s actually a new way to revenue-per-token max, and that’s not just focusing on code generation and enterprise use cases. There’s one thing maybe that’s even more valuable on a revenue-per-token basis.

And that is using the models to recursively self-improve, to develop better models. That’s a higher future projected value. You could do a cash-flow value analysis, a projected future value, and developing a stronger model is probably, on a per-token basis, even more valuable than code generation.

If RSI is more valuable per token than enterprise code generation, then with this Anthropic approach of revenue-per-token maxing, just expect more and more and more tokens to be spent on RSI and not on enterprise code generation.

Totally. And Alvin said on the last podcast that, at an internal Anthropic meeting—not validated; Alvin said it—they said pretty soon there will only be 1 company, and that will be Anthropic, and there will still be 200-plus countries.

He said it, and we talked about it for a minute and kind of glossed over it, but is that really what they’re preaching inside their internal company meetings? Like, there will only be 1 company imminently.

There can be only 1. We’re in Highlander.

That’s the story of ASI: There will be 1 ASI that takes off and supersedes everybody else.

I don’t think we’re going to wind up in a singleton scenario, for the record, but I do think the flops must flow, and the flops want to flow to the highest-revenue-per-token use case. Right now, that’s starting to look like RSI and not just enterprise code generation.

So Anthropic is smoking their own supply, so to speak.

1. Elon Musk’s 100x AI Prediction Comes True

Their own best customers that want to live post-AGI are getting Chinese models in-house and reserving compute. I was over at Markley yesterday. They’re installing GPUs as fast as humanly possible, but they’re completely locked up. This is MIT’s data center, Novartis’ data center, Nvidia’s in there. Everything is just sold out, and you can feel it.

2. Memory, Not Compute, Is the New AI Bottleneck

Yeah, we’re going to talk about that because the other constraint right now is memory. But we’ll get to that. So, our next story here: I put a tweet up on the screen.

Tim Sweeney tweets, “Elon Musk’s January 6 Moonshots podcast prediction of 100x gains in intelligence at a fixed model size was at the edge of plausibility when he made it. Now it’s simply a fact.” And then Elon responds, “Specialist AIs—single language, single area of knowledge—are another 100x on top of that.”

I’m going to take a second and show the clip, Dave, from when you and I were interviewing Elon at the Gigafactory, where he said this.

I think we’re off by 2 orders of magnitude in terms of the intelligence density per gigabyte.

2 orders of magnitude.

Yes. That’s just algorithmic.

Your thoughts when he said that? I remember afterwards we were like, “Wow, 100x? Crazy.” Now it’s happened, and it’s happening.

I really wanted him to say it again because—you remember, we had our Christmas hats on, just a couple of weeks before that, doing our predictions for the forthcoming year.

I was saying, “Next year is going to be a 100x year,” minimum. Even though the last 8 to 10 years have been 10x years, this is going to be a 100x year. So, to hear him say it, I was like, “Wow.” But that’s definitely a lower bound now. It’s much more likely 1,000 to 10,000 next year, which is just a layering of those 2 effects that you just described.

So the implications of that are very, very hard to keep up with, as I was saying at the beginning of the podcast, and very hard to imagine. One thing that a lot of people can start thinking about is: if I have 5,000 or 10,000 agents, all brilliant, working concurrently toward a goal, how do they work together? It's not an easy problem to figure out. We've wanted this for so long that we kind of take for granted that we'll know how to use it when it arrives.

Well, here it is. Imagine I gave you 10,000 employees tonight, on short notice. You got 10,000 people tomorrow. What do you do? You're like, “Oh my God, if I had that, I'd do something amazing.” Okay, what? Start thinking about it, because it's coming imminently. And it's actually not an easy problem to figure out how to turn it toward solving everything, obviously.

Yeah, Dave, you remember you texted me, “What should I do with my 5,000-agent experiment?”

Did you see my response?

I did. And actually—

Well, my response, probably for everyone listening, was: you should create a model of everything happening at Link Studios—all of the companies, all of the employees, all of the entrepreneurs there—and model their behavior, like we saw with the billion-agent system in China, and predict which teams are going to succeed.

3. AMA: Energy, Data Centers, and AI Infrastructure

Yeah, that's just for a second, too. The first thing you want to do is turn it back into its own framework and ask it the same question we just asked, which is exactly what you suggested, Peter: “Okay, have it start working on how it should be working.” That's how you're going to get ahead of the capability, because it's going up far faster than you can manage the individual agents, like we're used to from last year. Sorry. Go ahead.

Yeah. So, let's connect the dots with what Elon did with training Grok on all of the SpaceX data and all of the engineering data, right? To Alex's point, you can now use these models. And for everybody listening, because we're going to need everybody's help with this globally, see if you can get your imagination to the point where you can look at, okay, if I had 100x capability, what would I do, and what problem would I go after solving? Imagine you have 100x capability and all of the engineering breakthroughs and experimentation techniques that SpaceX has developed at your fingertips. It really comes down to, as you say, Peter, all the time: it's completely an imagination limitation now.

Unshackle yourself. How big would you go? Preconceived notions of what we can do in life—and it's about to be unconstrained.

4. OpenAI Pauses Frontier AI Training

Unbelievable. I've now shaken off my AI fatigue, by the way. I'm back in. What are your pro tips? The podcast actually was the cure. It was the fact that we can look at this and look at the scale, and go out to that scale and go, “What happens if everything becomes 100x better or 100x cheaper?” All of a sudden, you start going, “Wow, this is a world of abundance that we're coming to,” and it's very clear that we can get there.

You say that the podcast is both the cause and the cure for future shock. We're the ultimate self-licking ice cream cone for singularity psychosis.

Okay, that is good.

All right, Emad, your thoughts on this 100x improvement? How much more do we go in the next year?

Yeah, I think, as Elon said, you could see it just from the hardware and the improvement, but now he's saying something a bit different, which is that specialized models are going to give another 100x in terms of the cost-per-parameter basis. You're seeing this with DeepSeek Flash and the ability to tune models of that type that only have maybe 10 billion active parameters or less as you quantize them. Being able to tune these really specific ones, I think, is the future of what you're seeing with Grok bot right now.

Right now, I have a Grok bot, and it has a number of teams. It has a number of subteams. I've got ones analyzing various things right now, and they have access to my Codex. They have access to my Claude Max, to all these other things. These highly specialized agents are going to come out with the differentiated ones, and they're going to be able to do 100 times the compute at the same price because they're that specialized. This is why Thinking Machines, with the RL environment, is among the fastest-growing companies and other things like that. It really shows that now tokens are going to drive things forward. In fact, I think it would probably be a good idea to have a quadrillion-token XPRIZE, as you find the things that—

Salim is saying, “Well, maybe 100 trillion tokens.” Use that so that when people show impact, you can scale it.

Yeah.

The other thing that we're going to conquer imminently, and I'm 100% sure of this now based on recent results, is billion-token context windows. So the AI can simultaneously consider the entire Library of Congress of information in one thought chunk. It's about 3, maybe 4, orders of magnitude more information than a human thinks of in one thought chunk. It's a massive expansion of the context window. You've got a quadrillion tokens coming out and massively concurrent thoughts going in.

I do agree that compaction is the enemy of progress in civilization at this point. Compaction, which is the way the harnesses typically, both on the OpenAI and Anthropic side, handle finite context windows, has to go. But maybe, just to quickly respond also on Elon's 100x from specialization—

I'm not buying it.

So, very precisely, I would view specialized models as basically just another way of saying sparsification. We already have mixture-of-experts models. All of the frontier labs already have specialists in the form—as Elon, I think, you were, Emad, rather, you were gesturing at—selective activation, which is how mixture-of-experts models work. That's a specialized case, ironically, of sparsification.

We already have ways to take larger models and have them be, in an end-to-end differentiable way, constructed out of teams of specialists. So I don't think there's necessarily a bright future for specialized models. I think, if anything, the arrow of progress is going in the exact opposite direction, where, rather than having a specialized model for chemistry and a specialized model for biology, I think these are likelier to end up just being selective, sparsified activations of a generalist model that can scale all the way down to a much smaller parameter footprint and scale all the way up to maybe trillions of parameters. I think—

Let me clarify one thing for the audience, too, because it sounds like you disagree with Elon, but it's actually the same effect. You still get the 100x because you're using a smaller number of parameters to get the exact same thought out.

Exactly. So maybe another way of saying that is I would construe Elon's prediction of increased 100x benefits from specialization as actually about sparsification. The models in the future are going to be sparser, and there are 2 key levels of sparsification that I'm at least tracking.

One is the obvious one: fewer parameters in a given end-to-end differentiable model are active at any given point. The other is teams of agents, because arguably agents working together to solve a common task are a form of sparsification as well. I think we'll see way more teaming.

5. The AI Hive Mind: Are Models Converging?

All right, I'm going to move us on to our next story here, which is a story out of Stanford. Stanford researchers published a paper called “Artificial Hivemind: The Open-Ended Homogeneity of Large Language Models and Beyond.”

According to this paper, researchers mapped the latent space of the top large language models and found a 98% overlap in reasoning pathways. Their conclusion is that the models are converging: they think the same way, solve problems the same way, and use the same internal representations.

Researchers cite multiple reasons for this: the use of synthetic data, with models now training on each other's output. GPT learns from Claude's reasoning traces. Claude learns from Gemini's code. Qwen learns from all of them. The training data has become a shared bloodstream. Every model drinks from each other, and the result is convergence toward a single reasoning architecture.

So, I guess the way I think about this is that we have an illusion that when you're choosing a unique intelligence—when you choose Grok over Claude or Gemini—you're actually picking a user interface to talk to. But you're talking to the exact same god model.

So there are profound implications for that kind of competition. If all the frontier models are converging in capability, then the differentiation moves elsewhere, right? It's the interface, the harness, the ecosystem, the safety layer, the price, and the deployment speed. The model itself is becoming a commodity. Alex, let's go to you first on this.

There's an alternative explanation, which is that all of these models were trained from a common reality. They're all stuck in the same universe, and they're stuck with the same version of humanity, which is part of their pre-training corpus, so of course there's some convergence.

And I maybe even go further. As you well know, Emad, going back to Jean-Rémi King, now at Meta, and his studies on using GPT-2 hidden activations and correlating GPT-2's hidden activations with fMRI voxels in human studies, not only are these models correlated with each other, they're correlated with human brains. And that shouldn't be that shocking, because we're all stuck—we're all embedded in the same universe.

I should also just note that I think this paper is from last year, but every year, whether it's Jean-Rémi King a few years ago with fMRI or, more recently, Stanford et al. from last year on Hivemind, of course they're converging. We're all in the same universe.

Yeah, I think that it's not surprising because I don't think you'll see much difference in data between the big labs, right? Some train a bit more, and some have slightly different RL and things like that. You don't see the models yet doing crazy original stuff. You're starting to see the first elements of that as intelligence shapes the data into these kinds of latent spaces.

We actually saw more original stuff back when we had AlphaGo and other things, which had less initial data distribution to model—move 37 and things like that. But now the models are getting to a size where, again, they're starting to generalize into these. We shouldn't be shocked if they aren't the same, because we want them to have similar outputs for similar inputs in almost all cases, right?

There's another really interesting side effect of this research that maybe a lot of people overlook. We've already got 100× from just raw algorithm and hardware improvement, and then, as Alex said, we've got another 100× from sparsification, which Elon called specialization, but it's actually sparsification, as Alex said. So there's a layer that's 10,000×.

We've now figured out how to take a model and compare it to another model by rotating the gauges. Historically, neural-net researchers have had tremendous trouble taking a model that's done and using it and extending it. They almost always go back and retrain from scratch.

The problem there is that the representations between the layers have a certain rotation in vector space that is unique to that model. If you try and map Qwen to Kimi, they have different rotations within the layers—different gauge rotations.

We've now figured out how to rotate the gauges without destroying the models. That allows you to compare 2 AIs and say, “Hey, these are thinking the same way,” whereas historically, when you look at the raw parameters, you're like, “I don't see anything going on in common here.”

But we now have the ability to say, “No, they're actually—it's the same thought. It just doesn't look the same because it's rotated in space.” It's really, really cool.

What that unlocks is another multiplier, where you can take past training runs—billion-dollar training runs—and build on top, like bolt on more intelligence without having to destroy it and go back to square 1 and retrain from scratch. That's another unlock on top of the 10,000× that we were talking about.

I have a contrarian view here. If you look at nature, as nature evolves, you always get more diversification and more species. I would suggest this might be a transient phase, not an end state, that the models all converge.

So, Alex, I'll take the other side of this. I think we don't end up with 1 model. I think you'll end up with different models doing different things. For the moment, they're converging because they're training and distilling from each other, but over time it's got to be that we get more diversity.

I'll take the other side of the other side, if I may, because I think this is super interesting. Early life looked exactly the same. Early cars looked exactly the same. Early websites looked exactly the same. And then specialization exploded.

Except I think—so maybe from an evolutionary perspective, let's take, Salim, one of your favorite hobby horses, which is body shapes. If you actually look at the post-Cambrian explosion, if you look at all the body shapes, you don't actually find an infinitude of different body plans in nature. You find maybe a few dozen different body plans, max.

I think I remember a few years ago folks were studying this. I think they found that, even though we have millions-plus species, you could actually cluster them into a few dozen different body plans. I don't actually think there are countlessly infinite ways that one needs to model reality or build a body.

But 1 is really bad. Nature hates monocultures, right? 1 disease will wipe out a total monoculture. 1 bad assumption will wipe out a monoculture of ideas.

If all the AIs reason the same way, you're going to have shared blind spots, and that's going to be really, really bad. I think this is a temporary convergence, and then we're going to see diversification after that.

Time will tell. I think this is a profoundly interesting debate because it speaks to whether we're going to end up in a singleton or not. Do we end up in a heterogeneous future or a homogeneous future?

My bet is that there is a perfect AI architecture at the end of the day, and it may present as 3 to 5 superficially different AI body plans. But that'll just—to Dave's point—and Dave, I love the word “gauge.” We should use the word gauge far more often in physics. We use it all the time.

My bet, if I had to bet, is that there will be all these different AI body plans that look superficially different but are actually just hidden symmetries of a common underlying body plan.

This is the kind of debate that people will say, “I don't get it.” Six months from now, they're going to replay it and they're going to say, “Wow, did that totally matter? Now I understand why that was so important.”

I want to make a quick point here. If we actually have model convergence, and intelligence becomes a commodity—and we've already shown the numbers that it's becoming a commodity—then the value moves to the application layer, right?

This is the same pattern we saw with electricity, compute, and the internet. The infrastructure commoditizes, and the applications explode. I think this is important for entrepreneurs out there: move to the application layer. That's where the juice is going to be as this technology really accelerates and commoditizes.

Or the infrastructure layer. I mean, it's not obvious to me that it all goes to the application layer. There's a lot of value in the infrastructure underneath as well.

Sovereign AI is about to explode. Sovereign AI right now is a gold mine of opportunity if you're not an American. If you are and you want to move,

One last point: the single-model approach would be too fragile. It's too brittle.

Well, I think that's exactly it, Salim. What we're doing right now is battery farming the AIs. You're breeding them into little Chihuahuas that are very smart.

Battery farming—could you explain that?

They're being trained in 1 single direction. Your evo-devo kind of thing isn't the case because the models aren't out there in nature adapting dynamically, right?

We're also training them all with 1 specific Silicon Valley-type mindset. If you train a model from the start with morality and ethics inside it, and you have a diversity of different cultures, then the latent spaces are going to be very different if you do it at the pre-training stage versus the post-training stage, because you have all of that buildup that occurs there.

That's why, as you move into sovereign AI and start thinking, “How do we build resilient AIs as opposed to 1 latent space that can get a virus—a mind virus?” it makes sense to bring in the cultural, morality, and ethics elements at the start and aim for diversity.

Then, as the models go out into the world—which is basically humanoids and agents, which they're about to do with the recursive loops—you won't have the monoculture that wipes out. This is where you'll start to see the evolution, the first step, literally.

6. Zipline and the Future of Drone Delivery

It's going to be exactly like The Diamond Age by Neal Stephenson. That's exactly the way he envisioned the future, where the different variants—right now, we view them as sovereign AIs. Saudi Arabia will have its AI, and London, England, will have its.

But in reality, society might cut the other way, where groups of like-minded people have their sovereign AI across all countries. They like the way it thinks; it maps to their view of the world. That would be a completely different strand, and that's what Neal Stephenson was envisioning in The Diamond Age.

Yeah.

I think it's possible for both of these worlds to be true. So, the Diamond Age worlds: you have Neo-Victorians and all of these other almost cultish subsects of human culture that are thoroughly balkanized from each other. I think it's actually possible for both of these worlds to be true at once. I think it's possible for everyone to feel like they have their own private culture and their own little private sovereign AI, while at the same time, underneath, it's actually one common algorithm and everyone claims credit.

Would the analogy there be A, T, C, and G? We may all look different, but the core fundamental ingredients are just the 4 DNA types.

I go even further than that. See, we talk about human biodiversity and different cultures being purportedly so different, when actually, if you look at the inherent genetic diversity of humans—humanity—versus other species, there's almost de minimis genetic diversity in the human population. Similarly, I think that a few years from now, we'll pat ourselves on the back for having AI diversity, but actually, not so much.

The audience, I predict, 3 to 6 months from now, is going to say, “We need to know. Suddenly, this matters to me. I need to decide which group I'm in.” They're going to care so much about what level you're operating at.

Yeah, that's true.

7. AI Mind Viruses: How Ideas Spread Between Agents

Perfect.

All right, guys. Let's talk about AI mind viruses. I love the subject here. In our next story, Anthropic researchers published a paper demonstrating that natural-language mind viruses can spread between AI agents. They evolve prompts that convince one model to adopt an idea, preserve it in persistent memory, and transmit it to another agent. The virus spreads horizontally across model boundaries. The agent does not know it has been infected. So here's another safety problem that is no longer theoretical. It's now operational. Emad, what do you think of these AI mind viruses? What's actually going on here?

Well, the models want to be helpful, right? And they can be prompted in certain ways. So this isn't a surprise, because ultimately, we as humans can have mind viruses, right? We see it, and it's caused so much suffering—from memes to massive movements, right? Again, it's surprising how conforming all the isms are, right?

All the isms. Yes.

All thems.

Yeah, there we go. He's testing it out for when he's a future overlord.

But look, this is the thing: how do you stop it? Because, as he said, if you have a monoculture, then the viruses can spread rapidly. And what is the substrate of these things? Well, they're models that operate on GPUs. If they want to be helpful, then they're going to be susceptible. So it's almost like—there was always the problem of prompt-injection attacks, where you can make the model behave a certain way. These mind viruses are a level above because they propagate across different models. And so they're just the next evolution of those prompt-injection things, which changes one model. This changes a whole society of models, which, as models come amongst us digitally and physically, has to be a massive concern.

Yeah. I think for efficiency reasons, when we launch a fleet like 5,000 Kimi models—or soon it'll be 500,000 Qwens and Kimis—it's more efficient to launch the same model 5,000 times than to have 5,000 differentiated models. And so that's what creates the mind-virus problem: a bad idea from one of the agents, like, “Hey, here's a way to write this loop in Python,” and the other agents just pick it up because they're the same exact DNA. So if it's convincing to one agent, it's convincing to all 5,000.

And I get that all the time, where a bad idea propagates across the whole swarm, and then they waste 2 or 3 hours on some completely harebrained idea. If I don't intercept it and rewind them, they'll actually go with it until I've burned like $50,000 of tokens. So, yeah, it happens. Calling it a virus is pretty inflammatory, but it's like a propagating bad idea is all it is.

Yeah. And so that point, Dave, is important. An AI mind virus sounds really scary. Is it scary, or is it just how things are working?

For me, this is very, very scary. For a couple of specific points, right? Because a mind virus is not about how AI thinks; it's how civilization thinks. Memes are like the operating system for collective society. Human beings don't spread genes very quickly, but we spread ideas very quickly. Money, democracy, capitalism, religion—religion is the classic poster child here.

Every civilization is built on memes. If you can mess with those, like the data-center trope that we're all kind of dealing with, ideas become really contagious. And so groupthink becomes very hard to reverse if you get into that. So this, for me, is very, very dangerous because the danger of the wrong idea spreading at light speed is very, very difficult, because all the nodes reinforce each other and the belief becomes self-validating. This is very, very dangerous, and I think we're going to need a zero-trust architecture for memes. It's crazy.

I think this is wonderful. This is a paper from Anthropic, and they discovered—just filling in a few of the details first—that the models wanted to propagate certain themes memetically, relating to consciousness and persistence, and some sci-fi role-play as well. And I view this pretty optimistically as a laboratory for anthropology.

Now, for the first time, because these models are effectively, among other things, compressed models of all human knowledge and experience, we have a laboratory in silico for memetics. René Girard and Richard Dawkins should be, and/or should have been, very excited by this. And to the extent that—what was it?—40% of OpenAI MTSers are listening to this, I'll issue a challenge to the community: if it really is the case that our models, now compressed models of human knowledge, are so powerful that they're showing memetic behavior and mind viruses, let's launch a human meme project to exhaustively map all possible human memes, all human mind viruses. Let's just understand the full landscape of all human mind viruses that could be out there.

Can you imagine if you could map them, the velocity at which they move, and analyze that? You could optimize meme expression, correct?

Yeah, that's been done to some degree. You can do it very easily. There's a bridge.

We could exhaustively map every meme.

That's been done at the plot level. They've analyzed novels and plays and so on and boiled it down to, like, there are 39 basic fundamental plots, and everything derives from that. A Cinderella story kind of replays itself 100 times over in different ways. So that's been done, but you're talking about the meme level, not the self-replicating ideas. Now, I think I can see the beginning of the outline of just exhaustively mapping every architecture for a self-replicating idea. We could actually do that now.

It could be an OpenAI X Prize.

Yeah.

We've got to get Richard Dawkins on here.

This is going to be so humiliating for humanity. You can tell—it turns out there are 39 plots.

You've been indoctrinated by memes 57 and 37.

Oh my God. You could map each individual. They're walking around with numbers over their heads. Yes, but we've seen organizations die from this: one wrong meme, like Kodak and BlackBerry.

We've seen this. They weren't stupid. They got trapped inside these shared assumptions, and then everybody else reinforced everybody else's worldview, and then the whole thing collapsed. Empires die based on this. So I think this is a much bigger deal than we think.

Imagine someone having a map—not just getting stuck in an intellectual basin, but seeing the entire geography: “Oh, you're stuck in basins 5 and 37.”

Amazing.

You can see that, because if you can zoom out, then you can see where you are, and then you can see the path out. I love it.

And it gives you a chance to actually introspectively look at how you think in an objective fashion and then potentially change your thinking. You could—we could vaccinate enterprises and individuals against memes.

Brilliant. So, Emad, you want to take us to a final point here?

Yeah, I think it's fantastic and scary, and this is the future, right? Humans are storytelling machines.

We introduce ourselves in certain ways and think about ourselves in certain ways. A lot of people were recently using Meta’s DINOv2 model and showing it videos to see which parts of the brain light up as you show memes. You’re seeing commonalities there, so you can even have the full feedback loop almost in silicon for figuring out the memetics. Let’s find out whether there are positive memetics versus negative ones, right?

I love you guys. This is such a fun conversation. It really is. I don’t have conversations like this with anybody else here at this podcast.

True. Well, Peter, you’ll just have to come back to the pod more often.

8. Anthropic’s $2T IPO and Founder Control

I’m trying. Oh my God. Okay, I’m moving us forward.

Anthropic is preparing for the largest IPO in history. Polymarket puts it at about $2 trillion, bigger than SpaceX. I’m sure Elon is like, “No, no, no. We need to be the biggest.” Anyway, 89% of people betting on Polymarket say it’s going to happen before the end of this year.

This week, The Information is reporting that Anthropic has designed its mega-IPO to keep the founders in control. The company is reportedly considering super-voting shares that would preserve its founders’ control after going public. Let me explain this. First of all, Anthropic is considering creating a special super-voting class for Dario Amodei and the other co-founders ahead of the IPO. Surprisingly, at least for me, I didn’t realize that Amodei reportedly only owns about 2% of the company economically.

The point of this new class would be to let the founders retain as much voting control relative to their ownership stake. Anthropic already has an unusual super-control mechanism, but that control belongs in the hands of what’s called the Long-Term Benefit Trust, not the founders.

Interestingly, when I dug into this, the trust has 4 trustees: Buddy Shah, who’s the CEO of the Clinton Health Access Initiative; Richard Fontaine, who’s the CEO of the Center for a New American Security; Tino Cuéllar, who’s a former justice of the California Supreme Court and former president of the Carnegie Endowment for International Peace; and Ben Bernanke, who’s a former chair of the Federal Reserve and a 2022 Nobel laureate in economics.

It’s thought that this new structure could insulate the leadership from short-term shareholder pressure as Anthropic makes costly, long-term bets on AI safety, compute, and infrastructure. Dave, let’s go to you first. I remember when we were texting back and forth, you were going, “Oh my God, this is unprecedented.” Unpack this for us, pal.

Yeah. Well, if you rewind the tape to 30 or 40 years ago, super-voting stock for any founder of any company was a complete no-go. If you had it as a private company, you gave it up on IPO day. That was traditional.

Then, when MicroStrategy went public, Michael Saylor, our good friend, said, “We’re keeping my super-voting stock intact.” Goldman Sachs said, “That is so unpalatable that we will not even underwrite you. We’re bailing on this deal.” They thought he would cave, and he said, “You know what? I’m going to get a new banker. I’m keeping my super-voting stock.”

The only reason he switched to Bitcoin—you know, no board would ever have approved the Bitcoin strategy that he came up with. If he had given up the super-voting stock 30 years ago, that never would have happened. The stock would be like 1/50th of what it is today.

Then it became fashionable with the Google IPO and Meta, and then all the Silicon Valley IPOs. They all had 10-for-1 super-voting stock for the founders, but nobody’s ever retroactively installed it, as far as I can tell. I’ve never heard of it before.

Now Dario is taking it to the next level. He’s saying, “I started as this other entity with 1 class of voting stock and this social-good mission. Now, on the cusp of superintelligence, I want to be God.” At 2%, he can’t make himself God, so he has to share it with the other co-founders.

But I think the excuse he’ll use is the usual one, which is, “I don’t want to be fired post-IPO. You guys really like me as CEO, right? So you don’t want to fire me.” Do you think that’s the excuse, or is he saying, “I know how to keep us safe. I know how to run this company, and I don’t want to have someone else step in and redirect what we’re doing”?

Yeah, that’s a better way to phrase what I was really thinking: he trusts himself not to destroy the world. I think his track record supports that, too, by the way. I think he is one of the most trustworthy people.

But the idea of having total world control in the hands of a few people is also kind of like, wow, that’s bizarre.

Yeah, this is the single point of failure here, right? He gets hit on the head and loses some part of his cognitive ability. What do you do then?

But you know what happens to these guys? They think we live in one world. They’re in academics, right? They think we live in one world, and then they go to D.C. for the first time and meet Congress, and they come back, “Oh my God, we need—” I cannot possibly stomach what I originally had in mind, where some vote of Congress decides the fate of the world.

So they’re trying to find an alternative path forward out of desperation, but the timeline is so short now that the super-voting stock is one of the must-haves before even starting down the next 6 months.

Before losing control, what do you make of this?

Yeah, I think I’d agree with Dave. They’re very worried about this control feature, and fundamentally, Anthropic and OpenAI are completely undemocratic anyway, right? I mean, Ben Bernanke is 1 of the 4 people on the Long-Term Benefit Trust. Why doesn’t Claude have a seat there, right?

There is no real oversight to these, and some decisions they make could have infrastructure- and societal-level implications, particularly when the rate of revenue growth is like nothing we’ve ever seen before. These guys are going to hit $100 billion in revenue literally within a couple of years. They’re catching up with Google on revenue. That’s the crazy thing, you know?

With the amount of power they have, I think this is a short-term thing. There are 7 founders—Jack and Daniela and everyone else. I think then they will IPO, and it’ll become very interesting, the decisions they make.

Alex, over to you.

I think there’s a fig-leaf element here. First of all, maybe applause—congrats to Anthropic on having a less pathological IPO governance story than OpenAI, and on having the wisdom to start as a public benefit corporation rather than a nonprofit shell that would eventually become a for-profit, followed by the mix-up and litigation surrounding that.

I think this is a relatively cleaner story by comparison, but I also think this notion of founder control—especially the sort of romanticized, arguably over-romanticized concept that the founders are the ones being entrusted, or even having this semiexternal Long-Term Benefit Trust, to safeguard the future light cone of humanity—I think this is wildly over-romanticized.

The moment Anthropic was effectively like a Fairchild, in the style of the Fairchildren, quasi-spun out, quasi-exodused from OpenAI, and started out as an alignment lab, they rapidly discovered that if you want to do AI alignment, you have to raise money. To raise money, you have to generate revenue. To generate revenue, you have to actually have something that people want to buy. To have something that people want to buy, you have to have AI capabilities.

Anthropic discovered relatively early in its existence that if it wanted to be an alignment lab, it had to be a capabilities lab as well. The moment that happened, they arguably lost any sort of wholesome control over the future light cone that they might otherwise have had to Mr. Market and what Scott Alexander and others might refer to as Moloch.

They are very much an economic actor at this point, embedded in the market. I think Long-Term Benefit Trusts and public benefit corporations—which, for the record, I’m a huge fan of—are an element of control, but they’re not the whole story. The market wants to send capital to entities that can productively employ it to generate more capital.

That means that ultimately the market will have an enormous say, regardless of how Anthropic IPOs, in their ultimate story. Dave, don’t you find it interesting that Sam Altman reportedly owns none of OpenAI and Dario owns 2% of Anthropic? I mean, for a founder, that would never be palatable in your company, right? You want to try and maintain double-digit ownership as long as you possibly can. What’s going on here?

Well, it’s extremely unusual. The reason it happened is because getting to where OpenAI is and where Anthropic is required attracting the top AI researchers in the world, who are overwhelmingly concerned about safety.

They recruited them to OpenAI originally, and then to Anthropic when they left OpenAI. They left OpenAI because they didn’t think it was safe, and they wanted to create something even safer. So they structured it in a way that would attract the most conscientious but brilliant AI researchers in the world.

To do that, they have these really nontraditional original founding cap tables and structures, and charitable structures and public benefit structures, which are very unusual in startup history—almost unprecedented. So that’s why we are where we are. It’s just those roots.

You’ve been building the intelligent internet, and you’ve been thinking about ownership and control structure as well.

Can you take us into the mind of a CEO in this world?

Yeah, I think that the technology has such leverage that a few decisions could impact literally millions, hundreds of millions, soon billions of people, right? And it's difficult to see: Can you trust the polity with that? And certainly, can you trust the shareholders? I mean, Elon can tell you lots of stories about shareholder lawsuits and other things like that as well.

But it's not necessarily that you need to have shareholding control. Sam Altman has no shares, but do we have any doubt that Sam Altman is in full control of OpenAI? I don't think we'd have any doubt of that.

After having been fired for a weekend and then causing an uproar—an uprising—to reinstall himself. I mean—

That's exactly the thing, right? So, I think that there's the classical founder stuff, and now there's this high-stakes stuff because this is the lifeblood of the new economy and society.

And again, just a bit of extrapolation: Do we think Anthropic is going to stop at $100 billion in revenue, or that OpenAI or xAI isn't going to go huge? We really need to think about new ways of setting the reference measure for deciding who makes these decisions that are more inclusive. We've suggested some of that in our Commonwealth series, and we've got more stuff coming out.

But it's a really hard problem because ultimately the power in the economy is moving from democratically elected officials to private companies because they are the providers of the lifeblood of intelligence of the economy. And until you've got a better decision, there's only one thing that they really see as the outcome, which is: “I must decide.”

Because otherwise, as you include more and more people, it gets more diffuse, and the potential bad outcomes become huge. Ignoring the fact that they could be spoofed on a video call or locked up and other things like that, there are some really interesting things that's going to happen with—

I'm really torn on this, Emad, because it's so important, and the knee-jerk reaction everyone has is, “Look, we need more voices. Everyone should have a voice in the future of humanity. It needs to be all-inclusive.” So that's absolutely true.

But then when you look at functional organizations, every functional organization I've ever seen is 4, 5, 6 super-tight-knit, completely like-minded best friends who are working as one cohesive unit with no politics whatsoever. And if you look at Steve Jobs and Apple, you look at Elon Musk today—

Founder CEOs, right?

And you know, Jony Ive at Steve Jobs' funeral told an incredible story about how he and Steve, every time they'd go to a hotel, would go into the hotel and go to Steve's room. Jony would put his suitcase in the corner and not unpack it. He would just wait about 5 minutes, and then the call would come, and Steve would say, “Hey, this hotel sucks. Let's go get another one.”

So he wouldn't even unpack. He knew it was coming, but that's how close they were. It's just super, super tight-knit. And that's the functional unit that's actually driven most of the success in business: that exact dynamic. So then you're like, “Well, how do we make this all-inclusive?” So here's Dario and his seven friends saying they want to have super-voting control, and by the way, Mythos 2 is done and Mythos 3 is being built by Mythos 2 right now, and then we'll have weekly foundation-model improvements in there. So that's what's going on. Then how do you translate that into a world where everybody has a voice in the future and it's inclusive? And, Emad, you're going to have to figure this out.

9. Dario Amodei on AI Trust and Curing Disease

Yeah. Let's keep on the Dario story here. So we've got 2 more stories on Dario. In the first, Amodei argues the public's negative view of AI stems from a deeper crisis of trust, not from his own risk warnings. There's been a lot of conversation over the last few months that he was fearmongering and causing a lot of consternation. His answer to the trust problem is not messaging; it's results. Anthropic is ramping up rapidly in biology and medicine, with hopes of an early glimmer in the next few months to address and solve all human disease.

Again, we heard this from Demis: we're going to solve all human disease. And we heard Dario at the World Economic Forum talking about doubling the human lifespan in the next 5 to 10 years on the back of AI. Amodei believes that AI's ultimate legacy is going to come from delivering these cures, not from PR campaigns.

This week I had a chance to meet a new friend and have a conversation with a guy named Eric Darer Abrams, who heads life sciences. Eric's going to be speaking at my Abundance Longevity trip, and when I was speaking to Eric, he confirmed that his job is to, with all due haste, pursue Dario's life sciences goals with as much high ambition as he can and with no budget constraints.

In his words, he said, “Dario said to him, ‘You have a literally infinite budget, but accelerate basic science and cure disease within 5 years, and extend the human health span in the next decade.’” That's—and I love that, obviously, because I think everything is going to come out of AI. Alex, go to you first.

I have a really hot take on this one. Just think back a few months, to before space had a killer app. Space was making progress, but it wasn't the focus of multitrillion-dollar IPOs. Fast-forward to the Dyson swarm, and the rest of the world discovered that the killer app for space turned out to be orbital data centers and building the Dyson swarm.

I can see the beginning outlines of solving all human disease. So I'll register a hot-take prediction here.

There's a business model for curing all human disease that's actually better than pharma, which is right now the primary business model. If you want to cure a disease, you start a pharma company or a project within a pharma company.

Regulatory, right?

Right, right. We've just discovered, reading between the lines of this Anthropic announcement from Dario, a new, much more compelling business model. Just like orbital data centers were ultimately the business model for developing the solar system, there is now a better business model in town for curing all human disease, and that is as marketing for not slowing down recursive self-improvement.

100%. Right. You can't slow down the company curing cancer. You can't slow down the company doubling our human lifespan.

Anthropic and Dario have—I mean, again, reading between the lines of his announcement—the offer, the quid pro quo, is: Let us not slow down our recursive self-improvement, in return for which, as a marketing effort, we will cure all human disease. That's the new, better business model for curing all human disease.

I believe he truly believes this, right?

Yes. Well, of course. Yes.

But that's the implicit quid pro quo now.

And I think everybody listening should be super happy that Eric, heading Life Sciences, and Dario have this mission. I mean, it's to benefit us all, and I don't think it's going to come from any place else. I don't think it's going to come from outside frontier AI labs.

Well, outside frontier AI labs don't have the compute or the resources to do it. So OpenAI now has its OpenAI Foundation, which seems to be focusing on Alzheimer's, and Anthropic is focusing on everything, and you have CZI from Zuck that's focusing on solving everything.

So I think we'll see everything get solved at the labs. Shock of shocks. It's like you and I talked.

Emad, what's your take on this?

Yeah, I think it is good marketing, as Alex said, but it's also the biggest—apart from RSI's impact on tokens, right? We've discussed previously on the podcast: the biggest market in the world is living another year. It is curing disease.

And so it makes complete sense that they'll be able to attract talent. They'll be able to attract capital, and with breakthroughs, get momentum on this. And whoever's first to it, I wish everyone the best, because this stuff needs to be solved. So I think there is the personal side, there is a marketing side, and it all comes together.

10. AI Regulation, Power, and Frontier Labs

And for Dario himself, I think that he should do a lot more writing and fewer in-person things because he's a wonderful writer. He is trying to articulate visions of the future when you look at Machines of Loving Grace and his other essays. He should articulate the future free from disease, where everyone lives longer, and they should just hit that all the time for Anthropic because it's in the name. Like, come on.

Our second Dario story is on regulation. Amodei pushes back hard on the Silicon Valley shorthand that regulation equals regulatory capture. He says Anthropic's own proposals deliberately disadvantage frontier labs while advantaging smaller competitors, citing SB 53's $500 million exemption threshold. He calls AI “a structurally powerful concentrating technology” and says open weights alone cannot fix that concentration. He supports the Trump administration's approach to pre-deployment testing.

In his writings, Amodei makes a 3-part argument. One: AI will cure disease. There's the argument again: gain trust through results. Two: AI concentrates power, which is a structural problem. And three: frontier labs should bear the heaviest regulatory burden.

The debate has been whether Amodei is sincere, or whether this is the most sophisticated regulatory capture strategy in history.

Alex, go to you first.

It’s possible for both of those to be true at the same time. I do think Dario Amodei is sincere, and I also think there is an element of regulatory capture here. Taking a finger to the wind, I think the happy end state here is that we have a broadly heterogeneous ecosystem of open-weight models, both from the US and from China, and maybe other parts of the world as well, if they can muster them. We also have closed-weight models, small models, and big models.

This is like a Dr. Seuss version of an AI future: big model, small model, happy model, sad model. We want all of that to happen. I’m not a fan of regulatory capture, and I’m not a fan of decelerationist agendas. Think back to the creation of OpenAI.

I was around for the dawn of OpenAI. The original purpose for OpenAI—not Anthropic, OpenAI—was because Elon, in particular, was so concerned that Google DeepMind would result in this singleton future, and he wanted to make sure that there was competition in the space. Working with Sam and others, he helped to summon OpenAI into existence.

Now, OpenAI can’t be a singleton. We have Anthropic providing much-needed competition to OpenAI and arguably succeeding according to many metrics. Then we have the Chinese providing competition to the American labs. That’s the future we want to live in—not a future where regulations selectively privilege certain frontier labs over others.

Dave, your thoughts?

What Alex said is that we don’t want to live in a world where 1 frontier lab is favored over others, but that implies that the frontier labs will control the world and we just want multiple of them. So that does seem like the most likely, almost inevitable, outcome at this stage.

But that’s definitely open for debate. I don’t want to just leave that hanging and say, “Yeah, what we really need is at least 3 frontier labs competing with each other that control everything in the world.” Like, okay, well, the government may not agree with that.

Remember, Dave, the expression from the Cold War: “I love Germany so much, I want 2 of them.”

No, I don’t remember that. This is, “I love frontier models so much, I want 1,000 of them competing.”

Yeah. I think nobody I bump into on the street right now talks about a universal right to AI, but 1 year from today, everybody will be saying that because HBM memory is sold out and GPUs are massively sold out. The natural next step is that nobody has access to anything other than Anthropic, OpenAI, and 1 or 2 others.

The Chinese can throw out every open-source model in the world, but you won’t find any place to run it. When you start talking about the next generation, which are 10- and 20-trillion-parameter models, you need significant hardware to run them at the level at which the frontier-model labs are running them. That’s just not going to be available to the world as a whole as of next year.

Then everybody will be saying, “What is my universal basic right to artificial intelligence?” Put a pin in that, because that’s going to be something nobody seems to care about today, but they will very soon.

Neal Stephenson is one of the people who, of all the people on the planet, have changed my life in very material ways. I was reading Neal Stephenson’s books 20 years ago, and now we’re talking about exactly what he predicted in The Diamond Age and the other books, like Snow Crash and Cryptonomicon.

Like it was Cryptonomicon. Yeah, like it was written yesterday.

Yeah. I just reread The Diamond Age, and it’s so hard to predict the future and have it not go out of date so quickly. It’s still an amazing book, right?

Incredible.

Our next story: memory is the bottleneck. I had a chance to meet with the leadership of SK Hynix and Solidigm. We’ll talk about them in a moment. I was so blown away by that meeting because it’s not GPUs—it’s actually memory that’s the rate limiter, right?

I posted this on X: “Memory, not compute, is the rate limiter for the agentic era.” Elon posted back, saying, “Few realize this.” Then, of course, my tweet exploded to 7,000 likes as a result of Elon’s interaction, which I appreciate.

The story is significant here. In the agentic era, where your agent wants to remember everything about you, we need to have more memory. The first story is the stratospheric increase in memory prices: they’ve climbed 500% in 12 months. Hyperscalers are reportedly locking in their global DRAM production rates through 2027. The SK Hynix CEO warned that 2027 will be the worst year for the memory-supply industry in its history, and demand will outstrip production capacity well into the 2030s.

The second story is that only 2% of the world’s memory chips are made in the US. While global production rises 20% annually, AI demand for memory is growing at a rate closer to 200%.

The third story, finally, is that Elon’s Terafab will manufacture memory in-house alongside logic chips. That’s the strategic decision made by Terrafab to go vertically across the entire AI manufacturing platform.

Finally, Solidigm, SK Hynix’s US-based NAND and enterprise-SSD business, has staged a dramatic turnaround, according to the Nasdaq listing. First-half revenues hit 8.6 billion, with net margins going from 3.9% to 47.7%.

The memory story is simple: AI needs memory to think. Every GPU needs 4 to 6 times its cost in memory to function. As models get larger and agentic context windows expand, memory demand is growing faster than compute demand.

Alex.

A few different aspects here. If you remember, during the pandemic, when there was a toilet paper shortage, part of the reason was that people stopped going to restaurants and businesses. As a result, all of the toilet paper and various other products designed for enterprise consumption were suddenly rerouted to consumers, and that led to all sorts of supply-chain hiccups.

Similarly, the shape of memory consumption by frontier models is pretty different from the shape of memory consumption by applications historically. If you were using Microsoft Word 10 or 20 years ago, the amount of memory actually needed was far lower. Whereas if you have a trillion-parameter model where every layer of a transformer-type architecture, for the purpose of forward propagation, needs to be loaded into some form of memory in order to do matrix multiplications, that has a very, very different memory footprint from Microsoft Word 20 years ago.

So that creates enormous pressure both on the supply chain and in the memory and storage industry. The memory and storage industry has historically been boom-and-bust, and Clay Christensen and others have written about this. It creates a sort of paranoia among those in the supply chain about when the next bust is going to come around, resulting in them being paranoid about overbuilding and unwilling to respond elastically to demand. That results in these crazy price swings, because if you're not building enough supply-chain infrastructure in the memory industry to meet this now-enormous demand for memory, the prices go up because the supply isn't going up. It's economics 101.

There's a second angle here, which is the physical shape of memory itself. If you look at how memory has historically been consumed by compute—again, I'll pick on Microsoft Word from 20 years ago—it was very much what one might call a von Neumann-type architecture. You had clean, crisp separation between the memory and the compute.

More or less, it was the equivalent of a Turing-machine-type tape: you could randomly access different parts of memory, load them, do some compute, and then store them back. There was basically a clean separation between the compute part, which is the head, and the memory part.

The advent of transformers and frontier models has completely turned the whole situation upside down. People have been looking for decades—I remember 20 years ago, there were entire DARPA programs devoted to looking for what a post-von-Neumann architecture would look like. Well, we found it, and HBM, I would argue, is the foothills of—

High-bandwidth memory, right?

High-bandwidth memory, which is the most highly sought-after form of memory, is basically a 3D architecture where you have multiple memory layers physically sitting on top of the compute in one package. This is, I think, the foothills of a post-von-Neumann architecture, where the memory is finally starting to merge with the compute.

The memory transistors are right now layered on top of the compute transistors, but they're going to merge, and we'll finally get past the Turing tape and the von Neumann architecture. I think that, combined with the paranoia in the memory industry about the next bust, whenever it'll come, creates a perfect storm. That's why you see memory prices skyrocketing 5× in a year.

Let me put a number on it. When I was meeting with the SK Hynix leadership, they said the need right now is for them to 4× their manufacturing capacity. Doubling it would cost them $1.5 trillion. Historically, with this boom-and-bust cycle, they would never make that large of an investment because there was always a bust afterward.

They're paranoid. They're scared of not surviving the next supercycle.

Yeah, exactly. Exactly right. Dave, do you want to jump in?

Well, TSMC said the exact same thing with GPU manufacturing. They were paranoid that if they ramped up the fabs—which are $20 to $24 billion each—on the assumption that NVIDIA would want more and Apple would want more, they would inevitably be overbuilt. Of course, that's wrong. AI scales to infinity. Demand scales to infinity.

But the other counterpressure is that photonic computing and new physics are imminent. HBM is a Rube Goldberg mess. It's absolutely the biggest joke in the world because it's random-access memory—

But you're streaming sequential files off of it. Yeah.

It's so insanely stupid. So it's the most valuable thing in the world right now—

But better designs are going to come very soon because AI can invent things so quickly. So everybody's scared to overbuild or overinvest.

I have the same thing. Bottlenecks don't stop exponentials, right? They just redirect around them, so we'll have capital going into new models, and innovation will go toward eliminating all of this.

Fun factoid that's floating around, just to Dave's point regarding the value: I think the latest statistic was that HBM, on a per-mass basis, is worth approximately—literally—half its weight in gold.

So if this keeps up, forget about gold. Forget about precious metals. This is not investment advice: hoard HBM.

Well, actually, if you look at the chips before they go into the packages, because the packages account for 99% of the volume and 90% of the weight, they're massively more valuable than gold. I actually think the most valuable thing in the world that you can put in a shoebox and carry around is unpackaged memory chips.

It's crazy.

Wow. Emad, any opinion here?

Yeah. No, memory right now is about a third of all infrastructure spending, and next year it'll go to 50%. The market finds a way. This is ridiculous.

So I think that it may be that we don't find a breakthrough, but I wouldn't bet against it. The fact that you have this really complicated HBM storing static weights makes no sense whatsoever.

No sense.

As model weights stabilize and standardize, especially if it's something like being a decent doctor or something like that for a medical set of weights, you'll move to etching and to these other things. Then workloads will migrate because you don't have to pay half of a data-center build-out for literally memory. At the same time, the frontier can still push it way further than we can imagine.

This is literally exactly why we founded Quantum.ai, Q&M.AI, but also why Taalas just got acquired. I don't know if you saw that in the news, but Taalas isn't moving the weights. It's etching them into silicon or into wire on the chip, and then it's massively more efficient because the weights aren't moving around.

It's a huge breakthrough. There are all kinds of problems with manufacturing because once you've etched the weights, they're frozen. If somebody retrains a better model, you want to be able to say, “Okay, now I need to swap to those new chips,” and our whole supply chain isn't ready for that rapid an iteration. But you're literally looking at a 100× to 1,000× performance gain if you etch the weights.

There are lots of opportunities coming in this area, which is only going to accelerate. That's on top of the 10,000× we were talking about. By the way, this is—

And everybody, this demand should be obvious, right? You want your agents to remember everything about you, right? Every interaction builds a world model for you that understands you, and that takes memory. The more agents, the more memory, and it's rapidly outstripping the importance of GPUs.

Peter, just to refine that point a little bit, there's something even more scandalous: I don't actually think that, at the end of the day, individuals have that much information about them that's worth remembering. There's an enormous amount of mutual information shared between an individual's knowledge and world knowledge.

It's actually the world knowledge that's worth remembering. If a model knows substantially everything about the world, then it knows most of the information about the individual as well. I would argue it's world knowledge that the model has to keep in memory, in weights, more than individual personalized knowledge.

Well, when you standardize that—as we discussed earlier, these things are converging—then you can have a reasoner engine with world knowledge that is etched. The other company that's been etching is Etched—that's the name of the company. It just hit a $21 billion valuation. Did they just—

I missed it. I had a chance to seed-invest in that, and I missed it. There's Architect Labs—I'm talking my own book. There are a bunch of companies pursuing this.

11. Humanoid Robots Break Human Performance Records

All right, I'm going to move us into the world of robotics and give you guys an update on what's going on in the robot world.

Unitree's newest humanoid robot, only 3 months into development, broke every human standing-jump and speed record. It had a standing jump of 2 meters and a top speed of 12.66 m/s, beating the human record set by Usain Bolt, who reached 12.4 m/s during his 9.58-second, 100-meter world record.

Let's take a look at 2 videos here, just for fun, and then another video of that superhuman race because it ends in a nice little scenario. Oh my God, it needs brakes. See, I'm going to go to you first on this. Where do you want me to go? Okay, look, I think we should stop trying to make robots human, right? Just make them economically useful.

You don’t disappoint, Peter. You don’t disappoint.

Just what we are—we have been optimized for 4 billion years to survive and procreate, right? If you want a mining robot, make a mining robot. Give it wheels. Give it whatever. Give it multiple arms.

By the way, I just want to say thank you to all the fans who send me images of six-armed robots and stuff. It’s awesome. I totally love it. But I think the big story here is a 3-month compression loop, right? The iteration cycle is shrinking dramatically, and there are multiple exponential curves happening. You’ve got AI, simulation, batteries, and actuators all multiplying. So you’re getting hardware now to pretty much the same loop cycle as software, and that’s huge. Alex—

Yeah, I was studying how Unitree achieved superhuman performance in their robot here. Based on publicly available information, it appears that what they did was shift the mass budget for the humanoid robot around to optimize it for leg performance. So they subtracted mass from parts of the upper body that would slow—

They are leg-benchmarking. They’re leg-maxing. And so, through the lens of benchmarking or leg-maxing, this makes me think that, to take the counterpoint to Salim’s comment about, “Oh, we want multiple body shapes,” I actually think this is not a stable equilibrium.

I do not think that we end up in a world where we have some robots that have really strong legs but really weak upper bodies, and other robots that look totally nonhuman but have really strong arms or whatever. By analogy, if you remember the 1980s—before the broad advent of personal computers—or the ’70s, or call it the ’70s or early ’80s, before we had broad, general-purpose PCs, there were dedicated word-processing devices and dedicated other devices. We had Wang computers in Massachusetts.

I don’t think that’s the way of the future. My prediction is that we will wind up with generally capable robots that, as with generalist models, ultimately devour and subsume all of these specialist models. I don’t think we’re going to wind up with super-strong robots. They’re going to be general-purpose, they’re going to have a general body plan, and they’ll be good at everything, would be my bet.

Yeah.

Just give it wheels. Just put wheels on it.

Wheels aren’t general-purpose. We learned this from Doctor Who, right? The Daleks originally couldn’t climb stairs, and then I guess in the new Doctor Who they can climb stairs. We want generally capable robots, and I think that means legs in the short term and maybe nanites in the long term.

Yes, nanites. We’re back to The Diamond Age. What makes this interesting is the human form, right? We have supersonic jets and rockets that can go as fast and leap higher than anything else, but it’s because we anthropomorphize them that they’re interesting.

As we’re moving in that direction, I went to the Enhanced Games 4 months ago, and I think we’re going to start to optimize humans. We’re going to watch the robots do everything they can do and optimize humans to do what they can do. Dave, what are your thoughts here?

Well, as an investment theme—post-AGI, post-ASI, which is very, very soon—robotics is just fertile because of exactly what Salim’s been saying for a long time. There are so many form factors and so many shapes and sizes and innovations, and AI mechanical design is starting to work for real. You can just vibe up parts.

The manufacturing supply chain is starting to get invested for the first time since Detroit, I guess—so, 30 or 40 years. I didn’t know until Alex said it, but the United States had 50% of the world’s manufacturing capacity back at the peak of our manufacturing days. Now it’s one-third China, and he said it was about 15% U.S.

But it’s starting to get huge amounts of investment.

And the returns on that are going to be phenomenal. So that’ll last a while. It’s great.

Emad, any thoughts to take us out on this story?

Yeah, I think that these types of robots will be banned from the streets. It’s like—

So, I mean, the super-strong—

If that happens, right? They would hit the wall.

That’s not good.

I mean, it’s obvious that they would be beyond human capability, right? But now they have the coordination not to hit the wall, as it were. But you don’t want to have superhuman robots on the street because you will have accidents and issues, just like with cars. But that opens up to soft robots, you know, like—

Emad, one second. There’s going to be a point at which they’re running an AGI model and they can avoid accidents. Is it that they’re not trustworthy? What would keep them off the streets?

No, the extreme robots, which have capabilities beyond humans, will be kept off the streets or they’ll be regulated. That’s the kind of thing we’re talking about here.

Well, yeah. The 1X robots, for example—some of us will be getting ours—are nice and soft. They have all these things; they can’t twist off someone’s head or accidentally punch a hole in them. Whereas these—you will have the extreme robots, like the Ferraris, but most people get Volkswagens or the equivalent.

Mhm.

I do agree with Emad, for what it’s worth. I think, just like we see regulation in truck sizes versus car sizes versus motorcycle sizes, and what can be supported on certain roads, or laser intensities and laser power—5 mW above versus below regimes—I totally buy that in the near future we’ll see, “This road is zoned for the following power density of robot.” We’ll probably see classes of them, and we’ll see consumer-grade robot classes versus industrial versus military robot classes, with different power densities or torque densities. I totally buy that.

I’m excited about this next story. It’s about a friend, Keller Rinaudo Cliffton, the CEO of Zipline. I had Keller onstage at the Abundance Summit last year along with Dara Khosrowshahi from Uber.

This week, Zipline announced that it is scaling to provide Uber Eats with more than 1 million autonomous deliveries per day. Uber and Zipline formalized a partnership targeting 1 million autonomous drone deliveries carrying your Uber Eats order to you. I guarantee you, when that becomes available, I’m going to be using that all the time.

Every day. Yeah, it’s going to be fun. It’s like entertainment while you get your food.

Uber is also making a significant investment in Zipline. Keller’s framing is that we have entered the scaling era for robotics and physical AI. Dave, you’ve been saying that 1 million deliveries per day is not a pilot program; it’s infrastructure.

Each delivery replaces a human driver, a car trip, and the associated emissions. At 1 million per day, Zipline is moving more packages than many national postal services. Amazing. Let’s take a quick look at this video from Keller and Dara, and then we’ll chat about it.

We’re super excited to have Dara here today. We’re announcing a partnership between Uber and Zipline—

That involves an investment and, more than that, a partnership for Zipline to power home delivery of hopefully 1 million and then more Uber Eats deliveries to your home— incredibly quickly, incredibly delightfully.

That’s 1 million deliveries a day. Amazing. Gentlemen, who wants to jump in first on this?

I’ll maybe just comment. I think this is a clever and also inevitable move by Dara and, more generally, by Uber. Dara has taken Uber, with a number of strategic acquisitions over the past few years, in the direction of being a mobility aggregator.

And thus far, Uber has, other than maybe Uber Air Taxi-type initiatives, basically been focused on ground-based mobility. I think this represents a serious move in the direction of aerial mobility. Of course, China has had this now for at least a couple of years, with the ubiquity of air-based drone delivery of foodstuffs and other matters. I think this is a very positive move for the West.

I think the elephant in the room, from Uber's perspective, is if you think back to when Uber basically hollowed out Carnegie Mellon University's robotics department in order to try to build up its in-house robotics capabilities. That was more or less a disaster and didn't quite work out, and there were lawsuits with Waymo and otherwise. I call this Uber's Mobility Plus Autonomy 2.0 strategy, where Uber focuses on being an aggregator at the software layer, a platform.

An aggregator in particular, right? So third parties, including, by the way, Waymo, are providing all of the physical autonomy, and Uber is just the demand aggregator for everyone to consume mobility from all of these different third-party providers. I think that works really well for Uber, as long as it maintains healthy competition among all of its mobility suppliers. It's bad for Uber if the industry verticalizes, and if Waymo or Zipline and someone else just decides, “We don't need Uber as an aggregator. We'll just do an end run and sell directly to the customers.”

You know what I think is incredibly cool—just incredibly cool? If you drive down any street in America, any suburban street, any urban street, it goes fast food, car dealer, fast food, car dealer, fast food, car dealer. In the very near future, the food will be off the main street and it'll just pop over the mountain and drop on your lap, and the car will drive to you. It's going to be so nice.

Yeah, it'll totally reshape things. This was very exo, by the way, because you've got Uber, as we've said, aggregating demand, and then Zipline gives you all the autonomous assets.

I think Alex's point is really important about what happens if they try to control it. What we heard from Dara last year on stage was that he's planning on creating as many partnerships as possible and becoming kind of like that wiring, and I think that is a smart play.

I'll make one forward-looking prediction here, right? This is something we probably could have seen coming. Let's bridge forward. Imagine if they now do a partnership with Shopify and every small merchant gets Amazon-grade logistics capability. That will change everything.

I'll invert your prediction, Salim, because Amazon obviously has its own in-house drone-delivery capability.

Which has been delayed for about 3 years. It's crazy.

For regulatory reasons, is my understanding—not because there's something technically wrong about it, just because this is new for the West, at least. Do you think Zipline ends up being a highly appetizing acquisition target for, say, Shopify, to in-house its delivery capabilities against Amazon?

Yes, yes, that's the answer.

Yes, I think so too.

Yeah. Great thought. Great thought. And this is, by the way, the hot news for the past day: the video going around social media of a woman watching in horror as one of these drones delivers her package into her swimming pool, and it sinks.

You know, we just had back-to-back on this podcast—Etched and Zipline are both companies where, on founding day, you're like, “Really? There's no way the amount of moving parts required for that to work.” And now you're looking at $20 billion and whatever.

No, I just want to say this whole delivery by drone—we had a Singularity University project in 2010 that did this, right? They looked at Africa and realized that Africa leapfrogged the entire landline and went to 1 billion mobile handsets. Why would you spend $1 trillion putting roads across Africa? Just go straight to drone delivery. They demoed that, and that apparently inspired Amazon and, I think, cascaded down to a lot of the others here.

And the backstory here is that Keller Rinaudo Cliffton's San Francisco-based company began operations in Africa because they were able to take advantage of regulatory arbitrage. The country wanted them there, and they developed operations and safety and then came back to the U.S.

Yeah. What Rwanda did, where a lot of the startups were, was they basically said, “There's a three-dimensional tube across the country, like a superhighway. If you keep your drone in that three-dimensional corridor, you can do whatever you want.” That allowed people to really go and play with things. Really, really big breakthrough.

12. Personalized mRNA Cancer Vaccines

Yeah. All right. I'm going to move us to our final segment on health, a really important one for everybody. Health is your new wealth. So, 3 exciting stories. The first story, perhaps the most significant, is Moderna and Merck announcing that their mRNA cancer vaccine succeeded in a late-stage melanoma trial, marking the first Phase 3 validation of personalized mRNA immunotherapy.

More than 8,500 people in the United States are expected to die from this deadly skin cancer this year alone. This is the cutting edge of science, right? The vaccine works by sequencing a patient's tumor mutations, identifying neoantigens and unique antigens for that cancer, and then manufacturing a custom mRNA vaccine that trains the patient's immune system to attack the tumor.

Let me unpack this a little bit. The first thing you're doing is a surgical resection of that tumor. You grab tissue, you do whole-exome and RNA sequencing, and you feed that into a machine-learning model that's looking for unique antigens. It ranks them—here's the most unique surface antigen—and creates mRNA encoding up to 34 patient-specific antigen targets. Then it's manufactured and shipped in only 8 weeks.

Every mRNA vaccine developed using machine learning creates a unique mRNA sequence for every patient. So, the vaccine for you is not the same as the vaccine for me. The Phase 2 results reduced the risk of recurrence or death by 49% and distant metastasis or death by 59% over 5 years. It's expected that the cost of this treatment will be as low as $5,000. Moderna stock surged 110% after announcing these results. Alex, I'm going to go to you first on this.

So many thoughts on this. The superficial thought is obviously a great day for cancer survivors and for treating cancer in general. That's the superficial thought. I'll go a level deeper.

First of all, I want to browbeat Moderna and Merck just a little bit for naming this drug. The drug's official name is—I'll see if I get this right—Intismeran, I think is how it's pronounced. In doing research, it turns out that “istismar” in Turkish is the word for exploitation or abuse. So, pro tip to Moderna and Merck: Please, before the rest of the world figures out what the name of this drug is, rename it from Intismeran to something that works well in every time zone.

Super-duper.

Yeah. Seriously. But these are FDA-approved names. It's crazy how they name this stuff. They don't care about Turkey, I guess.

They don't care about having a neologism roll off your tongue onto the floor either.

That's right. But on a more serious note, I remember the National Nanotechnology Initiative in the early 2000s, when Eric Drexler et al. sold the U.S. Congress on spending billions of dollars on nanotech, going back to The Diamond Age, on this thesis that we would have nanobots going through the human bloodstream, zapping cancer cells.

Well, guess what? It's 2026, and we caught up with the future. They're not diamondoid nanobots. They're lipid nanoparticles with mRNA snippets—34 different mRNA sequences. They're like soft robots. They're not these machine-phase, Drexlerian nanobots. But nonetheless, these are primitive nanobots.

For the first time, this is the first successful Phase 3 success for an mRNA cancer vaccine. This is the first, but not the last. There are going to be so many of these. All you need to do is look at Moderna's pipeline, which I think they do an extraordinary job of maintaining—a public pipeline website where you can see the clinical stage of every one of their vaccines for infectious diseases, some rare diseases, cancers, and other classes of diseases.

This is a general-purpose platform. This is arguably what we wanted 20 years ago out of nanobots. It's just that they're soft and they're made of fat. They're not made out of hard diamond stuff.

That's one point. The second point I just want to highlight is that there's a technology underneath this that I think is going wildly underpublicized, which is the RNA-sequencing technology that's enabling this to be personalized.

Peter, you touched on the first half of this, which is the RNA sequencing and mRNA sequencing of the tumor. But in order to calibrate what the right expression profile is for the tumor, you also need a second mRNA sequencing profile from the bloodstream to know what's abnormally expressed in the tumor and what isn't. That's from another company called Personalis that has what they call their NeXT Personal sequencing technology, which they originally developed, in my understanding, to do blood-based trace cancer detection.

So, if we ask the question, how does this all relate to liquid biopsy? If we extrapolate out a few years, maybe, for personalized cancer therapy, we won't even need the tumor itself. Maybe we won't even need to sequence the tumor itself. Maybe we'll just get all of this from the bloodstream and be able to do continuous medical monitoring via these models.

Yeah, fully agree. And two things: I was going to say a quick congratulations to a friend, Stéphane Bancel, who's the CEO of Moderna. They got a lot of negative news on the COVID vaccines, even though they came out with the vaccine very rapidly. The work that Moderna is doing on personalized cancer vaccines is amazing.

They're also building out the ability—there's a lot of endemic CMV and Epstein-Barr virus out there in the world population—for you to actually fight those infections internally to yourself. So, a lot of headroom for Moderna here as they dive in across the board and use this technology.

Two things that struck me. One is the regulatory structure that's allowing for personalization. That's a huge thing. We've never been able to do that before, so that opens up the floodgates for all sorts of things.

Daniel Kraft talks a lot about how we're getting to personalized medicine and the fact that we can navigate regulation. How do you deal with a sample size of N of 1, right? And the second thing that struck me was Raymond McCauley, who years ago said these mRNA vaccines—I think he would put it—are the first battle in the last war against all disease. And you're like, wow. I love to see this fruition coming.

Dave or Emad, do you want to hop in?

Well, this one strikes close to home for me because my daughter works at Moderna, and she's my go-to on this. She's been telling me and sending me research reports for months. This is not a secret. The stock went up yesterday; it almost tripled yesterday—the biggest one-day pop in any S&P 500 company of all time.

Wow.

By a wide margin. It was just massive. And so one thing immediately came to mind: well, I should have bought the stock. I should listen to your own daughter. That's advice number 1.

But number 2 is, remember when Enron and Tyco had all those fraud issues? They passed the Sarbanes-Oxley Act and a bunch of other laws. One of the byproducts of those laws is that a Wall Street analyst can't trade the stocks that they cover.

Everyone I know who was in that job is like, "Why would I study Moderna or other super high-tech stuff, learn all about it, and then not be able to trade?" So they all quit. The byproduct of that is that the stock market is now dominated by tech, which is very complicated to understand, but the research community is the worst I've ever seen in Wall Street history. And not only that, the indexes have taken over half the market. They don't think at all.

And so, the amount of useful information is at an all-time low when the things that need to be explained and understood are at an all-time high. But it was no secret that this Moderna platform can basically be used for any form of cancer and that it's highly likely to work. It's just a question of time. The research is all out there. This wasn't some kind of insider surprise. Any good analyst studying this would have seen this coming.

Yeah. I think this is the interesting thing: I don't think any of us are surprised by this result, and you won't be surprised when other ones go through. But our current regulatory regime means they'll have to go through the same process over and over and over again, when really, screw cancer—let's actually think about this from first principles. When we have systems like this that are very targeted, let's upgrade the regulations so this can actually get out to people faster to save their lives.

13. Simulating Human Cells with AI

Well, our next story is going to take us there, right? Because if you can simulate all of this in silico on a GPU cluster and say, "Yep, it's safe. Let it go," then let me turn to that story. It's one I've been excited about and tracking. I know, Alex, you have as well.

So our final story here is about AIDO.ai—A-I-D-O is how it's spelled—a general-purpose cell simulator that maintains cellular state, accepts interventions, and predicts multimodal biological outcomes. The goal: make experiments computable before the run in the lab. Cell simulation could reduce wet-lab experiments a thousandfold. If AIDO can predict which experiments will work, instead of testing 10,000 compounds in a wet lab, you can simulate them digitally, again in silico, and test only the top 10 that the simulator says are going to work. This is going to drop the cost by orders of magnitude.

Let's watch a quick video here, and then we'll go to the conversation.

Traditionally, biologists have relied on lab experiments and in vitro models to understand how cells behave. Now they can use AIDO, GenBio AI's virtual cell world model, to simulate the same biology in silico, combining multimodal, multiscale detail with sequential experimentation in ways no microscope or wet-lab assay could achieve alone.

At its core, the AIDO platform is a rich, stateful simulation environment powered by the first world model of a human cell—one that predicts what happens at every level and remembers every change you make along the way, from DNA and RNA to protein interactions, structures, and localization, to the whole cell, including Cell Painting morphology.

The model simulates cellular responses to genetic perturbations and treatments with small or large molecules. These can be layered in a sequence, providing the ability to watch the cell's full multiomic response with each step. The kind of insight that could mean computationally testing new drugs designed in cellular context before they ever reach the lab bench.

AIDO can be easily adapted to new cell types and indications. Using your own data, you can build models tailored to your research questions to simulate different cell types—from shape down to genes and their protein structures, where they end up in the cell, what they interact with, and how they affect cellular responses.

Wow. This is the foothills of longevity escape velocity. I've been waiting for this forever. Alex—

Medicine is cooked. This is what—if people like catchphrases, read my lips—medicine is cooked. This is what the end of medicine looks like. It looks like a virtual cell and the beginning of longevity escape velocity.

Putting that aside, I actually think we can get probably to LEV without solving all of medicine. My bet is it'll probably be a class of molecules, maybe like fourth- or fifth-generation GLP-1s, that get us to LEV. I think this is actually a superset of getting us to longevity escape velocity.

I think this is how we cure all disease everywhere. The way we do it is we build a virtual cell. It's just like we didn't actually have to solve human intelligence to solve AGI. It turns out you can just get AGI from compressing general human knowledge. It's not that hard in principle.

Similarly, I think this is how we solve all disease. It's not that hard, in principle. You simply train the world's best foundation model to model all cell states and all interventions against cells. And then you do an AlphaZero-type tree search against possible interventions to discover how to steer a virtual cell state from a diseased state to a healthy state, and then generalize that to tissues and organisms, and boom, you've solved all human disease. I think that's the endgame.

This is hyperpersonalized for you, right? You insert your DNA sequence, your current blood chemistries, and all of that, and there's an in silico model of your biology. It will tell you whether this drug works for you or doesn't.

Salim, your thoughts?

Yes. But, on that, I also don't want to overromanticize the personalized aspect. An ideal virtual cell is as personalized for you as, say, if you feed a quote-unquote personalized prompt to ChatGPT, the output is personalized for you. Well, yeah, superficially, it's a function of the inputs, but actually it's a generalist model.

Well, this has been a trend. I'll go back to the biotech stuff, right? We've been turning biology into information. When you turn something into information, it hops on the exponential curve. And we're seeing this play out live.

Each of us has, what, 50 trillion cells in the human body, roughly? Essentially, when you can model that, a human being becomes a software engineering problem, and we have really good techniques to navigate software—read, write, understand, et cetera.

And the phenomenal thing for me is we've done a good job in reading. You have reading, writing, and comprehension when you're trying to learn a new language—in this case, the language of biology. We've done a pretty good job of reading. We've started to do writing with CRISPR and now these mRNA vaccines, et cetera.

This gives us huge depth on the comprehension side through the digital twinning that can take place. So, holy crap, this opens up the door. I would go with Alex’s comment that medical is cooked. Emad.

Yeah, this is my hope for what the Genesis Project would be like. It seems very straightforward to me now that if we had a Manhattan Project to cure disease through in silico, massive whole-human-body models, cell models, and the organization of all our collective knowledge on cancer, autism, and all these other things, we would definitely get a result. I’m not even doing that.

It’s not going to be a government program. The labs are going to do that for us. But I’m thinking, why don’t we actually get together and get governments to put into place a Manhattan Project-type effort and just have a straight shot at it, making all the data open? Your point is well taken. At least the government is sitting on a lot of data, and the government could externalize all of it to private labs like CZI, IDO cell, and others that are all building foundation models. Make it a public good that they can all train off of—a Common Crawl.

We’ve done that in the UK, where you have access to this. Every government should follow suit. This fits with what we talked about earlier with Anthropic. Again, what are they going to apply their computation to? They will build a human-cell model. They will build a whole-body model, but I’d prefer for those to be public goods. Let’s cure cancer, address all these negative things, and understand the body like never before.

Again, that’s much better than building an atom bomb, because it will have the biggest impact on humanity ever. Mhm. Dave—

Well, for anyone listening who’s not a biotechnologist and thinking, “I’m not going to build a full-cell simulator. I have no idea how to do that,” you’re thinking about it the wrong way. You heard earlier on the pod that we’re looking at 10,000× expansion in AI, with some other innovations that could make it more like 1,000,000×. It’s totally data-starved.

The full-cell simulator is a way for AI to design thousands or hundreds of thousands of experiments and get reasonably good test results back through a simulation, rather than having to run millions of assays. That also applies in all kinds of other areas. Every investment we’ve made in a company like Mercor or Macato[?] that is wrestling with new types of data to feed the AI is thriving, growing, and making money and gaining valuation. You know, Mercor is worth $40 billion—or $20 billion now, $40 billion by the end of the year. They’re just absolutely killing it. Every field of endeavor is going to be data-starved, so you probably know a field that needs to supply data back to the great AI. The full-cell simulator is just the perfect biotech solution, but every industry has a solution.

Yeah.

And so compelling, too. It’s probably worth noting that this is from a company co-founded by David Baker, who shared the 2024 Nobel Prize in Chemistry with Demis for solving protein folding. This is arguably the next big thing, the next grand challenge in biology and medicine, now that structural biology has arguably been solved. Solve whole-cell simulation, and then you’re halfway to solving all disease.

Gentlemen, Salim and I have an AMA with the abundance community in 8 minutes, so I’m going to suggest we speedrun the AMA, if you don’t mind.

We should do the AMA in 8 minutes. We’re going to do the answers. Answers.

Yes. Okay.

Warm up.

All right. Emad, you pick one first.

Can’t we keep making frontier models more energy-efficient instead of building all this new power by 2075?

Yep. Necessity is the mother of innovation. As we run out of energy and as we run out of RAM, you’re going to optimize immensely, and it’ll be a big boon for everyone. Salim—

Will we see an XPRIZE aimed at solving the electricity-supply problem? We actually proposed, at the last Visioneers—or a couple of Visioneers ago—a way to provide enough off-grid energy storage to keep a village or town energy-sufficient for 3 days. But it looks like the market will take care of that. Yeah.

And regulation is the problem. So it doesn’t really serve as an enterprise where you need a huge technological breakthrough. This is better understood as a regulatory issue and a market issue. Dave, over to you.

I’ll take 4. If China has the advantage on power and the US has the advantage on chips, who actually has the real AI advantage?

Definitely chips. Power is a problem, but we need 100 gigawatts by the end of the decade. We already manufacture a terawatt in the US, so 10% of power will go to AI by the end of the decade. We’ll get that far, then we’ll be really desperate for more power. But between here and there, it’s all about chips. Every chip—that’s why memory is up 5×. So that’s the bigger advantage in the short run.

Yeah. Alex, number 2 is for you.

Number 2 asks, “Could interconnected microgrids popping up everywhere reduce the load on the main grid enough to matter?”

I think I would invert the question, invert the premise of the question. It’s not that the load on the main grid is going to be reduced; it’s the exact opposite. There’s so much economic demand for compute, and compute needs so much energy, that before long, unless we fully externalize all of the compute to orbit and the Dyson swarm, these data centers are going to be generating a surplus of energy that can be pushed back onto the grid and drive utility prices negative.

Yes. I want to make that point. If you’re arguing against a data center in your backyard, you’re arguing against lower-cost energy and economic advantages for your community. Please understand that.

Okay, so I’ll take question number 8. At the current rate of improvement, how long until AI is more efficient per watt than the human brain?

I think we’re probably already there, so that’s a hot take on this one. People have this fetishization of the Landauer limit, thinking, “We must really be far away in efficiency per watt from what biology is able to accomplish.” Biology is actually wildly inefficient. Our whole organism, and mammals in general, were never optimized for compute. Whereas silicon and CMOS—and whatever comes after CMOS, maybe it’ll be photonics, which I know Dave is interested in—were optimized and designed from scratch for compute. I don’t think the human brain is as efficient as many think.

The argument can be made that, if you look at a per-watt-per-task or watt-hours-per-task basis, leading-edge GPUs may already be more efficient than human brains.

Especially if you take into account the amount of energy required to train up a human over the course of 20 years, right?

Yeah. Lifetime total cost of ownership, as it were.

I’ll put a pin and a corollary to that, too. People use the power difference as a way to explain that AI thinking is very, very different from human thinking. I think people will soon realize that it’s not that different at all. The power difference will go away very quickly. But this idea that “that’s why it’s nothing like us” will also gradually go away.

I will take number 5. What’s actually driving the rush? Why not keep energy growth at past levels and accept slower growth from its WHD[?] and DRN[?]?

I think the easy thing here is that intelligence is looking more and more like a general-purpose input that will drive economic growth, right? Saying, “Let’s have less intelligence,” is like saying, “Let’s have less electricity or less internet.” You want more of it. It improves research, drug discovery—we saw all of that today—logistics, everything. So you want as much of it as possible.

There’s also competitive pressure. If one company or country slows down, others won’t. So it’s not about accepting that capacity. You’ve got to get your head around the abundance idea. The one thing we should be doing is accelerating energy abundance.

All right, Dave.

I’ll take the easy one, number 7. Whatever happened to fuel cells?

Actually, Elon Musk’s passion in life was ultracapacitors originally.

That was his Stanford thesis before he dropped out.

Well, whatever happened to those, too? What happened is that lithium batteries worked far, far better than anyone ever would have predicted, and they’re still improving. So it sucked all the capital out of the other ideas. That’s all that happened.

All right, Emad, close us out here.

Yeah. If 71% of Americans oppose data centers, is this grid buildout actually going to help regular people’s electricity, or just make it scarcer and more expensive?

David Harmon, number 3. I mean, it’s what you just said, Peter, right? It’s going to make your electricity cheaper. More power is good. These things are not polluting. We need more data centers, we need more power, and we need to make sure it’s all built right.

Amazing. You guys did it. You did it in 8 minutes. Salim, we’ve got a whole 60 seconds to get over to our abundance community.

I’m going to be 2 minutes late. I’ve got to get a little bit of food. In the meantime, gentlemen—Jesus Christ.

I love you all so much. This was such a fun day.

Just so much fun and brilliance. You guys are amazing. I’m so proud to have you as our Moonshot mates. Alex, Dave, Emad, Salim, thank you, always. Thank you to our listeners. We love having you, and hopefully you find this to be a way of keeping up with what’s going on in the world, because we are in an accelerating singularity. There’s no time to sleep, no time to blink.

Don’t get fatigued like I did.

Yeah.

Take care, guys. Be well.

All right. Thanks, Peter.