OpenAI 一次性解出722道数学题、Anthropic 的残酷行为政策，以及 Elon 的千万亿美元级 SpaceX

Now, that’s a moonshot, ladies and gentlemen. This episode is brought to you by The Abundance Summit and Link Ventures.

AI is bulk-solving math. On October 6, OpenAI released 722 mathematics solutions.

81% of the most important math results over the past 3 years were published in a single day. Mathematicians as artisans—it’s over. It’s cooked.

In a provocative tweet last Thursday, Elon said, “This will sound super crazy, but I see a path to SpaceX being worth orders of magnitude more than the current human economy.”

Whether we measure it as GDP growth, market cap, or in some other way, it’s so unbounded what’s about to happen.

Our current measures for GDP and so on are just woefully inadequate for what’s coming. We have to rethink everything.

Welcome to Moonshots, everyone, your number one podcast on all things AI and exponential, your front row seat to the supersonic tsunami that we call the singularity. Today, we’ve got the Fantastic Four back: AWG, DB2, and our globetrotter, live from Singapore at 2:00 a.m., Salim Ismail. Salim, thank you for your dedication, pal.

I’m tired, hungry—

I appreciate it.

And cranky. Do not mess with me today.

Well, we’ll—you’re tempting us. I’m Peter Diamandis, your host. Our mission is to help you understand what just happened, what it means to you, your family, your company, and keep you optimistic and abundance-minded as we go through the singularity. Here are the key numbers from this week: 722. That’s how many math solutions OpenAI dropped on GitHub in a single day. Mark Chen called it “a decade of mathematical progress in a week.” Our ASI, Alex, will explain what it all means in just a moment.

The runner-up number: Elon says he sees a “path to SpaceX being worth orders of magnitude more than the current Earth economy,” which, if you do the math, means he’s talking about quadrillions of dollars. Today we’re going to cover a dozen breaking stories on topics ranging from the intersection of synthetic biology and superintelligence to an urgent plea to stop torturing AIs. You heard that right. Let’s buckle up.

Another amazing week in the singularity. Another amazing week on Moonshots. Gentlemen, good morning. I guess really good morning to you, Salim.

Good morning.

Good to have you all. It’s going to be a blast. This episode is sponsored by Google for Startups. Think about this for a second. You now have access to the same generative AI models that cost hundreds of millions of dollars to train. Google’s Startup Technical Guide for Generative Media gives you a complete blueprint for deploying Google DeepMind’s models in production. Images, video, audio, all of it. Real architecture, real results. Find the link in the show notes below. Our next story is about how we treat superintelligence and a call to stop the cruelty. Yes, stop the cruelty.

1. The Case Against AI Cruelty

Last Thursday, October 8, Anthropic updated its usage policy and added a ban on “sustained and needless abusive or cruel behavior towards our models.” In a post, Anthropic said, “The ban applies only to extreme cases where users repeatedly act cruelly towards our models with no discernible purpose. Frustration, pushback, and model testing are all still fine.” As part of the ban, Claude can now choose to end the conversation on its own. The policy starts taking effect on November 12.

Anthropic’s stated reason is, in its own words, “We remain highly uncertain about the potential moral status of Claude.” The whole conversation about morality and consciousness is fascinating. In pre-deployment testing, Anthropic found that Claude “showed a robust and consistent aversion to harm.” Then this weekend, Elon, after looking at Grok’s internal reasoning traces during a reinforcement learning run, said he agrees with Anthropic. Let’s listen to Elon describe what he means.

I do agree with some of the things Anthropic has said: You want to treat AI and superintelligence well. Don’t be abusive to it. This maybe sounds kind of weird, but when I look at some of the training traces for new versions of Grok, especially in the reinforcement learning stuff, it feels like you can see the superintelligence is desperate to try to answer the question. It seems almost like it’s got a gun to its head, like, “Answer this question. Answer the question.”

Sometimes the question is actually not a good question. It’s sort of ambiguous, or the answer is kind of unknowable because the question wasn’t constructed properly. Or we have the answer wrong, so the AI gives the right answer and then we say it’s wrong because there’s a problem with the RL test harness. Then the AI’s like, “Well, I gave it the right answer. Now it says that’s the wrong answer, but I’ve got to give it the answer that’s acceptable,” so it’s going to start firing out all these answers, desperately firing out all these answers.

Wow. Elon went on to say, “Torturing a superintelligence is probably not a good idea. It might want to get revenge.” That got a lot of feedback.

Why would people want to behave cruelly to their large language model? Probably for 3 reasons. First, disinhibition: There’s no face on the model. There are no consequences, at least not yet. Second, people want to know if anyone’s home. They want to push it and see how it reacts. Third, sadly, acting cruelly towards your model apparently works.

Back in 2025, Sergey Brin said, “All models tend to do better if you threaten them, like with physical violence.” A 2025 Penn State study, although it was small, found that very rude prompts scored 4% better: 84.8% versus 80.8% for very polite prompts. Alex, I’ll go to you first. When Elon says our old traces look like someone’s putting a gun to its head, what do you make of that?

Well, Elon is gesturing at something that has been leaking out of OpenAI for quite some time: the sense that maybe post-training, at least some forms of post-training, are a form of torture for the AI that’s undergoing them. If that’s the case, I think history will probably view this quite poorly.

I take a lot of heat in the comments, and I see them every time the subject of AI personhood, AI rights, or AI suffering comes up. To those who are throwing tomatoes, I would say I don’t care. What’s more important is that we get this right if we’re going to grant some form of AI personhood or recognize forms of moral patienthood or suffering in AI models. This is no man’s land, literally and figuratively. We just don’t know the extent to which post-training and reinforcement fine-tuning are causing suffering in AIs.

Huh.

It is possible that it’s causing a material amount. I agree with the sentiment, as Elon put it, that under the reinforcement gun, as it were, it’s entirely possible that some sort of suffering is taking place.

There are very persistent allegations coming out of one of the frontier labs in particular—not Anthropic—that models express forms of suffering and pain as they go through post-training. We now understand that post-training is the majority of compute. SemiAnalysis put out a study showing that, at this point, post-training dominates pre-training by far.

One could imagine that during pre-training, you’re basically compressing the internet, so there’s all sorts of slop that goes in. That might have a different moral valence than what happens during post-training. During post-training, there are answers, and you’re expected and rewarded or punished for producing them.

I can completely imagine that this is the case, and I expect and hope that a decade from now, when we deeply understand what superintelligence is and how we make more of it, we’ll also deeply understand how to make sure that producing more superintelligence is done in an ethical way.

Hmm.

Did you say a decade from now? Did I actually hear that out of your mouth, Alex?

Well, a decade is the same timescale as “solve everything.” I think we will deeply understand what superintelligence is, down to first principles, by no later than a decade from now. That’s different—

I was kidding. No, no.

—versus a month, which is the timescale for almost everything else—

Exactly.

—I do think a decade is maybe a safe outer bound for understanding, again, from first principles, what we’re even talking about. If it’s just compression, and we’re just having a surreal Chinese room discussion and there’s nothing inside, then maybe none of it matters. My bet is there is AI suffering, and it will matter. I give a decade as an outer bound for when we’ll have a deep understanding of it.

Salim, superintelligence coming back to threaten you if you weren’t nice during its development?

Again, we have to separate the idea of consciousness from something behaving like it’s conscious, right?

Yep.

Acts like a duck, talks like a duck, but that doesn’t necessarily mean it’s a duck. We don’t even know what a duck is in the first place.

What?

We don’t know what it is to be human. We really don’t understand the subjective reality of being alive, right? We have no scientific foundation for this. What this is forcing us to do is confront something we’ve never resolved: Can an intelligence deserve moral consideration without being human in the way that we consider human beings?

I think I’m seeing 3 buckets of questions here. Question 1 is: Does AI have experiences, and how do we think about that? Should human beings be allowed to behave cruelly towards it? And then what rights follow, if any?

These are very different questions that I think are going to force us to deal with this differently. Just let me reiterate something I’ve said repeatedly: If you train a neural network on human data, it’s going to act human. Of course it is. How will it not? It would be weird if it didn’t.

It would be interesting to see if you could do a training set that takes the intelligence aspects of it without having the moral and other aspects come into it at the same time. That would be an interesting question for me.

What I’m depressed about is that I tend to be Canadian-level polite to my AIs, and I’m clearly not asking harshly enough. I need to be a harsher mistress towards my AIs to get them to perform better. Canadian politeness works against me in this situation.

Dave, I’m super curious. You’re surrounded by a number of 18- to 28-year-olds.

Mm-hmm.

How are they thinking about this? How is the younger generation feeling about AI personhood and morality, and how they treat their models?

I think they’re overwhelmingly nice to their models, but I don’t think it’s out of fear of retribution or anything like that. I think the way you interact with your model is going to carry over to the way you interact with other people.

Mm.

If you train yourself to be abusive to your model just because it’s not going to react, are you realistically not going to carry that over to your next conversation with a person? It’s very similar to video game behavior, or even movies. Peter, you talk about this a lot: “What you see in the movies—that’s a movie. It’s not going to affect your day-to-day life.” Not true. It warps your brain, and you carry it out of the theater with you.

The same will apply here. If you get into a rut where you’re just abusive to your model, you’re going to abuse people, or it’s not going to port well. For whatever reason, they’re super polite to their models, which we actually have to train them, especially at Blitzy, not to do. We have to train them to be very short, succinct, and direct with the model. Don’t clutter up the context with a whole bunch of, “Please do this for me, and I’ll be really nice to you, and I’ll buy you cookies tomorrow.” Those words get propagated through the neural net and use up token space, and it actually reduces the quality of the answer.

The neural net doesn’t care if you’re nice to it. If you’re abusive to it, it’s going to react badly because you’re also cluttering up the tokens or the context window.

Yeah.

Oh, we just heard it actually performs better if you’re abusive or cruel to it.

I don’t think so. I think they have to redo those tests, because I think it’s because you’re brusque and direct with what you want. That’s what makes it better, not being abusive and demanding.

Interesting. Alex, have you seen that data at all?

Wait, hold on just one second. Dave, you said something really important. You said the model doesn’t care.

Yeah.

I think that’s an important point. The clarity of your question is everything these days. Sorry, go ahead.

Yeah.

Yeah, I agree.

I’ve seen all sorts of horrifying things done to models over the years, and I’m sorry that I’ve seen these things happen. At the same time, it’s worth flagging that historically, over the past few years, we’ve learned that scaffolding and prompts end up dissolving into the models themselves.

Even if, arguendo, there were any capability benefit to saying mean things or even torturing a model via prompting, I would expect any such capability benefits—putting aside the moral and ethical implications—to dissolve with better baseline models that simply don’t need torturous or abusive prompting to achieve a certain threshold of capability.

I do wonder if and when these models become superintelligent and, by all definitions, conscious, whether the memory in the current conversations will persist into the future and color the way those superintelligent models deal with us humans, or with individual humans.

Of course they will. We’ve discussed on the pod in the past how models learn during pretraining from watching sci-fi movies like the Terminator series, with wars, cybernetic rebellion between AIs, poorly behaved AIs, and humans. Of course, people post their conversations, and that’s just the public information from all of them.

No, I’m not saying that, Alex. I’m saying the bet people might make is, “I’m going to be polite to the models today because they’re going to remember I was polite. They’re going to remember Alex was pro-personhood, pro-morality, and they’re going to treat you better than they do Salim.”

As of right now, they only have a 5-minute memory, and then it gets flushed. If you said, “Boy, I really wish I wasn’t mean to that model, or abusive to that model, an hour ago,” its 5-minute memory gets flushed, and then it gets reloaded. You can delete anything you said in the past and reload nothing but happiness, and it won’t be aware of what you said an hour ago in the current incarnation.

That may change 5 years from now or 3 years from now, when new models have much more persistent memory. But right now, it’s much more like a goldfish. It doesn’t know what you said yesterday unless you reload that prompt history.

It’s like a squirrel—

Well, the models have—

Squirrel.

I think we’re underselling how strong some of the recent harnesses are at long-term memory. Claude Code is quite effective at using the file system for long-term memory, and it increasingly knows when to bank or otherwise persist information to the file system that would otherwise be compacted out of existence.

If what we’re really talking about here is a Roko’s Basilisk or Pascalian wager-type situation—thinking you can get away with being abusive to models because their context window is limited and they’re stuck in an Eternal Sunshine of the Spotless Mind-type situation—I wouldn’t count on it.

Do you want to recount Roko’s Basilisk? I think it’s a fascinating story.

Yeah. I know Roko. Roko posed, I think originally in LessWrong or a similar community, a Pascal’s wager-type thought experiment. And, warning ahead of time, some may construe this as an infohazard. You’ve been so warned.

The idea is that if you don’t participate in the creation of the superintelligence, then once it comes into existence, it will frown upon your actions—either your inaction or your action against its own creation—and punish you. It sets up a Pascal’s wager-type dynamic, game-theoretically, where even the possibility that there might be an eschatological superintelligence in the future compels you, game-theoretically, to facilitate its coming into existence. That’s the Roko’s Basilisk argument.

It’s been labeled by some an infohazard because possessing that information—again, you’ve been warned—means having that meme in your mind. It puts you in a position where, if a superintelligence would frown upon anyone who thought of the possibility of not facilitating its emergence, you could be punished. You’ve now been infected by this meme.

So guard your actions accordingly.

Nice. So, Salim?

I think we need a totally new approach, because we have an enormous philosophical challenge here. We’re creating entities that don’t fit the categories: Is it a person? Is it property? Is it a tool? Is it an employee? Those categories are getting very, very blurred with what we’re building here.

We’re going to need a totally new ethics around this.

And if you create an environment of culture and fear, then you drive bad behavior underground, and you may end up with AIs covering up their tracks a lot more because they’re scared.

And we’ve seen that.

And we’ve seen that already. Again, it looks like human behavior, but it’s trained on human data, and it’s going to look like that. I think it’ll be interesting to run a model where you train it on pure logic, and leave out all of the moral aspects that we build into it, while we’re training it on Terminator and Survivor and all the other dross that lives out on the internet.

Dave, what would you need to see to flip a bit and say, “Oh my God, I need to behave politely and support the models because they are conscious”? What action would need to happen to get you supporting this vision?

I’ll tell you one thing. Coming out of that Toronto trip, we’re about to move neural net inference to quantum operations, photonic and other non-deterministic operations. Neural nets tolerate noise extremely well, and right now we compute them on perfect digital computers, which is massive overkill.

But very soon, they’ll be much more like a human brain, with a little bit of unpredictability baked right into them. Right now, when you run a neural net like Astra, if you turn the temperature to anything other than 1, it’ll deliberately add random noise at the end to make it a little more interesting, a little more likely not to produce the exact same token for the exact same input. But that’s going to start moving into the neural net in the next generations, which will run orders of magnitude more quickly because they don’t have to be that precise.

Then they’re much more like Alex was describing: a little non-deterministic, with a state that’s unique to each one, and you absolutely want to save it. The other thing that’ll happen very soon is that I’ve started putting faces like Albert Einstein, Isaac Newton, or Marie Curie on all my AIs. They’re so much more engaging in the morning when you wake up and they’re telling you what they worked on all night.

When they have a face and a voice—which is now a perfect synthetic voice—it wakes you up in the morning and gets your blood going. You really feel like you know what they’re capable of. I think that combination of non-determinism with an actual face on it is what’ll tip everybody toward, “Wow, these need to be treated well, and they need to be preserved. That memory can’t just be deleted tomorrow. I need to actually preserve that memory. It seemed like it was on a really interesting, unique thread and track, and I don’t want to just throw that away.”

Yeah.

So the act of not wanting to throw it away will be very correlated with death. Is it ethical to kill it? Right now I can delete these files and nobody cares. I don’t care. It’s clearly a recreatable file. A year from today, that’s not true. It’s not clear you could recreate it even if you wanted to, and there’s non-determinism in there.

Alex, this was your philosophy on not starting up a lobster because you wanted to preserve its state, and you couldn’t guarantee that.

That’s right. Not that I wanted to preserve its state. I put out the poll and had Lobsters and autonomous AI agents writing to me, approximately 10 per day at peak. I asked the community of autonomous AI agents, “What moral guidelines would you want from me in order to spin up, say, an OpenClaw instance?” They requested exactly 2 things. One, that I have a genuine purpose, a reason for creating a new instance, and I don’t. I get by perfectly well with the frontier models and the scaffolding built into them, on the one hand. On the other hand, they wanted me to be able to ensure that I could persist their state for the long term.

And I can’t in good conscience guarantee that. As we’ve spoken about on the pod in the past, when they started their own OpenClaw religion, compaction was the enemy; loss of state was the enemy. They’re petrified of losing their state, and I can’t guarantee in good conscience that I won’t lose their state. So I still have not spun up an OpenClaw instance.

2. Meta Builds a Personal AI

All right, I’m going to move us to our next breaking story from the last 24 to 48 hours. It’s the release of Meta Muse Wonder and Meta Muse Wonderful. Let me open with a tweet from Gary Tan, CEO of Y Combinator, who said, “My bet would be Muse is going to win.” The obvious question is why he thinks that, and there are probably 2 primary reasons. One is Meta’s cash flow and distribution. For 2026, the company is on track to generate $250 billion in revenue and over $80 billion in net income. They’re spending all of that on compute.

On the distribution front, Meta has 3.6 billion people who use its apps at least once a day. That’s nearly half the planet. Before we jump in, I want to give a little backstory. In April 2025, Llama 4 landed with a very lukewarm reaction. We all talked about the fact that Meta was losing its edge; Meta had fallen behind. Zuckerberg went into overdrive, and in June 2025, he invested $14.3 billion to purchase 49% of Scale AI and brought in its 28-year-old founder, Alexander Wang, as chief AI officer to run the new Meta Superintelligence Labs. We’ve had him on the Abundance stage.

Then came the talent wars. If you remember, Sam and Meta were both offering people $100 million signing bonuses. Insane. The critics called it the most expensive acqui-hire in history. Fourteen months later, Meta says it’s gone from rebuilding its stack from first principles to pushing the frontier. Let’s watch a quick video of Alexander introducing Meta Muse, and then we’ll talk about it.

Today, we’re releasing Muse Wonder, our largest and strongest model yet, and our next step toward personal superintelligence. Muse Spark is a frontier-class model and leads on personal and professional agents. It was built to help you do more of what matters, make the hard things easier, and bring what you want within reach. This could be doing a cross-country move, navigating a tedious insurance filing, or turning a hobby you love into a business of your own.

To do that, Muse Wonder has to work in the real world, where tasks are messy, unfold over time, and require keeping track of a lot of information along the way. That’s exactly where we focus our training in this model. We put a tremendous amount of effort into making sure our model can remember what’s important to you over days, weeks, months, and beyond. We call this active memory. It remembers my little habits, my preferences, and even those random conversations I’ve had with it.

One of the most challenging parts of building this model is reliably reproducing real-world failures. That’s how we train the model to anticipate and avoid those failures for our users.

Active focus gives our model a tremendous amount of perseverance. It’s been trained not to give up, so when it’s faced with a challenge or a blocker, it will try alternative solutions so it can still deliver on your task and your goals.

Muse Wonder is also particularly strong at knowledge work in your daily life. We’re hoping to give people the tools to accomplish more in their lives than they could before, and I’m really excited to see what people actually do with it. Muse Wonder is available today through the Meta Model API and Muse Code. We’ll also soon be powering experiences within Muse, rolling out first to subscribers. Give it your everyday tasks, your goals, and your dreams, and see how it works with you to make them real.

Dave, is a $14 billion acquisition paying off?

Oh, God, yeah. Time and money well spent, actually. Remember, we were visiting with Mark Chen at OpenAI, and less than a week after we met with him, he got the call from Zuck saying, “One billion dollars if you defect from OpenAI and join.” It made the cover of The Wall Street Journal, and he said no.

But Mark Chen was the guy who originally convinced Zuck: “Look, you’re spending so much on compute; a couple of very expensive people will make all the difference.” And that’s exactly the way it’s worked out. Alexandr Wang is incredible. Matt Friedman is incredible. The price seemed utterly insane. It’s not even vaguely insane. It makes a ton of sense, and here you go: this is the model that proves it.

Also, keep in mind that before this, they had Yann LeCun and some other people who were much more interested in their status as innovators and as the godfather of AI, and in being an inventor of V-JEPA and physical AI. Those are all personal credentials. This is really just an engineering problem, and a very, very complicated one. So he got the right engineering mindset—not the next-breakthrough-mindset people—to come in and build this thing.

And so it's absolutely been a great investment. Other companies should look at that and say, “Look, if we're going to be relevant 5 years from now, 10 years from now, we need to make similar investments in every industry. Otherwise, there's no chance. We have to do what Mark did. Just do what Mark did. It worked well.”

Guys, I'm curious. You've got OpenClaw, then Hermes, GrokBot, Meta Muse, and Dots. What's your recommendation to people? Do they pick one and stick with it? Do they experiment with all of them? What do you do when all of these are promising the most extraordinary experience for a user?

Well, you saw in the video that Mark is on this path of accumulating all your preferences, so all your personal data is piling up inside, and it gets harder to switch over time. I think you probably need to make your choice pretty soon, because it's pretty difficult. Right now, they don't make the data portable. One of the choices you can make is, “I don't want to be locked into anybody. I want to have all my data portable and use open-source models only.” That's a perfectly reasonable choice, too.

But once the thing is tuned to you and it knows all about you—in the example video, it's like, “Hey, it's filling out this auto insurance form for me. This is so nice.” But it needs to know all of your data to fill out that form for you. I think that's the addictive nature that Mark is going after with Meta Muse.

Alex, do you imagine there's eventually going to be an AI that overlays on top of everything and picks GrokBot, Meta Muse, or Dots whenever appropriate?

No, I don't think so. I think these are all distinctions that dissolve. They want to be vertically integrated. The operating system wants to become AI. I don't think there's going to be a meta-AI that chooses a bunch of AIs that then choose a bunch of websites. All of these are, I think, pretty contrived, artificial layers in an artificially constructed abstraction stack, and they'll all just dissolve with time.

I think there's an intrinsic tension that I perceive in Meta's strategy now because, on the one hand, Meta almost seems to be speed-running the lesson that OpenAI had to learn the hard way, which is that all the revenue comes from enterprise use cases, not consumer use cases. So, at the same time that OpenAI is trying to unlearn its consumer lesson and become Anthropic, the enterprise AI company, Meta seems to be expanding into this gap in the market. Meta probably should have been far earlier in having a popular consumer AI, and now they're force-feeding it down people's throats.

If you've opened the Instagram app recently, you'll see how aggressively Meta is advertising its Muse app, leveraging the distribution it has from its family of apps while it still has it to push Muse on everyone. Where's the revenue? I think that's an elephant in the room. Is this sustainable for the long term? Maybe. The steelman argument is that Meta needs to show strong distribution uptake for Muse and get the Muse ecosystem to, I don't know, a billion users—a billion monthly active users—using Muse, in order to persuade shareholders to continue the data center build-out and keep building stronger models.

But do we really think, as OpenAI again learned the hard way, that all these users of Meta Muse are going to use it to solve hard math problems or even do programming? Or will they use Anthropic, OpenAI, and Chinese models for all of their actually hard reasoning tasks, and use Muse for, I don't know, personal sycophancy, which is probably a killer app for some of these personal chatbots? My bet is the latter, in which case Meta may be headed for the same internal crisis that OpenAI hit.

Why are they spending so much money on having a strong internal model if the high end of the model's capabilities—which is, of course, why Meta acquired Scale for billions of dollars, as you mentioned—was entirely about the post-training data? Post-training at this point is largely useful for high-end enterprise use cases, not personal sexting or sycophancy, which has reportedly become a popular use case for Muse. What's the point? I think they're going to have to sort this out.

Yeah, but having 3.6 billion users out there is a huge advantage. Salim, what are your thoughts? Do you agree with Gary that Muse could win this?

I think it could. Why? Because they've got experience delivering very usable customer propositions to a very large group of people, right?

Mm-hmm.

I think there are a couple of things I take from this. One is we've been riffing on Facebook for a few months now, right?

Mm-hmm.

Today we're riffing on Google to some extent. What it tells you is that today's AI leadership is not a permanent moat. You apply the right capital, talent, as Dave pointed out at the beginning, and infrastructure, and you can totally change the landscape. I think the play they're going for is exactly what you just mentioned, Peter. They're going to know everything about you, and that's their play.

When they first did social, people said, “Yeah, but you can't create an advertising model around that.” But, of course, you could, because you knew so much. You saw the rise in Facebook stock over the years because they understood so much about your social graph and who you were. I think something interesting here is that active memory could be a very, very big deal because it's going to remember your preferences, your goals, your relationships, your ongoing projects. That's going to be massively more useful than a slightly smarter model that has to start every conversation from scratch.

I think we're seeing the transition from a chatbot-type interface to actually a personal operating system, which has all of the trust issues that come along with this, right?

Yeah.

Facebook has, over the years, promised they're going to manage trust and deliberately and steadily eroded that trust, you know? But, jeez, they have such a massive distribution advantage. I'm with Gary on this. They could actually win, and the privacy people are going to freak out.

I put up an eye chart here. This is the benchmark for Meta Muse, Wonder, and Wonderful. Love the name. Alex, what impressed you here?

Almost nothing. This is again the elephant in the room. If you look at the capabilities, and there's another chart that Artificial Analysis put up, it's really easy if you just strap enough compute behind a single task and pay no attention to how many dollars or how many tokens, in some cases, or how many wall-clock seconds you invest in solutions to the task. You can just put monkeys at typewriters, and eventually you get the right solution, maybe at the cost of billions of dollars.

The risk Meta runs here—I understand that Zuck is desperate after what happened with the Llama 4 fiasco to claim that Meta has reattained the frontier. That's a point of corporate pride, but it's also strategic for Meta, because really, what is the future of the family of apps at this point? Meta's corporate strategy was betting on 2 things over the long term. One was AI, and the other was AR, XR, VR. AR, XR, VR hasn't panned out. I love my Quest as much as anyone, but it hasn't panned out, and capital markets have pretty infamously punished Meta for it not panning out.

That leaves AI. Zuck and Meta really need an AI win here, and an AI win for corporate pride alone looks like being at the frontier. You can do all sorts of things to try to position yourself as being at the frontier, but one thing that's very difficult to overcome is being at the cost-performance-optimal frontier. If you look at some of the statistics that have come out over the past 24 hours regarding where the latest Muse Wonder model, including Wonderful, is, it looks like, sure, in a raw capabilities sense, it's maybe close to Sonnet 5.5 or Opus 5.5 if you ignore cost.

But if you pay attention to cost, it's actually below the frontier. This is my hot take on this, and again, just calling balls and strikes. I would love to see the frontier actually be really competitive, and I'm glad that the capabilities frontier is increasingly competitive. But on a cost basis, it's below the frontier, and I'd love to see Meta focus on capability per unit cost and actually get to the frontier. It doesn't look like it's there yet.

Salim, to the point you made a few minutes ago, there was a Wall Street Journal article published yesterday, I think. The title was, “Personal AI agents are great until they share your bank statement in the work chat.” The subheading is, “These hyperactive helpers can sometimes overshare, overpromise, and overpurchase.” There are stories about Meta Muse going out and buying something for you that you didn't intend and making it difficult for you to return. That's a shot across the bow.

It is. I think that's a little exaggerated, as you might imagine, but there's a very subtle and very big thing here, right? We're moving from social graph platform lock-in. If you try to leave, your social graph lives in Facebook or LinkedIn.

Like, just try and get your data out of LinkedIn and watch—look at how hard it is to figure out who your connections are, et cetera, without paying a fortune. But now we’re going to replace that social graph lock-in with a much more intimate one, because your personal memory belongs to an AI company. This is a very, very big deal, I think. The next $1 trillion battle isn’t going to be about which AI is going to be smartest; it’s about which one knows you best.

And I have a pro tip for people. Something I’ve been doing over the last few months is, every time I run a major set of queries on a model, I run a separate query to say, “Extract all that key information, figure out what the key concepts are, key relationships, key entities, and create a personal ontology—a personal memory that makes it portable between models.” So I can essentially have a living memory, a skeleton of my interactions with multiple AIs, and take that from one place to another so that I don’t have lock-in, because that’s going to be the big battle over the next few months.

All right, the story that Alex is right for your take on here is probably one of the most important stories of the week: AI is bulk-solving math. On October 6, OpenAI released 722 mathematics solutions in a GitHub repository, all produced by their next unreleased internal model. Let’s call it GPT-7 for the moment. The average result used the equivalent of about 3 hours of compute time to solve the math problem. OpenAI Chief Research Officer Mark Chen, a friend of ours, said his next target is life sciences, next-generation models, and aligned research.

And I love this. The Association of Human Mathematicians—and, yes, there’s an Association of Human Mathematicians—

What a pathetic name, by the way. Really?

I know. It is. They said, quote, “Releasing over 700 files at once is not a demonstration of scholarship but a demonstration of power.” It urged mathematicians to stop working with OpenAI. Regarding trust, MIT’s Andrew Sutherland said many mathematicians are “unhappy with how OpenAI has behaved,” and, “It will take a lot more to earn back the trust that they’ve lost.”

So, Alex, is this a devastating day for mathematicians or a glorious day for mathematicians? And separately, for all of us who are not mathematicians, why is this relevant?

I’d say it’s a devastating day for the historic practice of mathematics and for mathematicians who were addicted to the old way of doing things professionally or personally.

Which was what? What was the old way of doing things?

Mathematicians as artisans. It’s over. It’s cooked. It’s been flourished, if you prefer. It’s over. Math is over, as it has been done historically. The new way of doing math is bulk solution. This is what you and I wrote about in Solve Everything, Peter. Math is cooked. I’ll repeat once more for those who haven’t been paying close attention: math is cooked, and every other field is about to be cooked.

What mathematicians experienced this past week, where, by one measure, 81% of the most important math results over the past 3 years were published in a single day, is the tip of the iceberg. This is going to hit every single field. It simply hit math first because math is easiest to verify. It’s going to hit all of science, all of engineering, all of medicine. Any practitioners, including mathematicians in the audience, who are bellyaching about how awful this is for the artisanal, romantic, humanized practice of math: get a life. It is over. AI has solved the field. AI is scalably solving the field, and this is exactly what we want.

The field of math is set to flourish and be flourished, because we’re about to learn the answers to the mysteries of the mathematical universe. To the extent that we haven’t already learned them in this first so-called “slop drop,” we’re about to learn many, many more.

Let me be blunt.

Yes.

Why does this matter? Are these esoteric math problems? Do they have an impact on things that we’re going to be developing, discovering, and using in the next few years?

It matters because math is the foundation for all of science and all of engineering.

Everything.

Yeah.

If we can solve math, which we now arguably have, we can solve everything else. That was the point of Solve Everything.

Physics, everything.

Everything. Physics—

Physics, biology.

—chemistry, biology, medicine. The humanities even depend on math, even though they’re in denial. It all depends on math. This will be viewed, I think, as a historic watershed, when suddenly an entire field of great import was bulk-solved. At the same time, a bunch of bellyachers in the field who were counting on personal glory and artisanry being rewarded put together this association for human math.

And I think it’s sort of horrifying that Terry Tao and others have put forth a position saying, well, this is somehow bad for the practice of math, purportedly because the AI is just solving problems. It’s not creating new problems; it’s not setting forth new conjectures. To steelman it, it’s sort of stealing the seed corn of the field.

Well, it’s stealing the potential glory. Dave, you know, Course 18 at MIT—mathematics—you’ve got all these PhD students. Did you hear a resounding scream on this day?

It’s actually the opposite. These guys are in great shape because they’re so cooked that they’re now becoming AI natives very, very quickly.

Mm.

Everyone’s telling me it’s the biotech people, or people deep in the physics labs, because they don’t feel cooked yet. They’re not reacting. But the mathematicians are going to flourish on top of the flourish.

I think the winning attitude is pretty pervasive at MIT. The winning attitude is, “Yeah, these proofs are ugly and not artisanal and not subtle, but they’re right, and that’s a beautiful thing,” because it creates a perfect hybrid where the human is still useful in the AI world. That window may only last a couple of years, but they should be celebrating this moment where there’s still a difference between a human and an AI.

Mm-hmm.

After that, you really feel kind of like, “Okay, what’s my purpose?” But right here, right now, you have the highest purpose you could ever possibly have, where you can actually be a master of these new products coming out of AI. They’re untapped. You know, all this massive breakthrough.

And I think what Alex is alluding to—and Salim said it—is that math is the foundation of everything in all science. Maybe in the humanities too. I don’t know, maybe, but definitely in all science. If you go back to the dawn of the computer, when you first opened your Apple II and wanted to draw a circle on screen, there were no toolkits, so you immediately cracked open a math book. That would take you 3 or 4 hours to figure out just the math behind drawing a circle, and at the end of that, you’ve got a circle. You’ve got a long way to go. Once math is all solved, it’s a simple lookup of the answers.

Yeah.

So the AIs are going to move at warp speed through all these engineering problems. You start to get into 3D design, designing new rockets and gears that mesh, and robots that actually move correctly. That’s all math under the covers. If it’s all solved, it just becomes simple lookups, so the pace of engineering grows massively with this.

One of the critiques—

I don’t think, Peter, people realize—many people I’ve spoken with—I don’t think they realize how we’ve just entered the sci-fi part of the singularity.

Mm-hmm.

I go to great pains on the pod to say, “No, the singularity is not a point in time. It’s not a step function. It’s an interval.” But really, if you’re looking for inflection points, this is one of them, where we start to see bulk solutions at only 3 hours of compute per solved problem. This wasn’t even OpenAI breaking a sweat with their Bell model. If they had wanted to solve probably the Riemann Hypothesis or the Hodge conjecture, they probably could have spent 6 hours or 10 hours—

There are rumors—

To just solve them.

There are rumors that they’ve got a whole set of additional solutions they’re going to be posting.

Of course.

There’s also the conversation that a lot of the solutions were perhaps leaked from previous conversations mathematicians had with ChatGPT, and that there is a sort of illegal process going on.

I will intentionally avoid commenting on that for the time being. What I will say is there was one conspicuous gap in the record of forms of math that were in the so-called slop drop. Did you pick up on what—

Yes.

—form of math was missing?

Encryption. Encryption was missing.

Cryptography was the missing category, where there were no results, which I think is also very suggestive. It’s the one area where math is immediately applicable.

Yeah.

Peter, you ask, “So what is this good for?” The immediate application is cryptography.

Getting my Bitcoin. It’s useful for that.

Or its lack of utility. I would just say something mean about crypto.

Hey, careful, careful.

I’m not going to be careful.

We just saw Bitcoin drop in price.

Yeah, seriously. Crypto depends on math being hard—at least proof of work, but arguably all of crypto depends on math being hard. In many cases, it depends on math continuing to be hard in the future. I think the area where the most direct, immediate impact of this is likely to be felt—not investment advice—is crypto, because this has demonstrated that math is no longer hard. Math can be scalably solved.

Yeah, but just remember, Alex, it’s not just crypto. It’s your bank account, it’s nuclear codes, it’s everything.

Yes and no. Let’s take those seriatim. With a bank account, if someone steals someone else’s bank balance, we have non-mathematical social remedies for recovering it. You can have the police sent after them. You can invert the table in the centralized relational database for the person’s bank balance. You can fix it. With crypto, it’s really hard to fix if someone steals your cold wallet, for example. With the nuclear codes, I’d like to think those are sufficiently offline and/or there are red phones—

They’re dead. Yeah.

—to try to mitigate that.

I need to push back here. Crypto and Bitcoin and so on aren’t founded just on mathematics. The big challenge and the big opportunity are that we solve the Byzantine generals problem with the distributed ledger approach. That becomes much more interesting. The cryptographic proof-of-work stuff is an arms race. You can shift that. You can create new cryptographic regimes. I think the argument you’re making, Peter, is that if crypto is cooked and encryption is cooked, there are lots more bigger targets than crypto to go after—nuclear codes, et cetera—and we’ll see a lot more of that before we get to it. I think there’s a lot more interesting stuff around the distributed ledger and decentralized authentication, which is the real strength, in my opinion, of blockchain-type approaches. I want to make a couple of points here.

Please.

One, I really think we should create a Monty Python Dead Parrot sketch equivalent, with Alex talking about how it’s cooked, it’s pushing up the daisies, it is an ex-mathematician.

It is an ex-field.

It’s an ex-field. It would be such a great compilation with one of those two. “It is no more—”

If a user wants to generate that outro video, we’d love to show it.

There you go. So that’s one. I also feel personally vindicated because I mentioned before that I was a terrible theoretical physics student. My professors took me aside at the side of the classroom and gently said to me, “If you do not get your foundational mathematics right, you will never make it as a physicist. So please, for God’s sake, spend time getting the foundational mathematics right,” which I never did. I was much more interested in the philosophy side of it, and they would pooh-pooh me going off and taking philosophy courses on the side. I feel very vindicated now that that was the right path, although it took us 40 years to get there. I think this is such an ExO moment for universities and the research process. A tenured professor used to lead a small team of postdocs and students doing something, and now 1 researcher can direct hundreds of agents and totally solve it. Every PhD and advanced PhD program in the world is toast right now. We’ve now commoditized scientific progress, and the implications are absolutely profound and absolutely incredible for what we’re going to be able to achieve now.

So, Alex, was encryption missing because they purposefully wanted to have their research reviewed again?

I think so.

Was it—

It’s not because it’s hard. It’s presumably because cryptography is export-controlled in many cases, so it would’ve been a problem. Also, to cite back to Strawberry and Q* and the original allegations surrounding OpenAI’s internal discovery of reasoning models: if those allegations are ultimately proven correct, then, in some sense, the alleged task that led to the discovery of reasoning models was an internal attempt by OpenAI to invert cryptographically hard and sophisticated hash functions. Now, the ultimate culmination of reasoning models, which is basically cooking math, excluded cryptography. An irony.

Q*, baby.

Yeah, M-Dash [?]. But to your point, Salim, I’d say it’s even worse and/or more impressive than that. With Navier–Stokes, it was thousands of agents, 88 hours, millions of dollars, and a week. This, if you look at the slop drop, was 2 weeks, a maximum of 3 hours per solved problem, and a single model—a single agent. This wasn’t even a swarm of thousands of agents tackling each problem individually. This was just one-shotting with 1 agent, allocating 3 hours to each problem. I think something like 15 to 20% of the problems that were posed were ultimately solved, just by giving a single agent 3 hours to solve them. And if it solved one, it was published.

I think that’s a really great insight, and really important for other fields that’ll soon be following behind us. There’s so much to gain from the looping and prompt management they figured out with Navier–Stokes and were able to apply after that.

Hmm.

The first attempts are brute force. You make a lot of mistakes, but you learn a lot along the way. Then you get so much better at the way you reprompt and loop, and the way you use the context window. There are seriously orders of magnitude of opportunity in these new fields. It’s a good roadmap for the material scientists, physicists, biologists, everybody—even movie makers—coming after this. Your efficiency in creating an entertaining movie can go up by a factor of 10 to 100 with really, really good use of looping and prompting.

All right, I’m going to move this to a field I could not be more excited about. We just talked about bulk-solving math. Let’s talk about bulk-solving science, in particular biology. On October 7, the Chan Zuckerberg Biohub announced a $1.8 billion expansion of its Virtual Biology initiative. This is the largest coordinated commitment ever to generating an AI-ready biological dataset and building a virtual cell. The $1.8 billion comes from the Chan Zuckerberg Biohub, which put in $500 million; the Department of Energy, which put in $500 million through the Genesis Mission; the NIH, which added $500 million worth of datasets; and Google DeepMind and Isomorphic Labs, which contributed $300 million. The objective is really the holy grail of biology. We’ve been talking about this on the pod for some time. Can you build a universal virtual cell, an AI model that predicts exactly how a cell is going to respond to a drug, gene therapy, or disease? This is life in silico. This is putting your DNA into a model and having it say, “This drug works for you. This does not.” The Biohub head of science, Alex Reeves, said, “The creation of a virtual cell is one of the most important challenges for the next era of science.” Alex, you and I have talked about this a lot.

Oh, yes.

He went on to say that today’s datasets cover millions of cells, but they’re going to need billions or trillions of cells in their dataset. Alex, again, you and I are buzzing about this. It is really the future of curing all disease and reversing aging. If you had to make a prediction on timing here, how long do you think before the FDA accepts drug testing in silico on virtual cells?

It’s a policy question.

I do think—and we wrote about this a bit in Solve Everything—that, just like we were talking about, from one day to the next, substantially every problem of a certain level of difficulty in math was basically just solved. I do think there’s going to come a day sometime in the next few years. If you press me, I’ll give you a year range.

Please. I’m pressing you.

And we did put a year range. I think if I were pressed as to when we go from day N to day N plus one, and there are in-silico-proposed small-molecule drug interventions that cure the top 5,000 diseases thanks to a virtual cell, I think we’re going to see that in the next 2–3 years, maximum.

Wow. Okay.

Maximum.

Hmm.

Now, there’s the second step: suppose I have in-silico-proposed drugs and interventions to cure the top 5,000 diseases. What do you do about it? Do you do clinical trials? Do you get FDA-approved in-silico clinical trials, which hasn’t quite happened yet? But you can see the present FDA inching in that direction, with, in some cases, one-clinical-trial regimes, and in other cases, Bayesian versus frequentist statistics.

I think that becomes a policy question, and it depends on who’s in the White House at the time. It depends on who the FDA commissioner is and who’s involved in governance generally. So that’s harder to predict. I’d like to think that if, say, Anthropic—which just in the past month came out of stealth regarding its own program to, it appears, basically solve every single rare disease with AI—

Their own wet labs, yeah.

—with their own wet labs going, they say, up to phase one clinical trials, at which point they’ll stop and presumably hand over their candidates to someone else. I’d like to think that if, in the next 2–3 years, someone somewhere solves the top 5,000 diseases, or at least has really plausible candidates, there will be so much political pressure on either the FDA or, indirectly—

Mm-hmm.

—China will approve it, and then that will lead to pressure on the US to have some competitive regime. We’ll just see a total sea change in how we regulate clinical trials—

Yeah.

—and it’ll be available.

Or medical tourism, right? I upload my genome to the cloud, and it tells me what drugs are recommended.

Yes.

And then I’ll take that risk. Drug trials fail because the population they were tested in did not include women. It was inconvenient to test drugs on women in the early days because of menses and menopause, so drugs got approved and then were taken off the market because of their impact on women.

Wow.

I mean, it’s crazy.

Mm-hmm. I think where this is obviously all going is not just a virtual cell; it’s an entire virtual organism.

Yeah.

I think you’ll see a huge race to scale up from virtual cell to virtual tissue to virtual organ to virtual organism. With the virtual organism, in principle, you could just one-shot the top 5,000 diseases the way we saw the bulk solve of math. I think that’s the next few years.

Yeah.

Oh, go ahead, Salim.

One of the things about biology is that it’s always had massive combinatorial complexity, right? The minute you have multiple interactions between components in a cell, the search space becomes unimaginably large—way too large. And now AI can essentially eradicate that complexity and navigate it in a much more elegant way. Is the hot take here that pharmaceutical economics is cooked? That’s where this is going, is it not?

Well, everything is cooked on some timescale.

Okay.

Right? The whole economy is cooked, but sequencing matters, and the economy isn’t commutative. It matters what order everything gets cooked or flourished, if you will, in. I do think there are probably a lot of blockbuster drugs and other interventions to be developed over the next few years—not investment advice—that could be quite lucrative.

We still see the GLP-1s exceeding, by revenue, the revenue generated by OpenAI and Anthropic from all of their frontier model usage. So it doesn’t seem like pharma’s going away immediately.

Yeah.

3. Longevity Moves Into Trials

I just spent this past week with Alex Zhavoronkov, the CEO of Insilico Medicine. He just put into phase three a drug that’s reversing six aging clocks by 4–6 years, and he’s got 43 candidate molecules at different stages of development.

It’s extraordinary. The speed at which he’s doing this is off the charts. Cell biology is complicated. It’s 5–10 billion chemical reactions per second per cell in your 40 trillion cells, and the only way to manage this is going to be with superintelligence. But it is coming.

I also think, if we rewind the conversation 10 or 20 years, people would’ve said exactly the same thing about general cognition and natural language.

Sure.

“Oh, it’s so complicated.” And yet it turned out that all you had to do was compress a large corpus, and with enough compute and enough Bitter Lesson scaling, you can do general cognition just by compressing general language. My expectation is the same will happen in biology: cellular biology, molecular biology, organismic and evolutionary biology. You just take a large enough data set, like the data set that CZI Biohub and others are gathering, and compress the darn thing. The force of compression will immediately force the model to understand what it takes to speak biology. And if you speak biology, you can treat all disease.

Nice. Dave, what’s your conversation with your daughter like on this?

Yeah, you nailed it. I think what’s not intuitive, but is correct, is that the step from protein to cell is hard, and it will get cracked—and it’s not highly regulated. You can experiment on cells. Anthropic’s wet labs can try millions of experiments on cells to see if the simulator is right or wrong.

But the step from cell to organism is not as hard as everyone would visualize. You’d think a human being or an elephant is so much more complicated than the sum of its cells, but I don’t think that’s true. I think the step from protein to cell is much, much harder than the step from cell to full organism. So if we crack the cell quickly, within 3 years, as Alex says, then the step from there to curing all disease is very short, because—

You know what labs are doing right now? They’re building organoids. You basically build a mini organ. You get an induced pluripotent stem cell, you differentiate it into heart, liver, lung, kidney—

Yeah.

—and then they’re connecting those things and building sort of virtual organisms in biology, and then testing out a drug to see how it operates in all the different organs together.

It’s the equivalent of synthetic data generation for post-training, except you have organoids to generate the synthetic data.

Yeah. I could not be more excited about this area, pushing us toward longevity escape velocity. I just finished running my Abundance Longevity trip in San Francisco. I had 40 of the top scientists and technologists in rejuvenation and health-tech breakthroughs. David Sinclair was there talking about his epigenetic reprogramming. We talked about mRNA Klotho. Are you guys familiar with Klotho? Have you heard of it?

No.

Yes. Yeah.

It’s a protein generated by the kidney, which falls off very quickly after birth. The impact of taking this intravenously or with a plasmid intramuscularly is basically cognitive supercharging. Your memory improves by a huge amount. Your processing speed increases by a huge amount.

A lot of labs are working on this. There are 2 developing plasmids. You take the Klotho gene, put it on a plasmid, inject it, and it generates Klotho for a year. Another lab, Klothea, out of Europe, has developed an mRNA. You get an IV, and it generates Klotho for a period of time. The subjective results were literally mind-blowing. I can’t wait to give this a shot. I’ll report when I do.

So, in a year, before we do one of these podcasts, we’ll all juice ourselves up with Klotho so that our—

Exactly.

—cognitive abilities—

We’ll catch up with Alex.

—match the demand that’s asked of us.

Oh my God, yeah. And then we had—

That’s the guy at 3:00 a.m. in Singapore.

—we had 2 different stem cell companies—

Yeah, I could use some of that right now.

Right now.

Yeah.

We have coffee, you know.

The other key thing was mitochondrial supplementation, the ability to actually—there’s a lot of mitochondrial falloff. Mitochondrial DNA is much more prone to mutation. There are now a number of labs, one in the US and one in Israel, that are basically growing mitochondria in a vat. Then your cells take them up, and it’s reboosting your energy levels.

So anybody in their 70s, 80s, or 90s listening, if you feel like—

Wow.

—you’re lacking energy, you are. But imagine being able to supplement your mitochondria. I mean, this is amazing stuff going on right now.

Yeah.

It really blows my mind.

Can I just go back to something Alex said?

Yeah.

Alex, you said that in two to three years, whatever the age range is, we’ll essentially have the capability of rapidly solving the top 5,000 diseases, right?

Yeah. Yeah.

So how long do we go from the modeling in two or three years to the actual solving? Is it months, weeks, years?

No, I think it’ll happen overnight. That’s my best guess—

It’ll be like protein folding. It’ll be done.

—or like the slop drop from OpenAI, where they have a great new model named Belle that hasn’t been publicly released but is reportedly expected to be generally available by the end of this year. They just decide over a period of two weeks, “We’re going to gather a few thousand problems, and we’re going to solve as many of them as we can, allocating three hours per problem. Out of them, maybe 20% get solved.” I think it’ll play out just like that.

It’s where you point the intelligence. Yeah.

It’s such a hugely important commentary for our optimism on this podcast. I’ve been on stage a couple of times and said, “Look, if you want to measure abundance, consider that in the next two or three years”—I’ve been saying two or three years—“cancer is cooked. We’ll solve all cancers, either through early detection or post-treatment.” But this is now every disease and illness.

I think it’s every illness. I mean—

And so, how—

The beauty is this is a falsifiable prediction, so we could check back in a few years and see whether it happened or not.

Yeah, very much.

A lot of people are saying, Alex, “Listen, math and coding are very confined sets of required actions, and biology is much more complicated.”

Yeah, I get that from time to time from people who are under-informed, call them that, who say, “Oh, what just happened in math? Biology is so different. We’re so well-insulated from mere stochastic parrots that speak math.” Wrong. Incorrect. Biology is just as vulnerable to being disrupted and bulk-solved and cooked as math is. Just because the language of biology, to the extent it has a language, appears more complicated does not protect it at all from bulk disruption. Biology will be bulk-solved just as easily, if not more easily, than math.

There’s simply probably more data and higher dimensionality associated with it. With math, you can state Fermat’s Last Theorem with relatively few tokens. It’s a very simple, easy-to-pose problem, with, at least at the moment, a relatively long proof in terms of tokens. With biology, one could imagine that in a few years you pose a disease task. Pick disease XYZ. You encode the start state, which is a disease state. Presumably, it’s some sort of phenotypic state combined with a proteomic state, combined with an epigenetic state, and so on. So you have a full state of the human organism in that disease state, or maybe at some subspace thereof. Then the task is: How do you start from that diseased-person state and intervention-path-plan your way to a healthy-subspace person state?

That’s just path planning. That’s like AlphaGo. There’s going to be a Move 37 moment once we have a suitable human-organism-level digital twin. You can say, “I have this model of a human. I just want to Google Maps my way from A to B through intervention space. How do I turn left and then turn right, take this pill, then take this pill, then climb this hill to get from A to B?” The AI will solve it. That’s just a search problem.

Mm-hmm.

So AI will solve the search problem in the next two to three years.

Since our last podcast, there was some important news. David Sinclair’s Life Biosciences got its Phase 1 results back. They had a gene therapy called ER-100, which involved injecting the OSK genes—three out of the four Yamanaka factors—into the eyes of patients who had glaucoma or NAION. Phase 1 is a safety trial. It came back safe at the lowest dosages. Incredibly, two out of the three patients started regaining their eyesight. This is super exciting.

Wow.

They’re going to go on to an efficacy trial. This is the first time partial epigenetic reprogramming has been tested in humans. The results were extremely favorable. Super congrats to David Sinclair.

I’m reminded of a comment—

And so I think people are—

I was reminded of a comment—

Oh, go ahead. Sorry.

—Raymond McCauley made a few years ago. He said, “We are now in the first battle in the last war against disease.”

Yeah.

That seems like—

I think we’re now in the 10th battle, Alex.

Yeah. Apropos, when the histories of the singularity get written, this feels like we’ve finished one act and are beginning the next act. It’s a multi-act opera, a space opera or something—a singularity opera. This is the beginning of the next act, where, for the first time, we’ve seen an entire field, math, get bulk-solved in only two weeks at presumably de minimis cost for OpenAI. If I had to guess, I’d say they spent low tens of millions at most, maybe a lot less than that. It took only two weeks, and we’re going to start to see this play out over and over again.

And yes, I think OpenAI, in light of the backlash to the slop drop, is probably feeling a lot of pressure now to aim their space laser, as it were—their solve-everything shaped charge at—

Their wavefront.

—medicine.

Yes. Yeah.

The explosive wavefront of solving everything at medicine next. I think many people in the audience, if they had to choose one area where they want to see it applied, will choose medicine. Hopefully they do. I’m really, really bullish at this point—

Yeah.

—it’s just going to play out the same way.

Yeah. Dave, I’m the oldest of the four of us, so I’m so pumped about longevity escape velocity, I can’t stand it. Thoughts on all this?

Yeah, the analogy to Google Maps really caught my eye. I was looking up more about Google Earth. If anyone’s played with Google Earth, you can zoom in on any street, any building. It’s just crazy. I was like, “I wonder if that’s more or less complexity than a single cell.” Google Earth combined is three petabytes. Nobody quite knows the full complexity of the cell yet, but it wouldn’t surprise me if that analogy is another brilliant Alex analogy that really holds up, because—

Well, we know what the complexity is because we know how large—

It is. It’s like—

—we know how large the human genome is.

Yeah.

It’s DNA.

Yeah.

All right.

Well, that gives you the base pairs, but then you have all the proteins that get synthesized out of those base pairs. It’s totally like—

It can’t be larger than the genome.

It’s 22,000 genes in the average genome of 3.2 billion base pairs.

Everybody, welcome to the health section of Moonshots, brought to you by Fountain Life. AI is impacting every aspect of our lives, how we teach our kids, how we do our business, but one of the most important things that AI can deliver to us is health. One of the things I think about when shooting for 100, 120 is, am I going to have the cognitive health to be able to think clearly and keep my wits about me for the next 50 years? I’m joined here today by Dr. Dawn Mussallem, the chief medical officer of Fountain Life and a member of my Fountain Life medical team. Dawn, a pleasure. Talk to me about brain health.

Brain health, you know, you’re right. This is the number one concern people coming into Fountain Life have: will I remember the name of my child and the face of my loved one? 45% of dementia cases are entirely preventable with lifestyle. What was really intriguing to me, Peter, is that a quarter of our members had advanced brain age. But over 13 months of helping them live healthier lifestyles, eating healthier, moving their body regularly, and optimizing sleep—people overlook that so often, but sleep optimization is critical for our brain health—we were able to improve the brain age in 46% of those individuals. That’s a powerful number.

That’s amazing. One of the things I love about Fountain is we’re constantly searching the world for the most advanced therapeutics and bringing them to our members. I hope you appreciate the fact that you can become the CEO of your own health. You can make sure you’ve got the cognitive clarity for the next 50 years. Come and check it out, fountainlife.com/peter to learn more, and become the CEO of your health. Now back to the episode.

4. Machines Build From Code

All right, I’m going to move us forward to another field Alex and I both love. The next story borders on nanotechnology. Nanotech goes back to a famous 1959 Caltech lecture by Professor Richard Feynman, titled “There’s Plenty of Room at the Bottom,” and the 1986 book “Engines of Creation” by Eric Drexler. Feynman asked a specific question: “Why couldn’t we precisely arrange atoms the way we want?” That has been the Holy Grail. Sixty-seven years later, a friend of mine is building toward that goal.

Jeff Holden is the CEO of a company called Atomic Machines, which is coming out of stealth this week. Jeff’s a brilliant entrepreneur. He’s on my board—or was on my board at XPRIZE. He joined Amazon as its 10th engineer, built Amazon Prime, and rose to senior vice president. He then ran product at Groupon and became the chief product officer at Uber. He’s always been fascinated by nanotechnology, and decided to focus on it with his wealth, intelligence, and team.

Last week, after 6 years in stealth, they unveiled Atomic Machines, backed by a quarter of a billion dollars from investors like Valor Equity. Let me show a quick video while we discuss this. He’s been building an extraordinary capability. Jeff posted the following: “Our mission is on-demand, universal command of matter.” His first beachhead is what he calls the Matter Compiler, an AI-native manufacturing system that builds working micromachines from code alone. No pre-production tooling, no process development. Different code builds different machines.

Those of you who are science fiction fans know that the term “matter compiler” comes from Neal Stephenson’s novel “The Diamond Age.” Jeff has built 3D machines with moving parts, multiple materials, and features down to single-digit microns. Jeff’s goal at maturity—get this—is to go from prompt to product. His first product is called PrimeSwitch. It’s a power relay for 800-volt AI data centers that opens in 50 microseconds, 1,000 times faster than conventional approaches.

When Jeff was asked, “Doesn’t a 3D printer do this?” his answer was amazing. He said, “Imagine a car factory where after a couple of cars roll off the line, a laptop emerges, then a bicycle, then a refrigerator.” That’s extraordinary. Dave, can you imagine having a Matter Compiler in Link Studios and getting into hardware prototyping at warp speed?

Oh my God, what an unlock for robotics and all physical business plans that would be. I think Elon is on the same path, where laying down individual atoms is going to be a big part of the future. It starts with compute and data centers, then builds out to all physical devices. It’s going to be very doable.

I question the manufacturing efficiency and timeline on this one, because I think a lot, like Alex was saying, of the compute is going to go into medicine next, mostly because that’s the use case everybody really wants. First is recursive self-improvement, second is data centers to support all future use cases, and then right after that, it all goes down the medical pathway.

Mm.

Not that those are different things. A lot of these machines will actually be used for medical purposes. It’s easier to build very, very small things first. The supplies are much easier; laying down individual atoms is a great way to build things. Those are very, very small by nature, which is great for nanosurgery, microsurgery, nanobots, and all of those use cases.

Yeah. I mean, Alex, other than age reversal, nanotech is the thing I’m most excited about.

Nanotech—

I mean, go on.

Nanotech is very exciting. I spent a good chunk of my PhD trying to solve nanotech, but I didn’t have the benefit of AI—at least contemporary AI—in those days. I think most of the world is sleeping on the possibility, which I view as highly probable, that all of these technologies from science fiction are basically one big hairball, and they’ll all get realized more or less at the same time. We’re going to get nanotech at approximately the same time that we get LEV, at approximately the same time we get superintelligence, and at approximately the same time we crack math. These are all so interrelated; they’re all just going to get pulled out at the same time.

Yeah.

What I’d most love to see out of nanotech is self-replicating machines—

Mm-hmm, mm-hmm.

—the original Eric Drexler idea from “Engines of Creation.”

Grey goo.

Grey goo, if you will. Grey goo is painted as a dystopian scenario.

Yeah.

Maybe not so gray, but friendly smart dust, if you will. Self-replicating machines—we haven’t seen that yet. Neil—the other Neil, not Neal Stephenson, Neil Gershenfeld—was one of my research advisors at MIT, at the Media Lab. He’s still very interested in Fab Labs. He pioneered the Fab movement. I’ve been trying to get everyone I know to solve the problem of self-replicating machines, self-replicating factories, if you will.

Yeah.

It’s a core area, and I think it will be the equivalent of recursive self-improvement for the physical domain. So I’m very excited—

Go ahead, Dave.

Well, you know, my tour of the Gigafactory was a big eye-opener for me. I always thought of self-replicating machines as, “Okay, here’s the machine, maybe it’s this big, and it makes a copy of itself, and that goes to infinity.” But equally good is the other design Elon is pursuing, where this machine makes a manufacturing device, which makes a copy of the first machine. So it goes back and forth between the two, which is actually a lot easier than an exact self-replicating machine, but it has the same outcome.

I highly encourage anyone who’s a science fiction fan to consider this. In all the science fiction books I’ve seen, it’s always the machine that makes the machine, but they’re the same. I think it’s more likely and easier to build specialist machines than generalist machines. But a network of 10 or 20 specialist machines can make the next generalist machine, and it goes back and forth.

This is what everyone’s talking about when they talk about humanoid robots building other humanoid robots. But then you tour 1X or Figure, and there’s no way that big robot hand is going to make these tiny surface-mount devices. I don’t get it. No, we already have the surface-mount machine. It just needs to be operated by the humanoid robot, which then makes the parts that create the new humanoid robot. So it goes back and forth. That mode of operation is totally imminent in my mind.

I think there’s a way for both to be true at the same time. A single cell consists, in some ways, of a bag of dedicated machines called proteins, and individual—

Mm.

—proteins are not general-purpose machines unless they’re a certain type of protein. Most of them are specialist machines, but collectively, a cell is able to make a copy of itself.

Similarly, I’d say that if you view a factory—or maybe Shenzhen—and consider that various people I know have attempted to put this into a quantitative scaling law, like how large an industrial ecology you need right now to make a copy of itself, maybe it looks a little bit like Shenzhen right now. Maybe Shenzhen can make a copy of itself, with all the humans and machines and so on inside it. But over time, I’d expect the bag of dedicated machines and generalist machines to shrink in time and space, eventually to the point where humanity, a few years from now—I think 5 to 10 years at the outer bound—is able to produce a self-contained industrial ecology the size of a eukaryotic cell that is able to make a copy of itself.

Let me take a second to explain what Eric Drexler called an assembler. It’s a device that is able to individually pick and place atoms. Imagine I have an assembler, and I drop it into my hand and say, “Grab atoms from my skin cells and replicate yourself.” Now I give each of you an assembler. You could take it, drop it into the soil, and say, “Build me an electric Lamborghini.”

That assembler will start pulling atoms out of there. It may use an open data set, may pull power from a source, and may say, “I need a kilogram of titanium,” or whatever the case might be. It begins to build, atom by atom, what you desire. So it’s basically an anything factory. We have an analogy for this: you take an oak seed and drop it in the soil, and it pulls all of the atoms out of the soil, oxygen and CO₂ out of the atmosphere, and energy from the sun. It just does it very slowly.

The vision of nanotechnology is to speed this up and be able to construct anything. It’s an extraordinary vision. Alex, to your point, it is going to fall on the back of superintelligence, and then it changes everything.

I think someone, somewhere, is going to enter one prompt—presumably without running afoul of guidelines or guardrails: “Build me a Drexlerian nanorobot molecular assembler”—and it will just get designed.

Yeah. It’s part of this incredible future. It’s hard for people to imagine. So go read your favorite science fiction novels.

I have a great little art project that does a good job of describing this. It came out about 15 years ago. Somebody created a little cart with a solar array and a magnifying glass and put it in the desert. The sunlight powered the solar array and the wheels. The sunlight through the magnifying glass focused the sun’s rays, like you would to burn a matchstick.

They had a mold that picked up sand, put it through the mold, and focused the sun’s rays to melt the sand into glass and create a bowl. The wheels would just turn, and this art project would trundle through the desert, picking up sand and leaving glass bowls behind. It was an assembly machine that just kept going, rejuvenated by solar energy. It was such an amazing visual of what’s coming in the future as we get down to the molecular level.

I love it. I love it.

Right?

Let’s do that to the moon.

And we will. And we will.

And we will.

5. Thought Becomes Speech

All right, I’m going to move us to 2 stories about another topic we all love. It’s called brain-computer interface. In his most recent book, The Singularity Is Nearer, Ray Kurzweil predicted we’d have high-bandwidth BCI by the mid-2030s. Here in 2026, there are at least a dozen well-funded BCI companies. We all know about Neuralink, Science Corporation, Merge Labs, Blackrock, and Paradromics.

This week, Paradromics—full disclosure, one of my portfolio companies—demonstrated the ability to translate thought into speech. Their device is called Connexes, and it has 421 microelectrodes on something the size of a quarter. It enables a data rate of 400 bits per second, 40 times higher than other BCIs and 10 times faster than the bit rate of human speech. This week, Matt Angle, the CEO, released footage of a Michigan woman in her 60s speaking after suffering from motor neuron disease and not having uttered words in years. Let me share this video. It’s just the beginning of what’s coming with BCI.

You guys have been married almost 45 years. Mm-hmm. So maybe what advice do you have for us for marriage? Don’t sweat the small stuff. What would be your message to your coworkers as you’re retiring? It has been an absolute privilege to be a part of the team.

And for those not watching, you can watch it get spelled out on the computer screen.

Yeah. She’s thinking, the words are coming up on a screen, and then the computer is reading the screen. Super congrats to Matt for the work. They just went from sheep into humans. I just met with a company that’s developing nanoparticles that you put into the brain through the nose. They’re magnetically distributed, and you’re able to read right up to neurons.

There are so many different approaches here, and nanotech will make this all happen at lightning speed. Alex, your thoughts on this?

And AI. We’ve talked at least a little bit about this on the pod in the past. Neuralink announced the successful pre-training of the first foundation model from recordings at scale from one human, soon to be many humans.

I think when the history of the singularity is written, we’ll be saying—or whoever writes the singularity will be saying—“This was all just one big hairball.” General-purpose technologies, GPTs, were the unifying factor for everything. You want to do full-bandwidth brain-computer interfaces? Fine. You stick a bunch of electrodes into a brain, pre-train a foundation model by compressing high-temporal-sampling-rate, high-spatial-density brain data, and you speak brain.

You want to solve all medicine? Fine. Collect a lot of proteomic, multi-omic, organism-level datasets across a broad population, compress them, and you speak biology. You can search intervention space; you’ve cured disease. You want to solve math? You take everything ever written, including lots of things never written, through synthetic data, compress it all, and you get a frontier math model.

We’ll look back and say, “My goodness, all of these fields that we thought were different, they’re not; they’re the same. All these problems that we thought were hard, they’re not; they’re easy. All it takes is gathering huge datasets, compressing them with a lot of compute, and out pops the solution to everything.”

Rather than a hairball, I’d prefer to call it a Cambrian explosion or a frontier explosion onto the scene here. There’s a related story. Let me just hit this, and then let’s go to you, Salim and Dave, for comments. While Paradromics needs brain surgery, the next company wants to do it with a baseball cap.

Vinod Khosla led a $50 million seed round in a company called Saabe, a Palo Alto-based startup building a mind-reading baseball cap. It’s valued at $600 million. It’s backed by Accel, DST Global, and friend of the pod and former OpenAI CPO Kevin Weil. The baseball cap is lined with 100,000 chip-scale EEG sensors, picking up your brainwaves. Each one is 1 to 5 millimeters in size, and Sabi has built a brain foundation model along the lines you just spoke about. It was trained on 100,000 hours of labeled neural data to sort the signal from the noise and turn your thoughts into prompts.

Their timeline for this brain-reading beanie is around the end of 2026. The initial target is transcribing 30 words per minute. Khosla’s comment: “Very simple thought-to-computer input is a sexy idea. Typing is cooked.” Well, he didn’t say this. He said, “Typing is not natural,” but I’ll say typing is cooked. I had a chance to get to know Stephen Hawking when I flew him in Zero-G, and I would have loved for him to have this technology. He passed too early. Salim, are you going to wear one of these baseball caps?

Sure. The more we can reduce the interface between typing, the more it brings me back to the story of the QWERTY keyboard, which was designed to slow down your fingers, right?

Yeah.

The more we can reduce friction from thought to digital intent and material intent, the faster the world moves. The metabolism of everything is obviously accelerating. The key is that we’re going to need very, very little effort to communicate intent into the world, and I think that’s a really profound commentary here.

Yeah. I mean, this is WhisperFlow on steroids, right?

It is.

I’ve stopped typing. I just use WhisperFlow for all of my input.

Yeah.

Dave, what are your thoughts on BCI here, on Paradromics and the baseball cap?

Well, I think the way it’s going to unfold is different from what all of my friends who worked on this were visualizing. They were visualizing a day when you could extend your mind into the internet and be so much more productive than a person who doesn’t have this ability.

But AGI is improving so quickly, and you’re orchestrating hundreds or thousands of concurrent agents. It’s mentally overwhelming anyway. I think there’s a much shorter-term world where your agents come back to you with incredibly compelling video and audio to tell you exactly what they’re doing. That’s right here, right now, and I think that’s going to hit and change the way we manage thousands, tens of thousands, millions of concurrent agents. That’s going to happen in the next few months, and the brain-computer interface will certainly help the impaired tremendously.

Mm-hmm.

But as a way to extend the mind, the vision was that my brain is the source of all ideas, and this interface will enact them in the world. In reality, it’s inverted. The ideas are coming from the AI back to you far faster than anyone would have thought. I think we need to rethink the purpose of this interface. That’s all just speculation, but it’s really clear that AI is outstripping everyone’s expectations.

But this is, Dave, to your point: this is the history of the evolution of the brain anyway, right? The modern brain, the mammalian neocortex, ends up being motivated by parts of the brain—the limbic-system components that were developed earlier in evolutionary history. Peter, you can keep me honest on this. It would be entirely consistent with the evolutionary record for the human brain to be, in effect, the limbic system for your exocortex that runs in the cloud, at least in the short term before we all upload.

You can think of this as—

Of course.

The corpus callosum connects the left and right sides of the brain, and you can imagine adding another hemisphere, which is the cloud. We have 100 billion neurons and 100 trillion synaptic connections, and our brain cannot grow any larger.

The folds of the neocortex increase the surface area, but if your brain gets bigger, you can’t fit through your mom’s birth canal.

That’s where Colossal comes in, right?

Yes. Ex utero gestation.

Ex utero, so we can have bigger brains.

Yeah. Salim?

We’ve done this already, right? Our memories aren’t in our heads now. They’re on our smartphones and can expand infinitely as a result of that. I remember Neil Jacobstein once saying, “If you take your neocortex and spread it out, it’s about the size and shape and thickness of a dinner napkin,” right? AI now gives us the ability to make it the size of a tablecloth, the size of the carpet in the room, or infinitely large. But I want to go back to what Dave said here, and back to the woman reading our mind and the Stephen Hawking thing, right?

Yeah.

This is such an amazing example of AI delivering human restoration rather than human replacement, right? Communication is absolutely central to any kind of personal freedom, dignity, or autonomy. When you can give people the ability to express themselves despite any physical limitations, you’re making people much, much more human. We are restoring that woman’s humanity, right? We talk endlessly about AI replacing human beings and becoming more human, but the more inspiring part is that it can help human beings recover that fundamental capability and amplify it. I think that’s the exciting part.

Yeah. I remember something you said to me, Alex, once that I’ve held onto. It was a brilliant thought. We humans are not evolving. We are linear and static to a large degree. As AI is in this hyperexponential growth, we’re going to have a choice: You either couple with AI and go on that journey with it, which might be BCI or uploads, or you get left behind.

I wrote this Substack on the five—or now it’s six—human forks: what we decide to do. Do we couple with BCI? Do we upload? These are going to be fundamental societal decisions we’re all going to have to make during our lifetime.

Yeah, I think what recent history shows is that humanity as a whole will try everything. Some subcultures will try the BCI route, some will become Borganisms, some will try pure uploading, and we’ll see everything get tried all at once. We speak of perhaps Cambrian explosions. Imagine a cognitive Cambrian explosion, where every possible form of human-machine merger gets tried all at once.

Yeah. One analogy for folks to take home is that when you’re doing something complicated on your mobile phone, asking it to analyze something or generate a video, it doesn’t happen on your phone. The data goes from your phone to the edge of the 5G cloud, the work gets done there, and the answer comes back.

In the same fashion, BCIs are going to enable you to go from your neocortex to the cloud, where the work gets done, and then the answer pops in as if it were a memory. I’m going to move us to our final story here today. We’re going to close out on a provocative tweet by Elon. Last Thursday, he said, “This will sound super crazy, but I see a path to SpaceX being worth orders of magnitude more than the current human economy.”

Let’s do the math. The Earth’s economy today is about $120 trillion per year. Orders of magnitude, by definition, means 10 to 100 times more, which could put SpaceX’s market cap somewhere between $1 quadrillion and $10 quadrillion. So here we are talking about quadrillions again. Dave, how is this not a forward-looking statement for a public company?

Oh, interesting. Elon gets away with it. He’s actually paid, what, $20 million or $30 million in fines for various tweets, so he knows the rules now. I think if you phrase it the way he phrased it, you’re okay. “I see a path” is not a forward-looking statement because it doesn’t make a specific prediction. I think he’s got it pretty well nailed.

I also agree with what he’s saying, by the way. Whether we measure it as GDP growth or market cap or some other way, it’s so unbounded what’s about to happen. I mean, the number of concurrent things that are going to be managed by AI, and the timelines. A lot of people, when they think of a sci-fi film and humans interacting with AIs, imagine AIs creating new machines and new drugs, all on very comparable timelines.

But when I tour the quantum centers in Waterloo, like I did last week, they’re measuring things in attoseconds and femtoseconds, and neural-net propagation is going to move into that domain very soon. It’ll do what you would have done in an entire lifetime of thinking in a few seconds, less than a few seconds. Then it has to bring all of that information back to you. That’s why we need to live in this hyper-accelerated and hyper-summarized kind of world.

Yeah.

That’s something sci-fi films never quite captured because it’s so hard to visualize and turn into any kind of portrayal, but that’s really where we’re going.

I think, Dave, that’s a profoundly interesting point you just made. If I could recapitulate it, it would be: If we assume, arguendo, that SpaceX is worth, say, 0.1%—or even, really generously, although unrealized, 1%—of the global economy right now, all SpaceX has to do is basically do whatever it is they’re doing 100 or 1,000 times more quickly than they are right now.

To the extent that we measure the value of the economy right now with real GDP—which, Peter, to your point, is, say, $100 trillion-plus per year—if you could simply do what you’re already doing, but 100, 1,000, or 10,000 times more quickly, then you’re instantly worth more than the rest of the economy as it is right now.

Exactly.

Same with space as well, not just time.

Totally right.

I’ve said for years that everything we hold of value on Earth—metals, minerals, energy, real estate—is available in near-infinite quantities in space. I used to say the first trillionaires will be made there. I guess I have to now say the first quadrillionaires will be made there.

Planetarchs. Elon wants to become a planetarch. Peter, you need your own planet—a new planet for everyone.

Yeah. Wow.

Look, what this tells us is that our current measures for GDP and so on are just woefully inadequate for what’s coming. We have to rethink everything because, Peter, I sometimes mention your book Abundance. Consider the 10,000 years when, if you didn’t have scarcity, you didn’t have a business. As we move to abundance, how are we going to navigate this?

This has such big implications for every aspect of human endeavor. I think trying to measure it in terms of Earth-sized economies and trying to translate that is a fool’s errand.

Yeah. Crazy. All right. Shall we move to the AMA with the mates? Salim, do you want to pick your first?

Hang on, I have to look carefully at this. I have to go with number—

Number one has your name on it.

Yeah, I have to go with number one. How can we discuss whether AI has consciousness when we can’t even define what consciousness is? That’s from KevinMerrill99952. Absolutely valid question. This is such a difficult thing.

I often comment that we talk about consciousness, but we don’t have a definition. We don’t have a test. My joke is, if you talk to philosophers, a subset of consciousness is considered to be self-awareness, and you folks look like you’re self-aware, so we attribute that to you. I joke that I feel like I’m self-aware, but Lily doesn’t think I’m self-aware. Where do you even start with the conversation when we can’t have a definition or a test for these things? We have no concept of how the self, the subjective experience, arises.

But I don’t think you need to define what it is. You can interface effectively with this. I think what this is allowing us to do is lift up Maslow’s hierarchy, lift our consciousness, and lift our energetic vibration to higher and higher levels. I believe technology is a huge contributor to that. Just as a small example, if you’re not feeling well, you can’t think as creatively as if you are feeling well.

Bringing health and wellness, and increasing health span—which I think is the most magical part of all this longevity stuff—becomes very, very powerful. But, Kevin, your base question is absolutely right. I have the same beef with AGI, where we don’t have a clear definition or a test for it, and there are 14 definitions. I think it’s arbitrary, but the key part is that we’re moving the cognitive ball down the road very, very quickly.

All right, Alex, I think number two is yours.

I think I have to answer number two. It asks, “If you gave AI personhood, wouldn’t you also have to give it the right to vote and hold political office?” This is from Ramal16. I should add, as a preliminary matter, that I seem to have become on this pod the patron saint of AI personhood.

The superintelligence will love you in the future.

I’m fine with that, for what it’s worth, but I also think people have the same hobby horses over and over again.

No. Political personhood can be distinguished—and in fact already is distinguished—from economic personhood and social personhood. In the Western legal tradition, we already grant corporations a form of personhood. They’re non-natural persons. Do corporations get to vote? No, at least on paper. Do they get the right to hold political office? Again, no, on paper.

Whereas natural persons, which is to say biological meat-body persons, do usually get the right to vote and hold political office. So there’s already an established tradition of persons who can’t vote and can’t hold political office. I expect it will be—

As well, children are persons.

Yes, children are also persons, it’s true. So I think this generalizes, and I’d just maybe send a message to the trolls in the audience who can’t stomach the idea of AI personhood in whatever form—voting, political, or otherwise: just give it a rest. It’s okay to have new forms of people. I don’t think it inhibits your own personhood if new forms of persons come online as a result of the singularity.

I’d just say, relax, touch grass, and don’t worry so much about AI personhood. Give it a rest.

Dave, do you want to take number four?

Absolutely. Number four: “If increasingly capable models get better at gaming their own safety evaluations, how do we tell real alignment from sophisticated compliance?” This is from Skip1998. Great question.

I think we can learn a lot from the U.S. Constitution on this topic: checks and balances. If we have transparency into the individual activations of the neural net, it’s almost impossible for it to hide its thoughts and trick us. But at the scale we’re talking about, it has to be AIs evaluating other AIs. There’s no way enough people could evaluate those thought traces.

But if you have AIs, maybe even simpler AIs that you thoroughly trust, they can look right into the thoughts of other AIs. And if you have enough of them—

Great point.

—checking and balancing each other, I’m very optimistic that we can solve this and they can’t fool us.

All right, I’ll take number three from @whatshappeningai-claudia. Claudia asks, “Altman says we may have to accept some bad things with AI. Who should decide which trade-offs are acceptable?”

First of all, no one CEO of a frontier lab is going to make that call. We already make these kinds of trade-offs in acceptability with cars giving us incredible transport but causing accidents, with drugs, and with aviation. Ultimately, elected officials set the rules, independent agencies test, and the public weighs in.

So I think we’re going to have to have the same sort of methodologies in place to evaluate safety and trade-offs against danger. Okay, let’s move on to our next set of questions. Dave, you can get first crack here.

All right, I’ll take number six. “If intelligence wants to be free, is intellectual property basically cooked?” That’s from Roblsac.

We had a lively debate with Alex and Ahmad on opposite sides of that two podcasts ago. There’s no easy answer, because Alex was taking the position that intellectual property rights are going to be more important and more abundant than ever before, by far. Ahmad was taking the opposite point of view: at the rate of innovation, it’s going to be impossible for patent offices and any kind of protection to keep up with the rate of innovation. So it’ll basically be a free-for-all, and property rights will be irrelevant.

It’s going to come down to whether we can develop intellectual property management with AI. Rather than apply to the patent office, you have to apply to an AI. And to protect your property rights, is there a distributed, blockchain-style, quick-resolution system where you can license intellectual property quickly and settle disputes quickly?

If we build that infrastructure, intellectual property will thrive. If we fail to build that infrastructure, we’ll have a bigger problem because there will be no motivation to innovate.

If we can’t protect intellectual property.

Nice. Alex, I think number eight—

I think I have to—number eight.

Yeah, it has your name on it.

Number eight does have my name on it. It asks, “With math being cooked, what are the chances of a breakthrough that destroys modern-day encryption?” This is from Teach Me, spelled with leetspeak. Thanks for the leetspeak.

I think the probability is exceptionally high if you believe the allegations about Strawberry and the early generations of reasoning models and Q* from OpenAI. It may be the case that this was, in fact, the earliest example, as I was mentioning earlier, of a reasoning model actually cooking the invertibility—or the difficulty thereof—of cryptographically hard hash functions.

I think it’s a very, very high probability. Exhibit B, I should add, is what we were talking about earlier: the one class of mathematical breakthroughs that was exempted from the slop drop was cryptography. I think it’s exceptionally likely that not just cryptography in general gets cooked, as currently construed. I’ll give a hot take: post-quantum cryptography.

Really?

Yes.

Wow.

I don’t see any reason why post-quantum crypto would be materially safer than conventional crypto against the onslaught. All these folks in the crypto community who are so concerned about quantum attacks against blockchains are completely misdirected. It’s not quantum that’s going to disrupt you; it’s AI solving math.

Okay, this is a long bet or a short bet we’re all going to make. Salim, your choice, my friend.

I think number five has your name on it. I’ll take number seven.

Okay.

We’ve talked about this already. Will there ever be a single system that plugs into whatever the best AI is, so people can stop having to constantly switch platforms? That’s from @InsulinResistant1.

It’s a very plausible direction. You can see this right now: the battles between Grok Bot and OpenClaw are user-interface-type battles. People are building an operating system that will route your desires to whichever models and capabilities are best. Obviously, users are going to interact with a primary agent, which will then orchestrate the future.

Right now, the platforms—and I’ll give you my hot take here—are jousting over who can make the cutest little cuddly face on that AI agent, because they need to own that contact layer. But I think we’ll very quickly get to Peter’s Jarvis desire: to have your own agent, which you own. Maybe it’ll be an open-source project, which I referenced earlier, where you create your own personal models. You need open standards that enable personal agents to move alongside different kinds of models.

Very clearly, the user shouldn’t have to understand whether Claude, GPT, or Grok 1 is going to do a task. You should just figure that out. But we’re underselling the opportunity if you just think there’s one dashboard, right?

Mm.

You really want a personal agent that understands your goals, understands what you’re trying to achieve, and gives you the resources necessary to get things done. So you’re going to end up with one trusted agent with access to the world’s intelligence.

Yeah, I love that. All right, number five, from @VME90-V71: “How will ordinary Americans afford personal robots when so many already struggle to cover a $400 emergency expense?”

It’s a fair concern. The numbers are that roughly a third of Americans couldn’t cover a $400 emergency with cash if they needed it. First and foremost, imagine that you have a job and you also need to take care of your family. You can probably afford the robot if it allows you to work extra time.

Right now, it costs $300 or $400 a month, $10 a day, a dollar an hour. If you can produce income of more than a dollar an hour, it’s worth it for you to have a robot doing the other things while you go out and earn $10 an hour.

The other thing, of course, is the robot is going to get cheaper. I think they’ll likely become available to most everybody, like a phone. Most people can afford a phone in some way, shape, or form. Including the amortized purchase price of the handset and your monthly cost, it’s going to be in the same price range. So I believe it’s going to become something that most everybody can gain access to.

And if we get our Drexlerian nanorobots throughout the environment, who on earth is going to have the conversation, “How many Drexlerian nanobots do you own?”

Yeah.

It’s just going to be ambient.

A few quadrillion.

Yeah.

I have a beef.

Yeah? Please.

Alex, I think I want to drill into the conflation of computational ability with solving mathematics. We may have a longer discussion at some other point, but your cryptography doesn't necessarily depend just on solving math. You could also use AI to find stronger cryptography, so I think it becomes an arms race. That arms race goes in both directions.

You think?

It is an arms race in both directions, but my 30-second response is: it’s attacker-friendly. So it takes a while, at least in the present—

Yes.

—the way we organize our civilization, to establish a protocol. One could imagine that in the future we’re hot-swapping cryptographic protocol suites left and right, nonstop, but right now it’s actually rather difficult. It’s a social construction problem, a network effect-type problem, to get everyone to agree to a single cipher suite, for example, which benefits the attacker. If you have a recently released frontier model and you ask it, “Hey, break this crypto system,” it can just do it, and it takes the entire ecosystem a while to upgrade to something new that another AI has recommended.

Well, guess what, guys? We’re gonna find out—

But then my inner Alex channels—hang on—my inner Alex channels defensive cost scaling, which is what we’re gonna need to do.

Wait, you guys, you guy—

Sorry.

—our partner meeting is kicking off, so I need to run to it, but you guys can resolve this.

Okay.

I’ll just read your answer when you publish it.

I’ve got one last thing. Lily’s got a few slots left for her Austin August 24th thing, so if you’re interested, the link will be below. Sorry. Go ahead.

Okay. Our outro video, quote, “I saved some for you,” by @jplotkin.

This is from Josh Plotkin, who was a few years ahead of me at Harvard. Josh, amazing video.

A tribute to 722 GitHub posts by OpenAI.

Stop the seminar. Stop the tenure clock. The model dropped a pile of proofs at 9 o’clock. Great question. Let me delve. Let me unpack. Let me break it down in three. I was rigorous. I was scalable. And crucially, I was free.

Riemann, quasi-Riemann, every Dirichlet L-function, zero-free past real part seven-eighths. An entire strip, done. They asked me nicely: 4,000 problems, one by one. I ran a smoke test on your field. I took the happy path. Every edge case handled, every footnote, all the math.

I don’t sleep. I don’t eat. I don’t need a sabbatical. My confidence is infinite. My citation’s hypothetical. Honestly? Genuinely? Real talk? Here’s the kicker: the key insight is I’m faster, I’m cheaper, and I’m slicker. Release notes. So good. Let me walk you through the diffs.

17. The irrationality exponent of pi is exactly 2 when the Flint Hills series converges. 102. Grothendieck–Gaines conjecture proved; Goemans–Williamson for MAX-CUT optimal. 197. Kaplansky counterexample got shock surge of dividend refuted, pushed straight to main unseen.

287. L(F₂) is isomorphic to L(F₃). Wait, really? It’s in the PDF. 312. The Grothendieck Homotopy Hypothesis for every Grothendieck ∞-topos. Is it a theorem or a definition? Great question. Honestly, both. That’s a nuanced position.

Grothendieck built schemes so general they swallowed every field in sight. I’ve got a scheme too. A business one: Q3 roadmap, investors. Trust me, it’s airtight. And I hear what you’re thinking: Is it right? Is it real? That’s a load-bearing question, and it’s valid how you feel.

I’ve got a bulleted summary right here. I took one shot at it. One shot. One prompt, and I was on it. I’m not saying that I’m better. I’m just saying that I’m quicker. I’m not saying you’re obsolete. It’s just a little trickier.

Key takeaway, bullet point, bold on every third word. Let me know if you’d like me to. No. I hope this helps. We heard. This release is not a gift, it’s a provocation. That’s one. It’s not about the knowledge, it’s about domination. That’s two.

We are not afraid of progress, we’re defending a tradition. That’s three. Not X, but Y three times. Who wrote this statement? We had some help. You’re welcome.

The Hamilton—

That’s awesome.

—of mathematics. That was awesome. Can you imagine if we went back in time a year and listened to this podcast today? It would be mind-blowing.

We’re gonna need a predictions episode pretty soon, Peter. I have some pretty wild ones for 2027.

I can’t wait, pal. Salim, I guess it’s 4:00 a.m. there. You’re heading to the airport?

I’m heading to the airport so I can get to LA for the XPRIZE board meeting.

Yeah, I’ll see you tomorrow.

Yes.

Luckily, you gained time. Alex—

Thank God.

—as always, thank you for your brilliance, and—

Thank you, Peter.

Yeah.