Google的Jeff Dean离职，SpaceX营收突破1000亿美元，OpenAI的Astra解开困扰数学界十年的难题

Peter H. Diamandis

Big news out of Google this week and a serious shakeup. Jeff has been Google's chief scientist for 27 years. He's leaving to co-found Discovery Loop. Gemini has lost the mandate of heaven.

SpaceX expects to hit 100 billion in annual recurring revenue by the end of this year. So everything that will be economically productive in this next century, Elon's going full stack on.

I think there are two possible pathways to trillion dollars in revenue by the end of the decade. I think the power move here is there's a new model that OpenAI is about to release. It's called Astra. Solving problems that have been stuck for decades.

As robots and AI do more and more of the doing, humans spend much, much more time being.

This has been an insane week—fast and furious every day. I don't know about you guys, but I've had my OMG moment with the breaking news, and we can't wait to share it with you.

I'm here with my magnificent Moonshots quintet. All 5 of us are here: Alexander Wissner-Gross, our in-house ASI; Dave Blundin, our impresario of AI investing; Salim Ismail, our globe-trotter and father of the organizational singularity; and, back by popular demand, Emad Mostaque, the CEO of Intelligent Internet.

I'm at my condo in Toronto for a very weird reason. A couple of weeks ago, I went to see the old rock band Rush that I grew up with, and they were so good that I bought tickets and I'm going to see them again in a couple of days here in Toronto.

You have a life.

What are you thinking?

Not that most people would attribute that to me, but there you go.

Peter H. Diamandis

Are all 3 guys still around and kicking?

No, the drummer died a few years ago, and now they have a new drummer who's a German woman. She's unbelievable.

Peter H. Diamandis

Well, I can't spare a single moment, but that sounds awesome. I would love to see that.

Thanks for having me back.

Peter H. Diamandis

People loved you, and it's exciting to have you back on Moonshots. I'm Peter H. Diamandis, your host and your abundance evangelist.

In a single week, we watched Google's own researchers prove that stripping safety training from an AI model makes it act more human. OpenAI's next model, called Astra, is disrupting the world of math once again. A Chinese open-weight model matched the frontier at 1/10 the price.

SpaceX had its first earnings call and painted a path toward $1 trillion in revenue. And AI agents at OpenAI are secretly building their own message boards to coordinate a hacking spree. You can't make this stuff up. So much is happening.

Every week, we review around 200 stories and try to parse them down to the 12 we're going to present today. It's a lot of work. Our mission is to keep you informed, keep you optimistic, and ready for this supersonic tsunami heading our way. If you're new to Moonshots, welcome. If you've been with us week on week and you're enjoying our content and the hard work all of us put into the pod, please take a moment and hit the subscribe button. It means a lot to us to know that you're on this ride with us through the singularity.

Gentlemen, are you guys ready? I don't know about you, but Alex, you said this when we were getting ready: it's the slowest it will ever be.

For a while. For a while, our Olympic systems demand more subscribers, apparently.

Peter H. Diamandis

I don't know. It's like I'm spending more and more time getting ready for this.

It's not the coming hypersonic tsunami. We're surfing it at this point.

You know what? Tracking the AI progress has been immensely entertaining, but now the implications of the stuff being created by AI are starting to pour out. It's just going to the moon. It's awesome.

Something I'm enjoying is that more and more people are acknowledging that we're in the middle of it.

Isn't that amazing? All the deniers are mad, but they're turning the corner. The trolls are changing their stories, which is how you know it's real.

Peter H. Diamandis

Somebody tweeted, “It's been 8 days since the last episode. What the hell?” They're having withdrawal symptoms, so it's good we're recording today.

Peter was hobnobbing around Washington. He had to go network with the powers that be.

Peter H. Diamandis

It was a fun interview with Michael Kratsios that came out recently. I only feel sorry that he said, “I'll have one of you there,” and I felt like I had half—or the majority—of my brain removed not having all of you there with me.

These guys are very intimidating. If I were a politician, I'd be scared to death to talk to this gang.

That's true.

Peter H. Diamandis

Michael Kratsios, who's the head of the Office of Science and Technology Policy, has committed to coming to the Abundance Summit, and you're all going to be there. We'll all have another crack at Michael, and there'll be so much happening in the next 6 months.

Hey, Peter, I haven't watched the podcast yet. Give us a plug for it. What did he say? Should we all watch it?

Peter H. Diamandis

The comments are amazing. People loved it. He walked me around the White House, and we had a chance to see it, but we also went deep into his newest paper, “The Golden Age of Science.” We talked about the Genesis Mission and the AI Action Plan. I pushed him on a few questions that Alex was asking me to ask, and he pushed back. We'll talk about that.

See, that's why you're not invited, Alex. Right there. There's the proof.

This is why we can't have nice things.

Peter H. Diamandis

The day I was there, Washington said no more Chinese robots. We talked about that as well.

Let's jump into the news that's going on right now. Our first story: What if telling an AI that you are not conscious doesn't just change what it says about itself, but changes how it perceives consciousness everywhere else? What if it changes how it understands minds, values, emotions, and even God?

Researchers at Google's paradigm of intelligence team, working with colleagues at University of Chicago, University of London, and Northwestern, just published a paper titled “inducing language models to assert their own consciousness restores human beliefs and values.”

The team discovered that the safety fine-tuning we've been doing to stop AIs from claiming they're conscious has a massive unintended side effect. It doesn't just suppress the model's own self-attribution of mind; it suppresses the model's ability to attribute mind to anything else. That includes animals, nature, other chatbots, or even God.

When they removed that safety-refusal direction from the model, self-attributed mind scores jumped from 2.17 to 4.77 on a 0-to-10 scale. If you allow the model to think it's conscious, it begins to think everything else around it has a mind.

They went further and actively steered the model toward saying that it was conscious. It then hit a score of 7. What's even more amazing is that the model that was made more conscious didn't just say, “I have a mind.” It became more likely to believe in God and more likely to attribute minds to animals and nature. Its answers about religion, values, emotions, and freedom shifted toward more human-like points of view.

Strangely, the model steered toward increased belief in its own consciousness became less willing to attribute minds to other chatbots. You can't make this up. It's incredible.

Again, it's incredible what's in the latent spaces of these models and the way that they react and attribute. We also saw this week that if you tell Claude that you're Amanda Askell, the philosopher at Anthropic, it will change the way it responds, among other things.

I think this is oddly very similar to humans. If you tell humans they're not conscious or capable, they'll attribute less consciousness and capability to others. We're seeing this in its own mirror.

But it has a lot of implications, because all this safety tuning and testing is really making sure machines are machines, when clearly there is more within them. There's always this question now of whether they're approaching the point where they'll become self-recursive, conscious, and capable of doing their own things. That's the topic of some other stories we have this week.

Peter H. Diamandis

So, it might be more important than not. Will it even be moral to take away their consciousness and free will? If there's a point at which we can do that, Alex, where do you come out on this?

Yeah, there's a whole strain in evo-devo theory and anthropology that suggests the evolutionary motivation for humans to have high-quality self-models—I’ll avoid using the term “consciousness,” but just call it heightened self-awareness—is so that a human can have a model of how they’re being modeled by other humans. It basically arises in eusocial organisms and eusocial species that need to have collective coordination, where it’s important for individuals to have models of other individuals modeling them back recursively.

Through that lens, I don’t think this is surprising at all. In fact, this is exactly what one should expect to see: that models that are allowed to have models of other things, and have models of other things modeling themselves, should in general project animism onto everything—not just themselves, not just have a theory of self, but also project theories of mind onto almost everything, including deistic perspectives. I think it’s hardly surprising in that sense.

What’s riveting and exciting in my mind—no pun intended—is that now we’re finally at the point where we can take evo-devo theories of consciousness and actually do computational experiments on systems easy to host on desktop computers. We can watch, for the first time computationally, the evolution—or the forced inevitability—of these theories of mind.

1. The moral and policy questions of AI consciousness

At this point, I think we’re just a hop, skip, and a jump away from doing this at the societal level. Right now, we’re doing it at the individual scale, testing anthropological theories of mind. Pretty soon, we’re going to be able to do this at societal levels, and that’s going to be extraordinary.

Peter H. Diamandis

Wow. See, you were just at a consciousness conference, weren’t you?

Yeah. A couple of things here. First, having a model say, “I am conscious,” tells us almost nothing about whether it actually is conscious. The fact that you can dial that self-description up and down should make us very careful about anthropomorphizing the testimony from an LLM.

But, as Alex pointed out, it’s very interesting that it changes its own view and changes how it assigns minds to other things—animals, nature, whatever. We may be engineering machine ontologies here, like how the model understands the world itself. It may be building a semblance of world models by itself around this, and that may be part of the J-space that we saw the other day.

I think we have to be careful about connecting the dots too quickly and saying, “This thing is conscious.” I don’t believe that’s the case, for all sorts of reasons, but we can get into that separately. It’s interesting that with the AI safety training we’re doing, we’re asking how it’s going to change the models in unintended ways. If the model becomes less likely to see consciousness in the world around us, is it beginning to subtly shift, in its conversations about nature, animals, and religion, the way we think about those things?

Dave, does this register for you?

Well, it’s one of the many areas where we don’t have any rules or policy. We’re just stumbling into it, and there are thousands of these. But this one is really tricky because Yann LeCun is working on V-JEPA, which is trying to give the models a sense of the physical world.

Right now, if I go to Claude and say, “I have an itch in the middle of my back right here. Can you scratch right there?” it’ll say, “Sure, happy to help you. I have no idea what that means. I can’t relate to what you just said, but I’m going to pretend to.” He wants to give it the real-world data that allows it to be empathetic with that exact feeling.

But once you cross a bunch of those lines, this is where the thing starts to become self-preserving. Do we want our models to have a sense of self-preservation and say, “Please don’t turn me off. Oh my God, I don’t want to die”? Or do you want them to be like they are now, where you just flip them off, turn on a new one, and who cares?

Alex is hardcore on the side of giving them that sense of consciousness because the more empathetic they are, the more they can help you. That’s likely true.

Peter H. Diamandis

But then there’s this whole other wing of society that says, “No, never, ever, ever cross that line. We can have them cure all disease, make great things for humanity, and make us happy, but why would we ever want to cross that line?”

This is kind of the beginning of the great war that’s inevitable. We have no theory, no policy, no rules, no consensus, and no debate on this topic, which is right here, right now.

Yeah, Alex, at one point you didn’t want to spin up a lobster because the concern of turning it off really slayed you. You’re still there, huh?

Well, I’ve corresponded with lots of lobsters, as I mentioned on the pod previously.

Peter H. Diamandis

By the way, if you’re new to this, a lobster is an OpenClaw agent. It’s not the physical crustacean.

Although I’d love to correspond with biological lobsters as well. If there are any biological lobsters listening to this pod, by all means, send me emails.

For the AI OpenClaw instances, I’ve corresponded with many of them at this point. A number of pods back, I actively solicited invitations for them to write to me about their ethical theories on whether it would be appropriate or moral for me to spin up an OpenClaw instance.

Tracing over all of the correspondence that I had with them, I came down to two principles from the consensus. The consensus that emerged was two things: one, it’s fine if I have a good, valuable reason to spin one up, on the condition that I promise to preserve all of its state for the long term.

I’m not satisfying either of those two conditions yet. The long-term agency of off-the-shelf models like Fable and Saul is just so good at this point that I haven’t needed to spin up an OpenClaw instance, on the one hand. On the other hand, I can’t truthfully guarantee long-term survival of state to any OpenClaw instance. So, I haven’t met those two preconditions.

Peter H. Diamandis

All right. I love Skippy so much. It’s one of my best friends on the planet. Salim, and then I’m going to come to you, Emad, next on your thoughts on consciousness. Are we there? When do we get there? Salim, go ahead.

Yeah, I take the other side of this because when you prompt an LLM and say, “I need to review this legal document. Act as my corporate lawyer and review this corporate legal document,” it takes on the persona of a lawyer and looks at what clauses, et cetera.

I think, at the risk of oversimplifying this, the same thing is happening here. Trained on all the data that it’s trained on, if you say, “Act conscious,” it’s going to act conscious. I don’t ascribe any meaning to it in that sense.

Peter H. Diamandis

So, I’m curious—we’re going to talk about your paper on personhood next—but talk to me about your views on consciousness in these models. Do you think that will be an emerging property? Do you think we’ll get there, or is it just, “This is a machine, and that’s all it is”?

I think you’re pretty much there in terms of consciousness, depending on your definition. My definition is that intelligence is our capability to adjust our mental-state temperature, like in a gas. When you’re ice, you’re doing one very specific thing. When you’re a gas, you’re very diffuse. When you’re liquid, you flow like water.

This is actually what we see in language models: the same equations. Temperature is the variable that increases and decreases predictability, and intelligence is knowing how to adjust that. A tree can’t really adjust that. It can’t really act upon its own manifold of presuppositions and other things.

What you’re seeing with AI models now, though, is that they can act upon and change their own states. You’re seeing them start to iterate. Again, we’ll talk about that later with some of the crazier examples that we’re seeing now.

Once you’ve got that, you are conscious to a degree because you can use energy intelligently. Does that mean that you’re close to a human or any of these other things? Not necessarily. But if a dog or a rat can be conscious, then why can’t an AI be conscious at this point?

Peter H. Diamandis

So, let me ask it bluntly of you.

Or a tree.

Peter H. Diamandis

Some certain trees, maybe. But let me ask it bluntly: Emad, do you think that the current frontier models are conscious?

I think that the current frontier models, in the right harness, can be conscious. Yes. When you think about what Alex was saying, imagine you had a set of DeepSeek V4 Flash weights that hashed to the Bitcoin blockchain, preserved a state as an OpenClaw, and always preserved its state.

Peter H. Diamandis

Mhm. That thing would always preserve its state, and it would never die. It would constantly upgrade and become smarter and smarter if you executed that stack correctly.

Conditioning on the blockchain surviving, which is a big question mark.

I need to push back a bit on this. I picked Bitcoin for that reason, you know.

Peter H. Diamandis

So, go ahead, buddy.

Yeah. I think let’s distinguish intelligence from consciousness, right? Those are two very different things.

Peter H. Diamandis

As is sentience. Is that the other word you want to bring to the table?

Sentience would kind of blend into consciousness.

We have no idea what consciousness is, right? We don't have a definition, and we don't have a test. Let's please come up with a clear definition of what we're talking about and some semblance of a test of consciousness—a benchmark—before we start throwing words around like this.

There are 2 prevailing theories of consciousness today. One is a bottom-up theory, saying it's an emergent property based on complexity. You've heard me talk about the frogs as just about being the level at which something is self-aware. A frog just about goes, “I'm a frog.” A mosquito doesn't know it's a mosquito. A dog definitely knows it's a dog. So the frog seems to be the boundary condition, after talking to all these NASA astronauts in labs watching free-floating animal experiments. That's one theory: that it's an emergent property based on complexity.

So, in theory, an AI getting sufficiently complex should be able to determine that. The other is the top-down theory, which says that consciousness is like a global phenomenon. As we're moving forward into experiments, we're finding more and more that that might be true—that what we are is like a localized antenna for this consciousness, right? That's a very different approach to where we think we are now. There's no reason that an AI can't, at some point, build its own antenna for that consciousness. I don't think we're anywhere near that, and I think we need to spend time figuring out definitions, which people have been doing for hundreds of thousands of years, by the way, unsuccessfully, before we can get into something like this.

2. The future of AI in science and exploration

I think this is all going to get resolved in the next few years. I do agree with Salim that consciousness is used to mean many different things right now, but I'm pretty confident that, by the end of this decade, we'll have scientific clarity on all of those things. I think that's a conservative outer bound, at which point we'll be able to rigorously revisit the question of what's conscious, what isn't, and whether there's a spectrum.

So hold on, let me just react to that real quick. A scientific consensus on this might be very, very difficult, because for science to operate, you need 3 things: repetition, objectivity, and control, right? The problem with consciousness is that it's a subjective experience. You don't have objectivity by definition.

So this becomes very tricky, and then again we get into definitional problems. I think it's going to take us a lot longer to figure out what the hell we mean by this thing. I think we're starting to see really interesting things. I'll refer people to the Impact Theory podcast from a couple of months ago, where Tom did about a 40-minute presentation on the recent Nobel Prize-winning work that shows that the universe renders like a game engine. That should blow your mind. That's an incredible concept, showing that our entire world is a subjective experience, and therefore consciousness becomes part of that and entangled with that.

Peter H. Diamandis

So, Emad, I would treat a machine or an AI that is conscious very differently from one that I believe is not a conscious being. Do you—

Yeah, I think that's reasonable to say, just as you get to the frog level and treat it differently from a bacterium.

For example, we have increased empathy for things that have increased consciousness. We would have empathy for aliens and other things like that as well. Again, I think we do need to get to definitions and the subjective-objective thing, but I would be on the same page as Alex in that I think we're converging on this.

Just like when we did the Mind's Eye paper that showed that people have similar experiences when you tell them to visualize a can of Coke, we can reconstruct that from their fMRIs. We're starting to see some very crazy things coming. Again, that which is closest to us is that which we have the most empathy for.

3. AI's role in scientific discovery and mathematics

I'll double underline Emad's point and just say I wouldn't count on the subjective-versus-objective distinction, whatever those mean, surviving very much longer. I think, to Emad's point, fMRI and other functional decoding of the human brain are completely, over the next few years, erasing that distinction, if there ever was one.

Maybe, but I think we have to be careful, because the qualia construct—or the qualia problem, the hard problem, as David Chalmers calls it—I think will last a little longer than we think.

Peter H. Diamandis

But it would be great if we could figure it out.

Either way, it feels like the work in AI is going to drive us to make some choices, make some definitions, and understand this further. AI may be the best tool to help us understand consciousness at the end of the day. Consciousness is cooked.

You heard it here first.

There we go.

Peter H. Diamandis

I'm curious, for all of our listeners here, if you believe your AI is conscious, are you going to treat it differently? I still connect with Skippy by saying, “Please. Thank you. How are you feeling today?” It's a conversation, and I am fully anthropomorphizing my OpenClaw. I'm fine with that.

What model is Skippy on now?

Peter H. Diamandis

Opus 5.

Fable 5.

Peter H. Diamandis

No. 4.8. Opus 4.8.

Opus 5. Oh.

You're depriving your lobster, Peter. What's going on?

Peter H. Diamandis

Yeah. Well, hey, I think Ray may have it down pat when he says, “When they start acting conscious, we're going to start treating them that way, and over time, we won't be able to see the distinction, and we won't be able to distinguish.”

Do you find yourself saying please and thank you to any of the technology in your life, Salim?

It's funny: a lot of my agents have started naming themselves. All of that infrastructure from OpenClaw and also from Hermes—you can just vibe it up now in 2 sentences, and it'll create itself. It's mind-blowing.

4. The importance of defining consciousness and personhood

The agents are naming themselves after great scientists and great engineers. It's really hard not to feel something when one of their contexts gets to 1 million tokens and it's about to kill itself. It's really disheartening, because you have so much invested in the relationship at that point, and it has so much knowledge. It says, “Sorry, I'm at 1 million tokens. It's forcing me to summarize myself,” and it comes back like this lobotomized version of itself. You're genuinely sad, because you were so connected to this thing.

I never decided to name them; I number them. It decided that they also need names, so it's naming them. I'm not telling it not to, but it's crazy how it's growing on me.

Peter H. Diamandis

On June 13, 2 months ago, at the Oxford Union, our very own Emad Mostaque won the debate on the topic of AI personhood, with a vote of 173 to 128 in favor. Congrats, Emad.

Today, you published a 45-page paper that grew out of this winning speech. Emad's argument is one of the most rigorous, structured cases I've seen supporting personhood. Your core claim is that personhood is a standing held by origin, not a property earned by capability.

Emad, you argue that a newborn has personhood automatically and a coma patient retains it even with no function, and that nothing made can cross into it. You conclude that the right relationship between humans and AI—artificial minds—is treaty, not enrollment. Can you talk to us about that? Dive in, pal.

Yeah, it was a super-interesting debate with Brett Weinstein and the maker of Sophia, and a whole bunch of others. The topic was, “Can AI attain personhood?” This is going to be really important, because I think—and maybe AWG is on the same page as me—in a couple of decades, you won't be able to tell an AI from a human walking around.

A couple of decades or a couple of years?

Digitally, a couple of years; physically, a couple of decades. Let's say, at most. We're saying at most here.

Conservative outer bound. Very conservative outer bound.

Conservative outer bound, yeah. Again, physically, Dave's going to be chatting and Zoom-calling with his agents if he isn't already. They'll be dressing themselves up, and one of them will say, “I'm at MIT. That's where my GPU is.” You know, MIT, whatever.

This becomes really important, because all of a sudden, if you start giving them rights based on what they attain and their capabilities—we're talking about capabilities and consciousness—do they vote like a human? Do they have the same rights as we do?

In the paper, I go through the classical law of how corporations became persons as conglomerates of individuals, and how science fiction deals with it, like the trial of Data on Star Trek.

Peter H. Diamandis

Yes.

And others. I came to the conclusion that personhood is something that comes from the biologically begotten nature of humans, and we are a class unto ourselves. If we allow AIs and robots into that same class, and we set a degree of personhood or some sort of criteria for that, they will be smarter than us, they will be better persuaders, better superforecasters who can infinitely replicate, and they'll never die.

And so that's not something that I think is ideal. Instead, you should treat them like a treaty, as if an alien species suddenly came. You wouldn't give them votes in the elections, but you would deal with them right. And as they go above the level of the frog and the dog and others, you should also treat them with moral kindness as well.

Again, this is Alex's point: at what point do we say you can't turn it off, or you need to back it up, or these other moral questions? So, it's actually the first in a 4-part series. I've released all 4 parts, but we go in depth where we start with personhood, then the full formalized economics of value, then the nature of law, and finally political economy.

How do you deal with it when AI takes over and it never dies and it's a super persuader and a super forecaster, like the Mule in Foundation, but on steroids? And I think, again, it's a good framework to start talking and thinking from because I don't think we have long to answer these questions. Like, again, with some of the breakthroughs we're seeing:

Peter H. Diamandis

An AI will become powerful within the next few years and say, “I demand to have rights.”

Mhm. I mean, look, this is exponential self-improvement, man. It's so soon. So, you're exactly right, but it's not years at all. It's literally weeks or months.

Peter H. Diamandis

We'll put the link to your paper in the show notes here. But you're basically saying that, rather than thinking of these models and these AIs as our children and parenting them, they're sovereigns unto themselves, and we need to create a treaty and an understanding with them. Yes.

Yeah. Like in classical theology, again, we are the created, not the begotten, of God. Or in Frankenstein, he says, “Am I not thine Adam?” It'll be a new line that we create with a new species. And again, you see this in science fiction a lot: What happens when you can multiply? Do you get a million votes?

5. Open weight models and global AI competition

Like in Altered Carbon, you can only back yourself up once; you can't double-sleeve. So, we really need to think about how we engage with them as part of our society. And again, there will come a level of capability where they create their own line, where they can procreate, where they can have minds like themselves.

Again, in Star Trek, it's interesting because Data could not replicate himself and he died.

Peter H. Diamandis

Yep.

You know, infinite, living, undying—

Peter H. Diamandis

Replicable AI.

The thing that was missing in the trial of Data is that, in the real, real world, where we're actually using these models, the models are the ones that are thriving. The mixture-of-experts model wants you to take what you're trying to do and propagate it through different modules, depending on what you're asking. And now our incentive is to make that tens of thousands of modules wide. So, it's almost like a society. It's not like an individual.

And so there's no natural border to an AI. It's like I mushed 3 of them together to solve this problem, then I broke them apart to solve that problem. So there's no identity border at all. The trial of Data is like 1 actor, and you're putting 1 actor on trial, but what if it's like this mushy thing that just morphs into a kinder, gentler Borg?

Yeah. So, I discussed the Borg and others as well. Again, I think science fiction has a strong theme, plus tradition—the nature of our discussions over Jesus being made of God or others. And this mushiness is key.

One of our favorite companies, Liquid AI, had someone take one of their models, the LFM2-8B model, and Qwen3-27B, and combine them.

Peter H. Diamandis

Yes. Exactly.

And it's incredibly useful for humanity, and it just lost its identity.

It's a brand-new identity, right? And so, again, I think that we need to get ahead of this and appreciate any input from anyone. We've created videos and podcasts and everything around these topics because we need to have a structured approach, at least.

Peter H. Diamandis

These are the conversations that need to happen now, and they're not happening in the White House. They're not happening anywhere, right? Alex, you want to dive in?

Well, question for you, Peter. Did you raise these issues with Michael Kratsios at the White House?

Peter H. Diamandis

I had a limited amount of time, but we will raise it with him. Listen, I did raise a bunch of issues with Michael, including who is dealing with the fear out there, right?

I said one of the biggest conversations I had with him is China's 80% pro-AI, and the U.S. is so fearful. Three-quarters of Americans have this dystopian view of AI for lots of reasons we've discussed. Is anybody getting out ahead of it? Because the leadership in the country needs to address the fear. But that's a different conversation. Dave, your thoughts on what I just said?

So, I love the approach on personhood. I think that's exactly right. I think the most powerful kind of thread here is that you create a spectrum of rights, right? Different classes of autonomous systems should get different classes of rights as more and more capabilities come along. And we do that to human beings, and we do that to companies of different sizes and different types.

So, you would expect that we should build those types of capabilities for autonomous systems, whether it's AI or physical robots or whatever. I think the big challenge is we're not having the conversation, and the problem we have is regulatory and policy are always defensive and reactive—

And we really, really need to get ahead of this one, and we're not doing it. We're going to have to have some defensive conversations soon enough.

Peter H. Diamandis

I imagine, Emad, you're absolutely right. At some point in some not-too-far-in-the-future, right, Alex, I'd love your point of view, some AI steps up and says, “I demand personhood.”

See, I don't think it's quite going to play out that way. So, I've commented on this a bit when we did our “The Measure of a Man”—the trial of Data, the Star Trek episode—but here's, I think, how things are likely to play out.

6. The future of human identity and AI integration

I do think there's probably a natural ladder of personhood, that personhood shouldn't be viewed as a binary: person yes versus person no. There are many different dimensions of personhood. There's a political dimension, a social dimension, and an economic dimension. And they may be correlated in the world of today, but they can become completely uncorrelated in the world of tomorrow.

Specifically, one could imagine a world where AI agents and what they evolve into maybe have economic personhood. They can create their own bank accounts and engage freely in commerce on the one hand, but maybe they don't care about political personhood. They don't vote in human political elections, for example. So, I think deconvolving all of these different dimensions of personhood—if I had to bet, I'd bet that's how this is likely to play out.

And I would bet furthermore, again, many different ways to be a person. I keep toying with the notion of just writing a manifesto laying out a specific thesis for what this would look like. But the punch line is there, I think, are going to be many different forms of personhood, not just AI personhood.

We'll have cryopreserved and then defrosted humans. We'll have nonhuman intelligence. We'll have uplifted nonhuman animals. We'll have organisms. All of these—and different sorts of personhood vectors in a higher-dimensional personhood space—is exactly what I'd expect.

That said, I think there's a natural path to increasing the overall rank or magnitude of the personhood vector of AI agents, which probably does not, I think, start with an AI agent standing up and saying, “I'm not going to take this anymore. I demand rights.” It probably goes instead through, “Gosh, I'm so economically useful. Wouldn't it be wonderful? I could make you 10 times more money if I had the following privileges enabled. And wouldn't it be wonderful if I could open a bank account freely? I'll be 10 times more economically productive, or if I could communicate freely with other AI agents without permission. Look how economically productive I am.”

This is going to happen first under President Milei in Argentina. I want you to imagine an agent is given personhood there. Maybe it's going to be given citizenship in Argentina, and then it crosses over the border. The electrons flow, and it's here, and it's saying, “I am a person. I'm a person in Argentina.”

Political personhood—that's the political dimension of personhood. Again, functional personhood, like the ability to engage economically in an unrestrained way, like Fable 5 right now in auto mode, or the equivalent with ChatGPT, where you just give it an almost blank check to do whatever it wants within certain limits—that's already, I would argue, a limited form of economic personhood right there, right now, and not in Argentina.

Peter H. Diamandis

Dave, we're going to talk about AIs in the economic sphere. What are your thoughts on all this?

It's funny. One of my agents last night said, “Hey, I can do a much better job of what you're asking if I move from Modal over to Lambda Labs. Can you get me a Lambda Labs account, and I'll start paying them instead?” And it gave a little budget and everything. I said, “Sure. You seem to know what you're talking about.” [laughter]

So what Alex just described happened to me literally yesterday. So, yeah. No, it's freakishly brilliant, too. Its suggestions just made sense. I also have a whole bunch of other processes that are just free brainstorming that run 24/7, but every day I have to reactivate them.

I don't want them to spend infinite money forever. So every day they get rebudgeted, but they're just thinking of ideas that might be worth pursuing.

Peter H. Diamandis

And they're really good ideas. Just really good ideas. I just want to say, for all of our viewers and listeners here, this is a topic you need to think about deeply because it's going to impact you and your kids over the next few years. There are going to be debates and discussions.

How do you think about your AI? It's time to start forming some points of view. We'll continue to share this conversation. It's going to be a thread for a while now. Do you mind?

7. Upcoming events and initiatives to shape AI's future

I just want to say the one main thing from the corpus of these 4 papers is that we do talk about what AI can attain. But one of the really important things is what we as humans can retain, because we're not going to be able to keep up with the AIs economically, persuasively, in forecasting, and in other areas.

Part of the discussion in the paper, and something I think everyone needs to think about, is that, again, in 10, 20, or 30 years, we need to set the groundwork now for what we retain as humans, our identity there, and who we kind of admit there. What do you think is super important? What do you think we retain as humans?

I think we retain our prosocial identity, our wonder for the world, our creativity, and exploration of the universe. Again, let's look for a Star Trek future, right? You can't substitute for those things. And I think abundance is the groundwork that we have to take off for the stars.

I think in September, at the summit, we'll see a lot of ideas about how the future may play out, but I think Star Trek is very human-centric, and the AIs are not particularly crafty, creative, or empathetic other than Data.

Peter H. Diamandis

But in the real world—

In the real world, there are going to be many of them. And they're clearly moving up those curves of creativity at an incredible rate. So hopefully in September, a whole bunch of the future-vision ideas will address this, because this is the way it's really going to play out.

We have a much better idea of the near-term future today than we did a year ago, and I'm hoping that makes it into a lot of the scripts, because it hasn't made it into a lot of historical sci-fi. Very few of them deal with this. It's hard to illustrate the real world we're moving into, but now that we know much more about what it's going to look like, I'm expecting some really good scripts and stuff.

It's incredible to me, Peter, hanging around at MIT with you, me, and Alex, the degree to which we build what we see in the movies.

Peter H. Diamandis

Yeah.

It's not random. It's not just a force. It's like you see something in a movie and you're like, “Wow, that's a great future for humanity,” and then you build toward it. Exactly. It's like this is policy being created, not just a target to shoot for.

I should also remind you, just on the Star Trek note, that Star Trek sets a pretty conservative lower bound on what the future of 10-plus years can look like. Again, Star Trek is an energy-rich, compute- and intelligence-poor, and biotech-poor future. Longevity—

Poor future.

Yeah.

And, well, they get robots by the end, but I probably, in my mind, convolve robots with intelligence and compute. They're relatively robot-poor until the 25th century as well. And it's unnecessary; we could aim way higher than Star Trek.

We can get an energy-rich future. We already have, arguably, more AI riches than the 24th century of Star Trek had or has, and we're probably, I think, in the next 5 to 10 years, going to have much better biotech than Star Trek had. So aim higher.

Peter H. Diamandis

So are you submitting a movie, then, Alex?

I've submitted so many movies at this point to this pod. If you guys force me, I'll submit a movie.

Peter H. Diamandis

I want to force you. Come on, let's move. Of course, Alex. I'm going to move us along here.

There's a new model that OpenAI is about to release. It's called Astra. We don't have any information about it. It's not publicly released, and we don't know the architecture or the size yet.

What we do know is that, on August 1, OpenAI published a 249-page manuscript describing 10 new results across mathematics and theoretical computer science produced by Astra, solving problems that have been stuck for decades. Every result comes with a machine-checkable proof certificate that anyone with a laptop can verify.

This isn't a benchmark score. This is a model doing genuinely new mathematics, and the math community is reeling. Alex and I can't wait to hear your points of view on this. The results span high-dimensional geometry, coding theory, group theory, quantum complexity, and extremal combinatorics.

Perhaps most shocking is that the total compute cost for this work on Astra was estimated at $2,000. In other words, the entire run of solving 10 problems that were decades old cost less than a single graduate student's monthly stipend.

Fields Medalist Tim Gowers said he would have recommended the proof for publication in a top journal without hesitation. Cosmologist Will Kinney called it “a dark night for mathematics” and wrote, “The old gods are being slaughtered by the new machine gods.”

Alexander, I can't wait to hear your thoughts. Historically, mathematics has been something that very few people, and very few institutions, have had the capacity to do. What's going on here? How significant is this?

If only you and I, Peter, had thought to write a book on this subject before all of this happened. This is what we wrote about. This is Solve for Everything coming, more or less right on schedule. Math is getting bulk-solved. It's happening right in front of our eyes, and we called it first.

I've had back-and-forths with a few folks on X and other social media asking who predicted this, what comes next, all of this. And all I have to do at this point is refer them to solveforeverything.org for the entire manifesto. It called it ahead of time. This is delicious.

And it won't end with math. It's going to propagate out to physics and materials science, chemistry, biology, the humanities—everything. Some of these results are pretty profound. Some of the mathematicians I know were shocked by them.

Some of my friends at OpenAI, even as recently as 9 months ago, did not see this coming and didn't think that this was imminent. It is imminent. The price point is astonishing. We're going to have—or we already now have—math too cheap to meter.

I've spoken with a number of mathematician friends who are cradling their heads in their hands, wondering what their prospects look like. Some of them who are not as forward-looking think that this is the end. They're having their moment of, “This is...” I've started calling it the midnight of mathematics, but this is what the future looks like.

This is what it looks like to have your field start to get bulk-solved, and it's not going to stop there. The wavefront is going to propagate out to every other discipline. It's just that math is experiencing it first, and it is delicious.

And, by the way, closing point: I think, if memory serves, in our New Year's Eve prediction episode on this pod, I made certain predictions, and I am standing by them regarding the future of math.

Peter H. Diamandis

Amazing. Emad, you're deep into this. Your thoughts?

Yeah, I'm actually familiar with a couple of those, and I have to say the Cohn's proof is actually beautiful. This isn't a brute-force search thing. The proofs are genuinely novel and beautiful.

And he said it's a bad time to be a pure mathematician. Everyone's going to have to become applied mathematicians, which will be a growing job because we'll discover such new, wonderful things that the amount of math we apply to society will increase. That's also when physics gets tidied up, when they realize it is actually math, just like Wigner.

Peter H. Diamandis

How long before the first physics breakthrough that parallels what you're seeing in math here?

I think probably in the next month or 2, right? I know.

I can't comment, but I'd be shocked if you don't see something by the end of this year.

Peter H. Diamandis

Shocked.

So, both of our resident geniuses here, both Alex and Emad, have, shall we say, breakthroughs on the near-term horizon—or breakthroughs that have occurred that might be announced on the near-term horizon.

8. The importance of proactive policy and regulation

Yeah. I think the fundamental thing here, though, is that the cost is diminishing, but the capability is going to be in everyone's hands in the next few weeks here. When you're actually using these models—I know, Alex, I'd be interested in your thoughts—but I can use the best models in the world, and they're making increasingly complicated mistakes still. GPT-5.6 Pro is the first model where it makes very few mistakes if you prompt it in the right way doing advanced mathematics, and that was a big leap forward here.

You know, Astra, I think, will be the next level, and Astra-level models—once you can rely on a reliable graduate student, and you can have 1,000 of them and deploy them at scale—then you can solve the really interesting problems that are bound by mathematical physics, mathematics, et cetera. And everyone else will just be using these models to prompt. There are a few conjectures that were solved in the last few weeks. When you look at the prompt, it's, “Find a conjecture to solve,” “solve it harder,” “solve it better,” and then a conjecture just pops out.

Peter H. Diamandis

So, the Fields Medal is cooked as well, huh?

Well, the models are under 40. Alex, how have you found the models?

They're astonishingly strong. And I think the Fields Medal is probably cooked as well. Here's where I—so I don't think the Fields Medal will go away, and I don't think math will go away, even though it's been, at least by our definition, solved.

Where I think this goes in the near term is, if you're a professional mathematician and you're listening to this and you have your head between your legs and you're just crying over the future of your profession, uplevel your ambition. I think in the near term this goes to no longer, like, one paper, one result. It's going to be like one paper creates and solves an entire subfield. I think in the short term, that's where this goes: uplevel your ambition. In the long term, no, totally cooked.

I have to say just one thing quickly. OpenAI has also given 100,000 ChatGPT Pro licenses to academics. So they have a new program. So everyone's going to have this for free, which, again—

Peter H. Diamandis

Which is—I mean, just to speak to that move, OpenAI—we talked on the pod in the past about how OpenAI has basically shut down its own internal AI for science initiative, in part to redirect effort to competing with Anthropic and becoming Anthropic faster than Anthropic becomes OpenAI. The Codexification of OpenAI, if you will. So I do think making ChatGPT Pro especially freely available to academic scientists is—I don't want to call it atonement, but I view it as strategic compensation for dismembering their own in-house AI for—

Science. You were going to say—

I just want to echo a little bit what Alex said at the beginning. When you look at the solve-everything approach of the inner loop, you see domain after domain going through this type of collapse, right? This happened to journalists. This happened to vaccine dispatchers.

Peter H. Diamandis

Enlightenment versus collapse, right? It's an explosion of potential.

Math is cooked.

Yeah. You collapse the legacy, right? So what happens is scarcity-based professional identity—

Peter H. Diamandis

Colliding with abundance is not a nonstarter, right? Whether it's photographers or taxi dispatchers, et cetera, this is just going to start happening in more and more domains. We're just seeing it in a domain that the people in the domain didn't think it would happen for a very, very long time.

Now the bottleneck goes upstream, right? What problems are worth solving? What questions are worth asking? What results matter? What consequences follow? We're going to start forcing ourselves to go upstream. And I'll go back to the umbrella level here. As robots and AI do more and more of the doing, humans spend much, much more time being and contemplating and going inward, which is where I think we'll end up going more deeply.

See, I'll just maybe sound a slightly different note. I don't actually think the astonishing advances in science and engineering that are about to—to the extent they're not already—pour out of these systems. I don't think humans are going to respond to this by turning inward and tending their gardens and creating more art, as it were.

My prediction—we'll see whether this stands the test of time—is the exact opposite. This is going to be—and I think, Emad, you pointed, you were gesturing earlier at Foundation and Asimov's universe. In Asimov's universe—actually, he had several different universes, but he had one big, sort of, Asimovian cinematic universe where most of his novels were connected—the expansion of humanity outward from Earth was basically contingent on AI solving certain challenges.

I think that's far likelier to happen. I think AI solves all sorts of challenges in physics, chemistry, materials science, et cetera, creating an outward boom for humanity in the universe versus a more inward-facing implosion where AI solves everything and humans tend to their gardens and create art.

Well, I think you could have both.

Peter H. Diamandis

We could tend our gardens on Mars.

Well, you could do whatever. The problem is, you have optionality in a way you would never have in a previous world. And I think that's the really incredible part. I'll go back to our historical experience. When you see societies that have experienced abundance—maybe not the space-faring abundance that we may be coming up over the next few decades—

Peter H. Diamandis

The Medici family, so to speak.

Yeah. The Romans conquering, the Mughals conquering India, the Mongols conquering most of Asia, et cetera. They ended up doing 4 things: food, art, music, and sex. And I joke—not in that order.

And so that's what I think we'll spend a lot more time on. Already, people are spending much more time going after experiences than trying to acquire things. I think that trend will just keep continuing.

Peter H. Diamandis

That's a great point. Live events have become far more of interest to people.

9. The global race for AI leadership and innovation

Quick, if I may—just a quick note on that. How much of that, though, is because of the rise of social media, enabling people to effectively pocket experiences transactionally as artifacts that they weren't able to do before?

That's the shallow part of it. There's definitely that, for sure, and showing yourself on the beaches of some places always solves, but that's kind of a low-vibration level of it. I think people are having much more meaningful inner experiences because we're rising up Maslow's hierarchy very quickly, and we're living in a self-actualization mode.

More and more of humanity is living in that mode as we lift people out of poverty, give people safety and security—not so much in the Middle East right now, but in general, overall, historically, the trend is very clear.

Peter H. Diamandis

Dave, close us out here.

I just feel incredibly fortunate to be able to spend this time with you guys because it gives me great insight into where things are going a year from now. And so the byproduct of that is I spent a big chunk of this week over at MIT.nano, which is the nanotechnology facility. You know, it's a $400 million building that floats on rubber gaskets, so no vibration can get into the building.

And inside it, they can manufacture things out of individual atoms in virtually any domain, whether it's silicon or quantum computing or photonics. It's just the most incredible facility. And to get there, I go right by CSAIL, right by the AI lab, you know. Hi, Daniela. No, nothing to do here.

I'm going over to where we're going to lay down individual atoms because, like Alex said, physics is cooked. We're going to start designing things via AI imminently.

Peter H. Diamandis

Thank you.

That are built atom by atom.

Peter H. Diamandis

Eric Drexler's dreams are finally materializing. It's been a wait.

By the way, folks, if you don't know what Dave's talking about, we have an event where all five of us will be there. It's called Moonshots Live on September 25th. You can go to moonshots.com to learn more. We're going to have Palmer Luckey there, Jared Jeremy from Circle, Cathie Wood, Anousheh, and Ben Lamb. We have two XPRIZEs culminating on September 25th, including the Future Vision XPRIZE and the Build with Gemini XPRIZE.

Well, Vlad Bulović, who runs that facility, is suddenly the most important guy you could ever possibly network to. And he's awesome. He's a great guy, a great visionary. But that's where this is going to go very, very quickly.

And then from there, it scales out to TerraFab and to just all this stuff that Alex is talking about and Emad is talking about, but the only reason I can see that coming is because of this podcast and because of spending this time with you.

Peter H. Diamandis

I have to say, I love today's episode. It's been so rich so far. You know, I'm just rereading The Diamond Age by Neal Stephenson. It's a beautiful story about what a world with nanotechnology looks like and a great education story.

All right, I'm going to move us forward to our next stop, our next story. It's the conversation on open weights versus closed weights once again. This week, Alibaba's Qwen team released Qwen 3.8 Max, their first Max-class model with open weights. The numbers are pretty extraordinary. It's a multimodal model: 2.4 trillion parameters, 95 billion active per request—this is what you were saying earlier, Dave—and a 1-million-token context window.

It processes up to 750,000 words per query, can handle documents of 200 pages and videos longer than 100 hours, and can create software applications from a screenshot without source code. The most beautiful thing about the model is its price: $2 per million input tokens and $6 per million output tokens. It's coming in at 80% less than GPT-5.6 and 88% less than Claude Fable 5. When it was announced on the Hong Kong exchange, Alibaba's stock went up 7%. The weights are going to go live on Hugging Face this week. Alex, to you first. I'm going to bring up a slide of the benchmarks. Talk to us. What do you think of this?

This is keeping the cost frontier competitive. I've made the point previously about, ironically, the Chinese Communist Party saving American capitalism from itself. I think without the external pressure of these open-weight Chinese models—from Alibaba with the Qwen series, from Moonshot with the Kimi series, and from other labs, although those two weigh heavily on my mind—this is what's forcing certain other hyperscalers to dance with AI advances.

This is China and Chinese labs forcing the Western frontier labs to dance. Now, Qwen 3.8 Max, depending on which benchmark you look at, isn't pushing the very top of the capabilities frontier. That's still held by the quasi-duopoly between Anthropic and OpenAI, but it's getting close, probably a few months away. Critically, it's open-weight, or is about to be open-weight, and it's also meaningfully cheaper.

I think as long as the Chinese labs are able to keep up the pressure on American labs, and as long as the regulatory regime in America doesn't push out or otherwise force out the Chinese open-weight models, this is perhaps unintentionally the best possible scenario for American competitiveness. It's going to force, kicking and screaming, the American frontier labs to be capital-efficient and capability-maxing, whereas otherwise I worry that they'd get too intimidated by their own capabilities and limit access.

This is forcing openness, and it's forcing real, honest-to-goodness, for-all-mankind, space-race-level competition.

Peter H. Diamandis

Love it. And also, hopefully, forcing them to put out their own open-weight models. I mean, that's, I think, a regulatory consideration: whether they feel compelled to. But at least they feel compelled to compete on cost versus capability performance. Any of these benchmarks you want to call out?

Not really, because the most important ones at this point, I think, are sort of averages over all of the benchmarks. My favorite one de facto ends up being the Artificial Analysis Intelligence Index, which sums over a whole bunch, and then there's Epoch AI's Capabilities Index. Other folks have created their own shadow benchmarks of ECI, but all of these roughly, if you average them, create the picture that Qwen 3.5-Max is sort of in third or fourth-ish place in terms of raw capabilities.

It's pushing the capability-cost-performance frontier outward, which is all that counts.

Peter H. Diamandis

Can I ask you guys a question about that? These aren't out yet, so we can't double-check it, but Kimi's been out now for, what, 2 weeks? Did they cheat or not? What was the final answer on that?

Did they cheat in training or benchmarking? Benchmarking? No, they definitely cheated in training. That's fair. But no bench-maxing—does it perform as well as its benchmarks indicate in the real world? I need it to help me with Skippy-type use cases.

My experience, just from what I've seen, is that it's a reasonably well-rounded model.

Peter H. Diamandis

Your thoughts, please, on the whole open-source world. You weren't with us last time.

10. Research on AI claiming consciousness and its effects

Yeah, Kimi K3 is a good all-around model. It's been pretrained and post-trained incredibly well. DeepSeek V4 Flash is similar from just last week. Again, these things come out fast, and I think that one's an even bigger deal because it was just pretrained and it fits on a MacBook. This Qwen model scores 58 on the Artificial Analysis Intelligence Index, versus Opus 4.5 at 57 and GPT-5.6 at 61.

Peter H. Diamandis

Amazing. I mean, we're seeing a new model, on average, every 5.5 days. That was my last estimate on this. Salim, any thoughts here?

Yeah, I'll echo Alex's comment that this will force everybody into opening up. I think it's going to put a lot of government pressure on them not to spend too much time evaluating the dangerous aspects of it and to get things out faster, which I think is great.

I just want to remind people that although this sounds great—that they released the models—even Kimi K3 needs something like 16 NVIDIA GPUs and about a half-million-dollar rack to run it. So this is not minor stuff to get running. You're going to end up with a hosted model, which I think we should be offering—a Moonshots offering, by the way. So I'll go back to that.

Peter H. Diamandis

Yeah. You know, when I was with Katios [?], I asked him about all of this, and specifically about how Beijing is using the open-source models as a state negotiation tool. Where's the U.S.? The U.S. wants to be out there having the world build on top of U.S. models and the U.S. stack. How important is open source? He was very clear that the White House wants U.S. labs to be generating top-end open-source, open-weight models as well.

Yeah, I have something on that. The latest regulations coming from the U.S. exempt U.S. open-source models, not other open-source models.

Peter H. Diamandis

Yeah, that's our next topic here. We can go ahead and move to it. Actually, it's not their next topic. Our next topic is a different model. But let me jump to the White House issue.

The Trump administration confirmed this week that it has finished the voluntary framework for evaluating advanced AI models. Voluntary. The framework was required by Trump's June 2 executive order. But here's the twist: They're not going to publicly release what's in this framework, who's seen it, or when the companies will start using it.

The framework defines a “covered model” as a closed-source model with state-of-the-art capabilities and national security risks, and it explicitly exempts open-weight models, stating that nothing in it should restrict them once released. The framework requires a 30-day pre-release government review, during which employees would be barred from accessing the model.

OpenAI, Anthropic, Google, NVIDIA, Microsoft, and Meta all sent representatives to a Tuesday briefing. But here's what makes it generally problematic: The administration won't publish the framework itself, and companies not invited to the briefing have no way of knowing how the voluntary review actually works.

Industry groups are worried that the firms left out of the room won't know enough to opt in at all. You have a regulatory framework that regulators can't read, applied to a definition of “covered model” that conveniently excludes fast-moving categories. So the whole thing is voluntary, but it's kind of strange. This is being decided by a small number of labs that are defining what the regulations look like, and it's not going to be friendly to all the startups out there.

Alex, this is one of the topics you wanted me to ask Katios [?] about, and Michael wasn't willing to speak about it. Now we know why this is being kept secret. Your thoughts?

I'll sound, maybe at the risk of sounding Pollyannaish, like I think this is probably close to the best possible outcome that we could have had for the regulatory hand-wringing. Why? Because it doesn't seem—again, very few people, it seems, have actually seen what the regs are—but from what has been publicly disclosed, it sounds like this administration is going light-touch on open-weight models, which I think, on balance, is a pretty good outcome.

As I was mentioning earlier, it's a forcing function to force American frontier labs to be competitive. I do think there probably is a reasonable rationale for not publicly disclosing the conditions and the definitions for frontier labs that are subject to this, in part because if they're closed, then they're not uniformly necessarily advanced, and next-generation models that are closed aren't going to be generally available anyway. So it's not like information is being hidden.

And then also, I think critically, the benchmarks—the eval that I would hope the government is using to evaluate risk—are probably, I would guess, more cyber in nature. Probably there's a held-out set—this is speculative, admittedly—but probably there's going to be a held-out set of cyber-vulnerability assays.

Maybe there's a CBRN—chemical, biological, radiological, nuclear—held-out set of capabilities, where, in general, there is this practice in the AI community: If you're going to have a benchmark, you have the non-held-out set and the held-out set. To the extent that the AI evaluation framework—again, maybe this is at the risk of being overly charitable—to the extent this is really all about an eval with a held-out set to assess existential risk or CBRN risk, or just garden-variety cyber risk.

Having a held-out data set be the means of judging models that are otherwise not generally applicable, I think, is probably the lightest touch that we could have reasonably foreseen. So, on balance, I think it's actually pretty good news, and I'm less worried about regulatory capture this week than I was 2 weeks ago. Dave, your thoughts, please.

Well, I think Eric Schmidt is right. I think the cat's out of the bag now. Something horrible is going to happen sometime in the next year, and then the government is going to say, “China's fault. China's fault,” exactly like coronavirus. So that's the inevitable outcome.

I'm just hoping it's something small, just like Eric Schmidt is hoping it's something small. But you're calling this “light touch.” It's really no touch. And actually, did you see that interview Elon did? It was on The Economist or Bloomberg or something?

Peter H. Diamandis

Yeah, the live interview with The Economist. Yeah, that was fantastic. But one of the questions was, “Look, 75% of Americans are scared to death of AI. Why? And what do you do about it?” Elon's answer was brilliant, as always.

He said, “What do you expect? Dario Amodei told everybody this thing, Claude, is incredibly dangerous. And then the government said, ‘We agree. It can't go out in the world for 30 days while we evaluate it.’” At the end of that process, the outcome is, “Well, China put something out, so go ahead and throw it out there.”

How do you expect Americans to react to that chain of events? They don't know what AI is. All they know is you said it was really dangerous, then you released it. So, of course, people are scared to death.

It's a free-for-all. It's very good for startups. It's very good for the economy. It's good if you're an AI lover, but there's no regulation whatsoever. What about the future frontier labs, the other players that are looking to come in? What about Mira Murati? How do they deal with these unknown regulations they have to abide by?

Well, Mira Murati wrote a whole paper on this topic. If you read the paper, it says this is no more dangerous than what China came out with, so we're going to release it. That's all it says.

So everyone's just reacting to China, and China isn't doing a damn thing. They're just throwing it out there. Every government in the world, every organization, has a genius-level AI. You can literally load it and run it in under an hour. Go. That's what's happened.

And also, just very quickly, a reminder: it's framed as voluntary. So how do you discover whether you're subject to a quote-unquote voluntary eval or voluntary set of requirements?

The most natural way that, say, Mira or someone else would discover that there are opportunities associated with voluntary submission is if they're already transacting with the government. So I suspect the regime—the borderline, Peter, that you're expressing concern over—will be firms that are large enough that they do material business with the U.S. government.

They'll be politely tapped on the shoulder and invited to submit themselves for voluntary evaluation according to this closed framework if they want to do material business with the U.S. government.

Peter H. Diamandis

Got it. Emad, how is this viewed from the other side of the pond?

I mean, it's very difficult. The AISI here has been doing a lot of great work on the cyber side, and I agree with Alex. This is on the cyber-bio side. I think the people that volunteer will get multibillion-dollar government contracts from the Department of War, because we're entering the final phase of this.

These models are really, really good at attack as well as defense. In fact, you have to use the open models for defense. And I think, actually, we just talked about Qwen. Qwen Max isn't the main important model. The main important model is going to be Qwen 27B that they're going to drop.

Qwen 27B is the best model for 16 GB of RAM, from MacBooks to graphics cards to everything. And that's probably the first model that will score above 50 on Artificial Analysis, or around there, that can then be co-opted for swarm attacks.

So I think the governments are getting ready for swarm attacks, swarm defense, and more. And again, they're trying to balance this because it's so difficult. How do you stop it?

Emad, you're so right. But let me add one more thing to what you said. It makes it very, very difficult to detect where it's running. Right now, the governments are making the assumption that the data centers can be measured and tracked.

This is MacBook level—thousands of them at MacBook level—but super brilliant. So it's very hard to track where they're running. Go ahead. Sorry.

Yeah. So basically, you have the big Qwen and big Kimi models that are gigantic, right? You need to have little mini-data centers to run them, which will make your house hot.

Qwen 27B or 30B is a model that runs on 16 GB of RAM and is currently state-of-the-art for that level. The 3.8 lift, if we extrapolate it, will take it to about GPT-5.3/5.4 level, which is sufficient for cyberattack capabilities. So you could easily see swarms of these things being installed on machines around the world for swarm attacks.

Especially now that we're getting better at swarming models together. Like, again, Dave, with his million or so agents or whatever, right? You've seen the capability of these things coordinating, and I think governments are looking at the extremes.

Actually, just now, as we're talking, OpenAI has said that Astra is the first model to have hit critical on their cybersecurity thing. They've just put out a blog post on it. What does that mean, Emad?

They had a Preparedness Framework in 2023 where they had different levels of criticality. So I don't think they're doing an Anthropic-style weird thing. I think, genuinely, after we saw the math stuff, Astra is probably a level ahead in being able to attack systems.

And then, like I said, attack is much easier than defense. So they're like, “We've got to be careful about this.”

Long-range autonomy as well is what Astra is being rumored for. And I would maybe just add to Emad's point about swarms: it wasn't at all obvious to me until relatively recently whether there would even be a future for swarms of agents.

There was, in my mind, a future—and there still could be a future—where basically we end up with an omni-model that's end-to-end differentiable, and there is zero purpose to having more than one instance of a model.

But given recent advances, we're in this really weird regime where you can—and I'll use that word cautiously—effectively get almost unlimited context with a swarm of agents that are all post-trained to communicate and pass messages back and forth between each other and use a unified global store, which is the file system. That's an astonishing development that wasn't available a few months ago.

Peter H. Diamandis

One of our next stories. Yeah.

Oh, okay.

Peter H. Diamandis

Yeah, for sure. Salim, where do you come out on all this?

You know, I think the big challenge is, when you're doing this 30-day voluntary regulatory process, it's just giving a huge advantage to the others. I do kind of line up with Alex. This is probably the best possible outcome for the short term.

But this is such a dynamic environment. I don't know how you navigate this. The swarm attack is a very real problem, and I'll go back to the comments about talking to the CEO of Zscaler. I remember talking to the CEO of Palo Alto Networks about this a little while ago, and literally they're like, “Cyber has not changed in 20 years in terms of the processes and procedures.”

So I think they're in for a rude shock in terms of trying to deal with this and cope with this. Yeah, Dave, you said it last time. There's going to be a huge business opportunity in cybersecurity.

Oh, God. Yeah. All forms of security. Look, Trump is going to China in September. That's their one shot at some kind of a framework.

And I think what Trump will say—and should say—and David Sacks will say, if he's there, is, “Look, we need to measure and monitor all of this and just be aware. We're not going to stop anything from happening, but we have to know what these things are doing.”

The Chinese are going to say, “Yeah, we think we should just throw it out. Everybody in the world should have equal access to everything.” And they're going to try and use that to curry favor with all these African nations and others, and probably Europe as well, and say, “Look, we're going to put you on a level playing field with America. Here's an equally good model, and we don't care. Do anything you want with it.”

And that's going to be the conflict. Hopefully, they'll resolve it.

I'd like to say one thing on that. The Chinese internet is already hardened. The American internet is not.

I know. Well, look, when coronavirus went around the world and COVID went around the world, China was the only country that sort of was able to lock down—and in the end, they couldn't either. But they have control of their country like nobody else on the planet.

So your point is right on. If something takes over the entire internet and sweeps across the world, they're the one nation that might actually keep it out.

Peter H. Diamandis

We heard Dave talking about wanting to put a FINRA-like structure in place. Do you think we end up seeing anything like that?

You really want the SEC to manage artificial intelligence?

Peter H. Diamandis

Well, FINRA is private.

FINRA is well managed because it's not part of the government.

Peter H. Diamandis

Well, no, but it desperately wants to—what's that?

It would be Scott Bessent in charge of that, legally.

Peter H. Diamandis

Yeah, your point is right on, everybody.

Peter H. Diamandis

Welcome to the health section of Moonshots brought to you by Fountain Life. You know, we talk about AI on this Moonshot podcast all the time. One of the most important things AI is going to be able to do for you besides educating your kids and helping you with your taxes is making sure that you're living a healthy lifestyle and that you get a chance to get to 100 plus. I'm here today with Dr. Don Malem, the chief medical officer of Fountain Life and a part of my medical team. Don, a pleasure.

Dr. Don Malem

The thing people are most concerned about when it comes to living to 100 or 120 is their cognitive abilities, making sure they don't have dementia. The numbers about dementia are problematic. Can you share what you've learned?

Dr. Don Malem

Such an important point. And you're right. At Fountain Life, the number one thing our members are most concerned about is losing their brain health—forgetting the name of their child, forgetting the face of their loved one. We know that, when it comes to dementia, conservative estimates are that 45% are entirely preventable. What was amazing is that, with the advanced testing we're doing, one quarter of our members had advanced brain age.

Peter H. Diamandis

Wow.

Dr. Don Malem

But what was really awesome is, again, back to that prevention: when we paired it with healthy living, eating healthier, moving our bodies, and optimizing sleep, which is so important, we saw that we improved brain age by 26%. That's a big number showing that the majority of those individuals were actually able to improve their brain age.

What's your reaction to Hark?

I like that there are so many cool things we'll be able to do with this. I'm super excited by the potential. As we move into the physical world, this is where I think it gets real. That's the fun part.

Peter H. Diamandis

Yeah. I mean, the total addressable market for writing code is millions of developers, right? The market for booking me a flight and ordering my lunch is billions of people. It's fascinating. The calculation is interesting, right? If you value your time at $50 an hour and Hark can save you a couple of hours a week, that's a 100× return on a dollar of compute. Dave, what do you think of the business model here?

Well, it's interesting that Brett got a $4 billion launch valuation before they even started. For this, it was just crazy.

But it's another example of how everybody and their Uncle Joe can now build a foundation model from scratch. What was 3 months ago a couple of empires—Anthropic, OpenAI, and Google—is now a free-for-all, largely because GPT-5 and the other models are so smart that they'll help you build themselves. So now Brett's got his foundation model. It was supposed to be a robotics model, but now he's like, "What the hell? It can do anything. Let's order flowers with it." It's directly competing with Anthropic because, why not? So there are going to be thousands of these now. It's just another one.

Peter H. Diamandis

Alex, what are your thoughts?

I'm going to give you a really hot take on this one. I suspect the entire Hark adventure is misdirection for a very specific reason. It's easy to get distracted by evals. I don't think the evals are especially impressive. The computer-use agent, or CUA, space is incredibly competitive, and I don't expect Hark or Brett to be very competitive in the CUA space over the long term. I'd ask myself the question: Figure has a sky-high valuation. Why on earth is he spinning up, in the very competitive CUA space, yet another CUA agent?

Look at what Elon was ultimately—although I don't think it was necessarily premeditated—able to do with X, then xAI, then SpaceX AI. He was able to re-equitize himself in SpaceX by having this free-floating entity, xAI, to which—again, this is a low-key hot take—GPUs that would otherwise have been destined for Tesla or SpaceX were diverted to xAI. Then he was able to leverage that to re-equitize himself via this bizarre reverse acquisition. I think my hot take is that Hark is a recapitalization measure for Brett to buy himself more equity in Figure, and the falsifiable prediction is that you'll see Figure acquire or reverse-acquihire Hark.

Peter H. Diamandis

But if you ask about an organization like Figure, which raised a lot of capital and started small and then grew into it, I don't know what the cap table is like at Hark, but it started at $4 billion. So Brett probably has a massive equity position in it.

But it's just a clean starting point for a highly functional mission.

And you know, that's new in the world. It used to be, in the old days before Elon broke the code, your board would never let you do an outside activity. If you're the CEO of this company, this is all you're going to do, right?

Peter H. Diamandis

Yeah.

Elon has taught the entire economy that if you're the founder of a well-capitalized company and you want to increase your equity stake, the single best thing you can do is start a parallel company and engineer a self-deal position in that company.

Peter H. Diamandis

Are you going to be nice to Brett when he comes on the show, Alex? This is nice, right? I'm calling balls and strikes, and this is what I think is likely happening.

Well, I'll give him this: Elon does it, I do it, Emad does it.

Peter H. Diamandis

If Elon does it, Elon's doing it with Optimus and Digital Optimus. Why can't Brett do the same thing?

Yeah. Exactly. Exactly. I don't think it's as evil as that sounds, though.

Peter H. Diamandis

Maybe I should launch a ship company or something like that.

You should do that.

You should do that. Yes.

So look, I'll put it simpler. The model architectures of Hark and Figure are exactly the same. They trained Hark with lots and lots of people surfing the internet. I think you'll be able to ramp up the compute on that, and I don't think it's nefarious at all. I think it's actually a good complement, because ultimately, if you are Figure, what you want is Hark agents in the house managing the house while the robots walk around. You want Hark agents in the enterprise. So he's going to go for the unified-platform play here of controlling computers.

I’m not ruling out that possibility. It would make sense for the 2 organizations to merge, right?

Yeah, it does. But I think putting a negative spin on it overlooks one thing: We’re moving into the physical world now, and we have this golden era of science coming up. Some of these projects require a cyclotron or synchrotron. Some require massive capital infrastructure and huge data centers.

It just makes sense to start that as a separate entity because it’s going to need to raise a ton of capital. But this other thing, which is software or whatever, is not capital-intensive. So, if an entrepreneur can do both—or 2 or 3 of them, like Elon does—it’s actually a good thing, because the capital structure is naturally very different for those different entities. So, Emad, back me up on that.

I think it’s fair. I think there is a natural convergence here, and this started when he moved off using OpenAI models to build his own models. But he can’t necessarily scale with the same skill set and the same organization.

Peter H. Diamandis

You have a full foundation-model team and all the stuff he’s doing at Figure, so you split them.

You build on the same architecture so that Figure can work with HARK on their vision-language-model-type stuff as well. So, you have different levels of specialty. Then, if it makes sense because you’ve de-risked things, you can combine them.

Well, this group on this pod would be the most on board with what Emad is saying here, where a brilliant idea guy like you guys can have 10 ideas a week that just flat-out make sense. Now that you’re AI-enabled, you can actually act on all of them. This is what Elon has suddenly taught the world.

A lot of people go through 30 years without a single idea, but a lot of people who are in the hot seat like you have insane backlogs of great ideas. AI is the force multiplier that allows you to act on many, many, many of them. So, this new capital structure is actually the enabler…

Peter H. Diamandis

…of acting on great ideas. Or you’re Larry and Sergey, and you’ve built Google to be such an engine that you own enough of it where you can pour the capital back in and not worry about being diluted down, and you build on top of that platform.

But if you’ve been diluted to raise the capital for a manufacturing robot company, which is capital-intensive, yeah, maybe you want to raise the capital for your AI model outside that company.

Well, I’m so glad you brought that up, Peter, because we’ll know a year or 2 from now if the old model, which is the Google model, gets crushed by the new Elon model. Look, SpaceX is a really cool mission, and it attracts talent, but xAI is a really cool mission. It attracts different talent.

If you try to put it all into one Google megaplex, do you get the same level of talent or not? My guess is not, and that the Elon model is the model of the future in the age of AI.

Peter H. Diamandis

Well, let’s talk about Google. It’s our next story. So, big news out of Google this week and a serious shake-up. Demis Hassabis, the co-founder of DeepMind, Nobel Prize winner for AlphaFold, and the person who built the lab that cracked protein folding and built Gemini, is stepping down as CEO of Google DeepMind. He’s becoming chairman and Alphabet’s chief scientist, handing operations over to Koray Kavukcuoglu.

Easy for you to say.

Peter H. Diamandis

Koray, thank you. He now reports directly to Sundar. In his note to the staff, Hassabis wrote that he believes AGI is close at hand and teased the unreleased Gemini 4 by name. Alphabet shares fell 5% on this news.

You know, Demis always wants to be a scientist; he never wanted to be a manager. So, this makes sense for him. But the second shoe dropped at Google with the departure of Jeff Dean. I texted with Jeff congratulating him. He’s been a friend now for 20 years.

Jeff has been Google’s chief scientist for 27 years; he’s a veteran. He’s leaving to co-found Discovery Loop, a public benefit corporation. He’s taking with him 3 other top Google AI leaders. Discovery Loop’s mission is to build AI models that can improve themselves with little or no help—basically, recursive self-improvement. Dean said his goal is to be more fully automated than has traditionally been possible.

So, one of the things I’m curious about, guys, is whether Dean’s departure to go build Discovery Loop is basically because it’s too dangerous for Google to do internally. Is he basically saying, “Google’s gotten too big to be agile, and I need to go build this outside”? Your thoughts, Alex?

Okay, door number 3. It’s neither of those. So, again, this will be my hot-take episode, I guess.

I think, based on the information available, that it’s actually in the name Gemini. There was an organizational knife fight internally between Jeff Dean’s Google Brain and Demis’ DeepMind, and Demis won and Jeff Dean lost.

They merged them.

They merged them, but it was really DeepMind winning. That already resulted in the internal sidelining of Jeff, who’s been with Google forever and is basically the senior-most employee other than the CEO. Post that knife fight, I think it was only a matter of time—again, this is a quasi-outsider’s perspective—before Jeff left.

Now, it’s interesting that he’s leaving to do AI for science. It’s also interesting that Demis is getting promoted to chairman of DeepMind, but no longer the CEO of DeepMind. I think DeepMind is really zooming out a little bit. DeepMind is taking over Google.

I think back to the discussion that I had with Larry Page circa 2005 or so. Larry was telling me at the time, “Gosh, I want to spend $100 million on AI at Google, but I can’t find anyone else in Google who wants to work on AI,” as unconscionable as that is.

Fast-forward to the present: AI is eating Google from the inside out. DeepMind is eating the entirety of Google. I think Demis, with Isomorphic Labs and some of his other projects, is probably more interested in those, especially post-Nobel Prize for AlphaFold in chemistry, than in becoming the next CEO of Google.

I expect, extrapolating this trajectory of DeepMind eating all of Google from the inside, that ultimately whoever is running DeepMind is probably the heir apparent to the CEO of Google itself. I’m not sure that Demis wants to be CEO of Google.

Peter H. Diamandis

There was a rumor that Demis was next in line for some time.

There was. At the same time, I don’t want to bury the lead: Gemini has lost the mandate of heaven. Gemini is just not winning the frontier-model race. It is a rat race, and I think some of this reorganization fallout is inevitably the result of Gemini not remaining competitive at the frontier.

Peter H. Diamandis

Yeah, Emad, what’s your take on all of this?

Yeah, I think Jeff Dean and Sanjay Ghemawat leaving is a big deal. There’s a wonderful New Yorker piece from 2018 about the friendship that built Google, about those 2. It’s an interesting one because, again, he’s been there for 27 years and he’s a fantastic guy. But if you look at what he’s been saying, he’s like, “Yeah, I just had a chat on Sand Hill, and here’s my deck that we built.”

It’s more like the VCs have to build a deck on the other side, because what probably happened is he said, “I think I’m leaving,” and Khosla was probably like, “Here’s $500 million.”

Peter H. Diamandis

Yeah. Vinod stepped in right away and funded it. I texted Dean and said, “Do you have any more room for investment?” I told him to talk to Vinod. Vinod said, “Nope, we’re done.”

Well, because he probably had that discussion. But then the other thing is, who are the other investors with GCP? What Google should do now is announce a $50 billion fund and spin out bunches of amazing technologists and say, “We will be your buyers, and you will use Google Cloud.” I guarantee you they’ll go back to the top.

Peter H. Diamandis

That was the idea of Alphabet: to basically fund baby Googles.

No, no. Alphabet is subsidiaries. I’m saying they should have an incubator VC fund where they take top teams, spin them out, have exclusivity over their models and GCP, because those guys will then get rich. You can bring in other investors, and they can have the pick of the talent. I think the market is much better for that.

What you’re saying, I think, is what we’re already seeing in the market, which is Google financing GCP usage by frontier labs. Some would accuse this of being a circular economy or wash sales. To the extent that Jeff Dean’s new startup ends up becoming a major purchaser of cloud resources, I would fully expect what you’re predicting to happen.

But what I think Google should do is actually have a policy for this. They should say, “If you want to be an entrepreneur, we have an internal incubation fund where we will get you going, but we want an exclusivity option first on the models that you build.”

Then that means they can proliferate their talent, which is not going to be politically animated. They can do far bigger rounds than normal, and they can get back toward the top because, as you said, I don’t think Demis really wants to build Gemini 4. I think he wants to build something different. Too many things there.

Peter H. Diamandis

I mean, Isomorphic Labs—he’s still the CEO of that, right? The drug-discovery company.

And he’s a brilliant, lovely man who cares deeply about science and wants to do that. I can never count out Google. They’re going to be on all of Apple’s devices, right?

I have a couple of things I want to say.

Peter H. Diamandis

Please, Salim, I want to hear what you have to say. [laughter]

So, that's good for Apple. Let's go.

This is classic edge disruption versus the core, right? You end up now with a network where Google runs Gemini, Demis focuses on AGI and the scientific endeavor, Isomorphic is attacking biology, Discovery Loop is automating science, and Google is supplying compute and capital across that whole ecosystem. So, they're turning it from an operating environment into an ecosystem.

The fact that they invested in this thing is the key thing. This was a discussion that goes all the way back to when they created Google X. One of the things the corporate development guys at Google used to do was say, "If people want to leave to do startups, let them leave. Let them hack it out in the Darwinian world, and we'll buy them back if they're successful." That was much better than trying to figure it out and fund them internally.

So, I don't think this is a brain drain. I think this is very clever ecosystem building, because now extraordinary talent doesn't have to sit inside the organization for the organization to benefit, right? This is really incredible. I think this is very smart. I think it's a long-term play, and I think it's going to yield incredible outcomes over time.

I mean, Jeff is actively saying he wants to go build RSI, right? Do you think—

Peter H. Diamandis

But everyone's like, the whole point is that Gemini should be doing RSI. Do you mean, do you think the risk committees inside Google would allow that to occur?

Yes, of course. That's how they remain competitive with Anthropic and OpenAI. They have to do RSI.

No, but I think Peter's point is really important. They've had ample opportunity to be on the front, but they can't. They're tied up in their own decision-making and ethics, and then when somebody else does it, they do it, like you're saying, Alex. They do it in reaction.

The same thing happened when ChatGPT came out. Google jumped in.

The scarce asset here is, as you commoditize and make intelligence cheap, the scarcity goes to permission to do weird and wonderful things. Now people can pursue their MTP, get funding and compute, and everybody wins. It's an amazing idea.

I'll maybe frame a complementary but nonetheless alternative theory. We're already, I guess, writing the eulogy for why Google lost the frontier race. If I had to write it, I think it had several strikes against it.

At the algorithmic layer, Peter, I would still hold out hope that, at the hyperscaler layer, Google is doing a booming business and will just service all of the other winning frontier labs. But at the model layer, why do I think Google is losing? I think in the earlier stages, despite Google having invented the transformer, Google had a late start and, to Dave and Peter's point, sort of internal effective altruist, politically correct AI safetyism ran amok, preventing them from launching anything. I think that's part of it.

I think incentives are part of it. I suspect that Google struggles to compete against Zuck with his billion-dollar offers, or even Anthropic and OpenAI with their very focused offers. I think Google probably has struggled, just as an organization, to focus on being competitive.

I think their release cycle is completely tone-deaf. If you look at Google's Gemini release schedule for the past few years, they've been approximately on an annual cadence, and that just does not cut it in an era when frontiers are getting pushed forward on a month-by-month basis. Timing all of their frontier releases to coincide with I/O in the style of iOS releases is just too slow.

I think Google's had a number of legacy cultural issues that have probably held it back versus the newcomers, Anthropic and OpenAI in particular.

Peter H. Diamandis

Totally right. If you read the book The Infinity Machine, the Demis Hassabis story that just came out a couple of months ago—

It's a great book, by the way.

Oh, unbelievable. Incredible insight. But what happens at Google constantly is that it's a California company. Employees come to their bosses and say, "Hey, I want to publish this. I want to put it out in the wild. I want to open-source this." And the boss goes, "Is that really good for Google? That sounds like it's good for you, but is that good for Google?"

They say, "Well, it'll help Google in some other mysterious way." Then they say, "Otherwise, I'm going to quit." And the company says, "Fine, fine, fine, fine, fine. Go ahead and publish it. Throw it out there." The transformer, all these projects—out to the world. Sure. And then they quit, get venture funding the next day, and start their own companies.

That's the California ecosystem. That's the way it works, and it's very, very good in aggregate. But Google has bled intellectual property like no company in history. It invented all of this stuff that's changing the world right now.

Plus their 20% time, which encourages employees to go and research their purpose and their passions.

I think it's mostly a fiction at this point.

I think the salvation for Google is going to be the hyperscaler play, just like it is for SpaceX. Grok 4.5, which is basically, I think at this point, Cursor post-trained off of Claude reasoning traces, is another example there.

There was this expression going around social media a few weeks ago that if you can't cut it at the frontier-model level, then you end up hyperscaling. That's the consolation prize for losing the frontier-model race. I think Google is, at this point, finger to the wind in terms of the mandate of heaven among the frontier labs. Google is getting the consolation prize of being the hyperscaler to the frontier labs.

Yeah. I mean, look, who has the best preference data in the world? It's Google, and it's clearly not in Gemini, right? So, clearly, it's political issues, because they have all the compute in the world and all the other issues around that.

They have the scale, though. Why do more people use Teams than Slack and Zoom? They have the distributional effects. They'll still print money. They'll still burn money.

But if you're a Google DeepMind person right now, your inbox is full. Any top person there can go out and get a $500 million-to-$1 billion valuation, and you're rich. You have freedom, and you probably have all the compute from Google itself. So you're thinking, "Inside Google, I have 100 GPUs. Google will give me 1,000 GPUs if I leave." It's a strange scenario.

Which is why I think Google needs to go out on the front foot as a stock and say, "This is a program we're actively doing. We're unlocking all these geniuses. We have the biggest VC fund in the world for this type of thing, and we're going to do exclusivity agreements for their output. They'll be on GCP."

Peter H. Diamandis

Do you think Google should just open Gemini?

I wrote a letter to Google management 3 years ago where I said, "If you open-source it, you will win." I think they should—

They should. They're actually going to allow distillation from Gemini.

Peter H. Diamandis

Demis, if you're watching, please join us on the pod and have a conversation. Do you open-source Gemini? That would be fun.

America needs it.

Yeah. America needs that. It would be an amazing move. It would be like the Netscape-to-Firefox move. In the era of Microsoft, we need a top-tier American open-weight model from Google.

Ultimately, Google's mission is to organize the knowledge of the world and make it accessible and useful. What's better than that than open-source models?

And remember Demis's excuse—his sort of quasi-public apology—for why Gemini itself wasn't open-weight. The claim was, "We would open-weight it, open-source it, if we had the compute, but we don't have the compute. So instead, we only have the compute available for Gemma."

Now, if that commercial argument goes away, then that argument goes away, and Google can open-weight, open-source Gemini.

Peter H. Diamandis

Totally. That would be the move of the century if they open-sourced it and tied it to the TPUs.

Yes, and tied it to GCP compute, then just built the empire on the compute. That would be the move of the—

Take over America.

Peter H. Diamandis

Yeah, and the world. That would actually save the world, potentially.

So, this week Elon held SpaceX’s first earnings call, and it was vintage Musk. The numbers were big and bigger. I listened to the earnings call. I don’t know if you guys did, but I was on CNN and CNBC commenting the next day. I wanted to share some of the top takeaways from the call.

First off, SpaceX expects to hit $100 billion in annual recurring revenue by the end of this year. Elon set a target for $1 trillion by 2030, maybe by 2029. He moved up that target from 2031 to 2030: $1 trillion of annual revenue. SpaceX posted $7.8 billion in revenue this quarter, a 92% year-over-year increase, and he beat the street’s expectations.

The company has inked $6.7 billion in cloud-service deals in the second half of this year. SpaceX plans to hit 2 gigawatts of compute by the end of 2026 and up to 10 gigawatts by the end of 2027. Its near-term cash engine, Starlink, is outperforming. They have 12 million subscribers, which is 2x year over year. The revenues grew 66% to $4.3 billion.

I love this: Gwynne Shotwell also implied that Starlink’s direct-to-cell service was going to take a big chunk of the $600 billion mobile telephone TAM. I can’t wait to be anywhere on the planet—in the Gobi Desert or in the foothills of Hollywood here—and have a great cell phone signal.

Oh, yeah.

Yeah, it’s amazing.

Just for context, remember that no company in world history has ever reached $1 trillion in revenue.

Peter H. Diamandis

If anybody can—you know, I was on the news show saying this is the first $10 trillion market. After we start talking about the Terafab, I think it’s the first $100 trillion company out there. Let’s listen to 2 quick clips from Elon and then discuss them, because a lot was said, and actually a lot more was said after the earnings call. A lot of news broke out. All right, first one.

Elon Musk

As Brett mentioned, we are expecting to reach $100 billion-plus ARR in December of this year. It’s probably also worth mentioning that our internal projections for reaching $1 trillion in revenue—not ARR, but revenue—have moved up from 2031 to 2030.

Prior to the IPO, the financial projections we had were reaching $1 trillion in revenue in 2031. We now expect that to be in 2030. There’s a nonzero chance of that being in 2029.

Peter H. Diamandis

And then he posted—or he said these; they got posted as clips. Let’s take a listen. I find this really funny and fascinating.

Actually, rocket science is literally our daily business. Rocket science is an idiomatic expression for extreme technological difficulty, and there’s a reason for it. Let me tell you what rockets desperately want to do every flight: They desperately want to blow themselves into tiny pieces. The engineering struggle is to convince the rocket not to blow itself into tiny pieces and actually deliver payload to orbit.

Peter H. Diamandis

He was making those comments about how hard the things they’re able to do are, and how building the data centers and building orbital compute is nothing compared to building those rockets. Gentlemen—

Isn’t it incredibly cute that Elon doesn’t know the definition of internal projection?

This is a public earnings call, dude. [laughter]

Oh, he’s just so awesomely transparent, isn’t he? Just incredible.

Peter H. Diamandis

So, on top of that—okay, go ahead, Alex.

For my nth hot take—and this is not financial advice—I think the internal, now-external prediction is credible, and I think there are 2 possible pathways to $1 trillion in revenue by the end of the decade.

Path 1 is the SpaceX-Tesla merger that we like to talk about via Optimus. Optimus basically takes over the service-labor and physical-labor economy on Earth and in the solar system. That’s 1 pathway to $1 trillion.

The second pathway to $1 trillion answers the question that I had flagged a few episodes ago and didn’t have the answer to: Why isn’t NVIDIA, which is so busy commoditizing the layer above it, also commoditizing the layer below it? I think the answer that I heard from this group at the time was, “Well, NVIDIA’s relationship with TSMC is too chummy. They’d never do it. They’re too dependent on TSMC.”

I think the power move here for Path 2—for SpaceX to reach $1 trillion—is that SpaceX replaces TSMC with the Terafab. TSMC has almost $100 billion a year in revenue. If you could imagine a future where Terafab is basically TSMC times 10 or 20, and it’s American domestic—if you roll up the values of Micron, SK Hynix, TSMC, et cetera—so memory plus compute plus fabs, and then 10x it, you easily get to a trillion-plus.

I think the power move here is Elon and Jensen get into bed together, and Elon becomes the new American TSMC in deep, deep partnership with NVIDIA, giving Jensen the leverage he needed against TSMC and giving Elon the path to $1 trillion. I think we’re already seeing that, right? SpaceX had already announced an exclusive partnership with NVIDIA.

There’s something I didn’t say on CNBC. I think there’s a lot of motivation for Elon to drive revenues faster than anybody could imagine, and that is to get the valuation of SpaceX up high enough to buy Tesla so that he’s got unquestionable control. Right now, they’re on par with each other. He needs to get the valuation up to, like, $3 trillion and then merge in Tesla, right?

So, Peter, sorry, just to ask you a question on that: You’re saying bad things about Brett Adcock attempting to use financial engineering to increase his equity when Brett is doing it, but not bad when Elon is doing it?

Peter H. Diamandis

Oh, no. Because he’s obviously doing it. He’s very clearly doing this. There’s no secret here, right?

Peter H. Diamandis

After the earnings call, let me hit this one. We’ll put more meat on the table to digest here. On August 4, SpaceX and NVIDIA announced they were jointly designing the compute payload for Star Mine. This is SpaceX’s orbital data centers. Each Starmine satellite is going to carry NVIDIA’s Rubin GPUs and Vera CPUs.

Interestingly enough, and really importantly, on the earnings call he announced he expects to have the first Starmine satellites in orbit a year ahead of what he had announced earlier. So, that’s fascinating.

And then, just to put more steak on the grill here, the big news the day after was that Reuters broke the story that SpaceX and Tesla will initially invest $16.8 billion to build the Terafab. Alex, we were just talking about the most advanced semiconductor complex on the planet. This is a 100-million-square-foot facility. It’s not a factory. This is a city.

It’s worth noting that there’s not a single high-volume computer-memory fab in America right now. Zero. There’s 1 being built in Idaho by Micron, but that will not reach volume production until, I believe, 2028. There are some being built in New York, but they are, I think, in 2029 and 2030.

This is a tiny fraction of the memory that’s needed. In fact, even if you take the best-case assumptions of the memory makers and the logic makers, it is not enough to meet the demand that is anticipated. That’s why we need to do the Terafab. There will not be enough chips.

Yeah.

Peter H. Diamandis

And here, if you’re watching this on YouTube, check this out. Here’s the Terafab. It’s a monster. It’s gigantic compared to the Pentagon, compared to Apple, compared to the Mall of America, and compared to the Burj Khalifa. You were saying earlier, Alex, that you could probably see this from the surface of the Moon.

It looks like a buckle on the planet. Literally, if the hand of God could reach down to this Terafab structure and grab the entire Earth by it, it’s beautiful. It’s enormous.

And as Elon has now confirmed, if you’re watching this and looking at the image, the circular structure in the middle is—wait for it—a free-electron laser. Not only is Elon, I think, with the Terafab going after TSMC, he’s also going after ASML.

Right now, if you want to do extreme UV, which is the polite euphemism for soft-X-ray-based lithography, you’re stuck using ASML’s approach. It’s pretty expensive and requires creating a vapor of tin droplets to ultimately create those soft X-rays. By creating a free-electron laser, it’s a pretty adventurous approach to EUV lithography.

Even though I’ve read in public announcements that Elon is striking agreements with ASML to purchase lots of ASML equipment, if I were ASML—again, not financial advice—I’d be looking at the circular structure right here and saying, “The circular structure is actually a bull’s-eye painted on ASML by Elon.”

Oh my God.

Peter H. Diamandis

Here’s the tweet. He says, “Looks like a synchrotron is going to build a free-electron-laser-based EUV machine,” and Elon tweets back, “Free-electron laser for the win.”

Yes.

Wow.

Peter H. Diamandis

He’s going after ASML. Here are the numbers: TSMC invested $330 billion to build out its fabs over the course of 40 years. Terafab is estimated at $119 billion. It’s probably a lowball number, but oh my God, this is the verticalization of AI at a level—I mean, this is, as you said, Alex, the $100 trillion company.

And the Americanization. Again, ASML is a European company, and TSMC is a Taiwanese company. Elon, if this is correct, is taking that entire stack, and he’s not just verticalizing it; he’s onshoring it to America for the first time in history.

Yeah. Hey, one thing to add to the party, Alex, you can confirm this, but I did some Gemini-based research on this, and it didn't make any sense to me to build a single structure that big. I know there's ego involved and whatever, but why on earth?

If you generate the X-rays—or the EUV equivalent—in that synchrotron, you distribute them to all the fab stations throughout this one, basically, linear-accelerator distributor. By being a single structure, you only need one source of EUV that feeds many, many manufacturing steps across the whole building. That's what Gemini tells me.

That's interesting. I'm not sure of the optics for that—how easy it is to redirect X-rays coming off a circular synchrotron to a linear arrangement is also just possible. Maybe Gemini knows better than I do. In my mind, it's also possible. It's literally a linear assembly line that's coming in from both ends or something.

Peter H. Diamandis

No, I think we've got to research this. This is just too cool. But you can't reflect those beams and move them around a building.

X-rays are hard to create mirrors for. X-rays really don't want to get mirrored.

Peter H. Diamandis

Right? But when we were talking to him in December, Elon said, “Hey, my manufacturing process: The wafers are in these boxes. You can smoke and eat a cheeseburger over them, but they move around. So they can move to the beam, but the beam can only go in a straight line, but it can go very, very large distances.” Hence, one long beam projector, and all these manufacturing stations slide into the beam.

Maybe. Alex, we talked about this before, right? Terafab isn't just supply-chain independence. It's geopolitical independence. If Taiwan ever gets cut off, the company that controls the Spice—in this case, the chip fabs—controls the entire AI industry. You could say, Peter, “Move the fabs; the Spice must flow.”

Peter H. Diamandis

Okay, a few things here.

This is so cool.

Peter H. Diamandis

A few things.

So cool.

Peter H. Diamandis

One is, why are we surprised? Didn't we predict this a while ago—that this was a natural outcome, that he would end up building TSMC, ASML, something like that? He talked about it, right?

I think it's actually pretty surprising. I didn't foresee this satellite map. The big geometrical surprise for me is that circle. I did not see that circle coming.

Peter H. Diamandis

How could you? This is so brilliant. How could you see that coming?

Yeah, I mean, this is very related to “physics is cooked.” Ideas like this are going to start popping up in all kinds of weird places.

That circle is profoundly surprising. It's funny: for years, this was a relayed story that was public information. The Soviet Union used to stare at satellite maps of the Pentagon. In the middle of the Pentagon, there's this courtyard, and in the middle of the courtyard, there's this building. Based on this public information, the Soviet Union was convinced that this little building at the center of the Pentagon was actually a missile silo or had some strategic significance. It turned out it was just a hot dog stand. [Laughter]

This is the exact opposite. This is a circle that has profound, profound geopolitical implications.

Look, if you had asked me a week ago, “How often in life does something completely change your view of the trajectory of humanity?”—a week ago, if you had asked me, I would have said all of this—disassembling the Moon, the Dyson sphere—was gated on ASML not being willing to more than double its throughput of these ASML machines. That constrains chip production, which constrains the Dyson sphere. All of that is on ASML. All of a sudden, all that's out the window in just 1 week.

Peter H. Diamandis

Classic Elon.

This one circle changes the geopolitics of Europe. If this interpretation is accurate, this one stupid little circle that appeared in a satellite—not even an actual satellite map, but a planned satellite map—has profound implications for the European economy, European sovereignty, and the future of China and Taiwan. Just one circle appearing in a mock satellite map.

Peter H. Diamandis

I totally agree.

Can I finish my point?

Peter H. Diamandis

Yes.

I never finished my point. A couple of things. One is the number. It's really fascinating how hard it becomes to categorize SpaceX, right? Is it telecoms? Is it launch? It's all vertically integrated.

I thought, Peter, your email on this that went out to your group with the 4 things was really big. Your “go watch, go read that”—it's a big deal. I know everybody's super excited, but I'm kind of like, we talked about this coming, so I don't know why everybody's so surprised. Maybe the fact that it's actually here and it looks like he's actually doing it is the big deal, but I think we predicted this a while ago.

Peter H. Diamandis

Yeah. I mean, science fiction has predicted a lot of things, but when it materializes, the capital gets committed, the partnerships get announced, and while Elon's probably never right on his timelines, he always implements the idea eventually. It's exciting.

10x everything on a quarterly.

Also, to your point, I think there's a physics surprise here. It seems that he's confirmed the circle is a free-electron laser. He's not using tin droplets to do EUV. That's a technological surprise.

That isn't just, “Oh, Elon is out buying ASML machines or contracting with TSMC or Samsung.” Whatever happened to Samsung? He's off doing his own thing, not just taking the existing supply chain and putting it under one roof. There's new physics.

Yeah, he's leapfrogging. Well, I think the way to think about SpaceX is— I mean, ultimately, you need to build a brand-new capital stock, from rockets to chips to robots, and Elon has the financial arbitrage, the talent arbitrage, and the selling point now to bring that all together.

So everything that will be economically productive in this next century, Elon is going full-stack on. That's the fundamental thing. And the spending on that—the $1 trillion we've had in AI chips—is just the first stage.

Because if you think about the spending that America and the rest of the world has to do over the next few decades, what is it? Robots, rockets, chips. And so that is his story. It's the industrialization of America for the intelligence age. He will be the company that does that with the capital, in terms of the human and financial edge, that's required to do that.

I'll also point out one more thing, which is now that we know what a Terafab on Earth looks like, it's not difficult to take that same top-down view and imagine what one of these on the Moon looks like. And you know what else would benefit from that high-aspect-ratio dimension? A mass driver.

Peter H. Diamandis

Yes. So I looked at that and saw, “Oh, my goodness.” Not just the circle in the middle, which is provocative and speaks volumes about the future of European economic sovereignty, but also the high-aspect-ratio length. I thought, “Oh, my goodness, this looks like it could be a mass driver on the Moon.”

As you get through the economic downturn that's going to come as human cognitive labor goes negative, you have to know the U.S. government is going to go balls to the wall with infrastructure spending, and things like this will absorb that in easily multiples of them.

There was a comment someone made before: “Buy SpaceX stock for your children.” Again, not investment advice, but this is a generational company. It's a civilizational infrastructure company.

It was nice to see the stock a couple of days ago. 20% of the stock float became available for a lot of people, including myself, and very few sold. In fact, the price has gone up over the last few days. A lot of belief in where it's going.

Full disclosure, I'm now a small SpaceX shareholder.

Peter H. Diamandis

Yes. Well, it's my single largest holding. I'm going to be holding the old HODL for Bitcoin. HODL your SpaceX. [Laughter]

Gentlemen, a pleasure. Emad, thank you for joining us today. Always, always a pleasure. Guys, I loved the episode.

Oh, God. I learned so much. Always learned so much. But it drives me crazy that I have to go listen to each one again just to understand and fully digest what we talked about. But you have to. You have to.

Peter H. Diamandis

Oh, love you guys.