Bernie 要求各实验室停止，华尔街将 GPU 变成债券，Grok 4.7 登上 #1，特邀 Emad Mostaque

Bernie Sanders sent a formal letter to the CEOs of Anthropic, Meta, and OpenAI.

AI capabilities have reached a critical threshold. Pause AI development.

The cat’s out of the bag. It’s too late, right?

NVIDIA just announced a partnership that redefines what GPU compute means as a financial asset.

This is like the very first pitch of the first inning of the build-out of the Dyson swarm.

Financial assets want predictable depreciation, and exponential technologies don’t give you predictable depreciation.

I’m not as concerned that this will end up being another mortgage-backed securities fiasco for a few reasons. One is—

Elon just dropped Grok 4.6, and he’s right: it’s a banger. Grok 4.5 came out 2 weeks ago, 4.6 just this morning, and 4.7 is rumored to be coming out in 2 weeks.

4.7—and this is what I think is going to be really interesting—is going to be trained on all the SpaceX physics and engineering knowledge. You can’t get beyond frontier with this stuff, but what if you have the best engineers? Now that’s a moonshot, ladies and gentlemen.

1. Longevity Enters Its Breakthrough Era

Welcome back to Moonshots, everybody. Uh, we're gonna be covering nine stories today that span the frontier from longevity, AI infrastructure, synthetic biology, AI filmmaking, urban air mobility, and oh, by the way, Grok 4.6 is crushing it. Uh, the through line is the same as always. You know, accelerating singularity is compressing the distance between the impossible and the inevitable. We're back this week with the Moonshot quintet. Uh, AWG, our in-house super intelligence. Alex, welcome. Good to see you and your normal heart.

Thank you, Peter. Good to be super intelligent.

You are, my friend. Dave Blundin, our investor in AI extraordinaire. Dave—

I was over at MIT all day today, hence the garb. But, uh, our tech track teams just got back from SF, and they brought back huge amounts of knowledge, and they're all over at MIT Z Sale dispersing it across the campus right now.

Love it. And Salim Ismail, our empresario of exponential organizations. Salim, where are you today? Still in Toronto?

I'm still in Toronto, heading back tomorrow.

Hmm, okay. All right.

I should be going back now, but this podcast happened, and I can't fly— ... you know, within like six hours of it.

Yeah. Yeah.

So—

We're gonna, we're gonna put a Starlink on your, on your hat and have you walk around with it.

Okay.

Uh, and back by popular demand, Emad Mostaque, the embodiment of the intelligent internet. Emad, I hope you're reading the comments on the last pod, uh, that we did together. People are loving you. Have you seen them?

Uh, yeah, I did see some of them. I don't normally read the comments. I made an exception, and it's very kind what everyone said.

Yeah.

It's just great being—

It's just the British accent.

... amongst you guys.

That's all it is. I mean—

It is. It is. It worked for my American wife, it works for the others, you know?

Um, Peter Diamandis, your host, your optimism amplifier. Please remember that having an optimistic and abundance mindset is a choice, and our mission here is to deliver data-driven optimism that helps you make that choice. So guys, uh, last night, I was in Utah at the University of Utah. It was the finals for the $101 million longevity—or, shall I say, Healthspan XPRIZE. Super psyched. We awarded $1 million to 10 teams and recognized another 10 finalists. This is on the march to winning the $80 million grand prize. We’ve given away $20 million so far.

The mission of these teams—and everybody, please get excited about this—is to add 20 healthy years to your life, asking teams to reverse the functional losses that we get through aging. In cognition, giving you the ability to think and have memory like you did 20 years ago; in muscle, the ability to build muscle like you did 20 years ago; and in your immune system.

It’s extraordinary. We had 800 teams enter this competition, with every possible approach, from mitochondria to stem cells to gene editing. It’s extraordinary. This is a competition that will be won by 2030, so keep an eye on this. Salim, I don’t know if you want to jump in on this one.

You know, 2030 is infinity, right?

Yeah. Well—

I think the XPRIZE was such a huge inspiration when I was writing the Exponential Organizations book because you’re reaching outside in. I think that’s a record for the traditional prizes for us, right?

It—

It’s kind of an incredible thing.

Like Elon’s $100 million prize for carbon removal ended up with something like 1,500 teams. But reversing aging has got to be tougher.

By the way, the teams here don’t do this in theory. They don’t do this in mice. They have to actually do human trials. So they’re all going to be doing trials with control groups and humans, probably around 150 people in the trial. It’s real data, and we’re going to actually know which of these 20 to 30 approaches that actually make it to the finals work.

For me, Alex, we’ve been talking about being in the midst of longevity escape velocity. This is accelerating it.

It’s going to be spiky, I think. I still think—just as with AGI, either it’s in our rearview mirror, as I think, or some might say it’s either here or almost here—I do still expect that healthspan and longevity escape velocity are going to be spiky. I’m just optimistic that we can even out the spikes, even if there’s a subpopulation that achieves healthspan on the margin a few years before everyone else.

You know, one of the most important things I think about this competition is that when we launched the $10 million Ansari XPRIZE for spaceflight back in 1996, people did not believe in commercial spaceflight. They didn’t believe that individual teams could do this and carry humans compared to the government.

As it progressed, the confidence level in this grew, and then, when it was won, money flowed in, regulations changed, Bezos and Musk started Blue Origin and SpaceX, and it changed the game. So I’m feeling the same thing going on right now. Longevity—the idea of reversing aging—has been a crackpot idea for most of the last few decades while I’ve been in the industry, and it’s beginning to change. People are starting to believe, yes, it’s going to happen, yes, we’re in this healthspan revolution.

We’re getting to a when, not an if, question, right? And that’s really huge.

That’s really important here, Peter. When you launched the Ansari XPRIZE, there was no space industry to speak of, at least on the commercial side, and now we have a multitrillion-dollar industry that did not exist, right? Because of the collective innovation—

Multitrillion, yeah.

—plus all the members of all the teams end up going to SpaceX, Blue Origin, et cetera, et cetera.

We could expect the same thing to happen here, where essentially you’re creating a $101 million bet. You’re building a portfolio of experiments and allocating capital to wherever there are demonstrated results. It’s an unbelievable model that we’ve now seen repeat over and over again. It’s incredibly exciting to see.

Question for you, Peter, on this one. I always ask the counterfactual question. Having now run this Healthspan Prize, or at least the beginnings of it, do you think that if this prize had been created, say, 20 or 30 years ago, we could have made enormous progress on the margin? Or do you think there’s some historic contingency that means right now is really the first time in history where we could make—

Great question.

—credible progress on it?

We had a lot of comments. We had a lot of the top scientists there. Aubrey de Grey was there last night, who coined the term longevity escape velocity, and then Ray popularized it.

I think there was a consensus that the timing is perfect, that the tools for genome sequencing, for making specific molecules, and for being able to measure and report with AI are the tools that are required today. There might have been some approaches that could have been done 20 years ago, but I think today is when we’re going to make the greatest progress.

I’m starting to see capital flowing in aggressively. At the end of the day, longevity is going to be the biggest market, right? If you could add 30 years to your life, I’ve had this conversation, Salim, probably you have as well, onstage with YPO audiences or family offices. I say, “How much of your wealth would you spend for an extra 30 years of life?” The honest answer is nearly all of it, right?

Yeah, and the powerful distinction there is not so much lifespan, but the healthspan effects are really, really powerful.

Yeah. The numbers today are, for the United States, if you look at it, basically, 100 years ago in 1900, the average life expectancy was 47. Today, it’s 79. We added about 2 months per year over the last century.

Today, while the lifespan is 79, you’re healthy until an average age of 63, and you spend the last 16 years of your life in poor health.

When I was talking to Khosla about this, I said, “Listen, the biggest benefit the U.S. budget could have and the U.S. economy could have is adding 20 healthy years to people’s lives. They’re retiring later, and they’re not spending as much government money on sick care.” So it could be a huge transformation. Alex, you were going to say? Yeah.

Yeah, maybe just, again, coming back to this historic counterfactual, one of the things that irks me the most is that so many of the abundance-oriented futures that we want to find ourselves in—like LEV or superintelligence solving everything—just take forever. I do wonder, again, is it 20 years post-Yamanaka? Yamanaka, I think, was 2006. We’re in 2026.

What did we blow these 20 years on? Why couldn’t we have done this 10, 20, 30 years ago?

I’ll answer that. I think when you get multiple exponential technologies that can address this particular space, then that’s the point to invest or put up a prize, because then you get radical outcomes at very low cost. So maybe there’s a benchmark: the minute some domain has 2 or 3 or more exponential technologies converging on it, that’s the point to put dollars into something like this.

Mm. Dave?

Well, there’s no doubt in my mind that I’m going to do more productive work in the second half of 2026 than in my entire life combined up till 2026. So throw it back at you, Alex: even if we had worked really hard on this 20 years ago, would anything that we did between then and today even hold a candle to what we’ll achieve between now and the end of the year?

Because it’s so funny to me. My daughter’s over at Moderna. She’s a biochemical engineer, and she listens to the pod, of course. It’s so obvious to her that we’re in an AGI hard takeoff now, and the people she works with in biotech are about 1% AI-aware and 99% not aware. But the AI-aware people are now spending over half their day, maybe 80% or 90% of their day, talking to AI agents—not in meetings, not running gels, and not running assays—because it’s just a different mode of living that’s accelerating tremendously.

But the aware subset is tiny. I almost think this is an important point, and I almost think we need a new term for it. I’m just thinking off the cuff: maybe “retrospective hyperdeflation,” this very singularity-oriented idea that with superintelligence, you discover that everything that you spent the past decades on was just a total waste, and you should—

Yeah.

—have instead just done nothing, twiddled your thumb for decades, and waited for superintelligence to solve it for you.

Or just work on chip fabs or something that will be very useful on that day. Or buy land in Texas or something. Something like that.

All these people who spent 6 years on PhDs trying to divine protein structures just mostly wasted their time.

Well, that one was really outed by Demis. I mean, that one—it was like an average of 4 years to discover 1 fold, and now it’s—

I knew there was a reason I didn’t do a graduate degree. That was it, because it would’ve been irrelevant.

It would be worthless.

This is your post-hoc justification, Dave.

That’s right. Post-hoc justification.

The past is rigged too.

What’s your take on all this?

It could be worse. You could be a pure mathematician, right?

They’re all cooked too. They’re having this same moment of ennui.

Yeah, but at least with the biologists and the longevity people, you can do assays and things physically. I think that this is the biggest market in the world, as you said, but it’s the first time that it’s tractable.

I think if you go back to when there were the Yamanaka factors and all these other things, you didn’t have the infrastructure necessary or the talent pool necessary. I think that as you’ve seen the various breakthroughs in other things and everything come together, right now there’ll be a shortage of people who can really work on longevity properly, even though it is the biggest market in the world. And 10 years ago or 20 years ago, that would’ve been even tinier. There were only a few people actually looking at some of these things back at that time.

So I think it is this confluence factor all coming together, and, like I said, it’s finally tractable. So the people who are winning in the Longevity XPRIZE—if that’s any indication that those things will work—I don’t think they’ll be short of capital.

Yeah.

But again, the XPRIZE is the catalyst, right? And that was their whole idea.

Yeah, and—

Peter, you’ve pointed out something that I think is incredibly important to highlight, which is that we spend huge amounts of money today treating chronic diseases at the end of life.

Do you know what the number is, Salim? The global—

80%.

—the global cost of age-related disease is $20 trillion per year globally.

Holy—

It’s a huge amount.

Yeah. Daniel Kraft said, like, 85% of health care costs for the last 5 years of your life, something like that. Just a staggering number, right? Because now the business model—

Just for context, the global economy is $120 trillion or $130 trillion total.

Yes. Just look at that number, right? So now the business model becomes: how do you maintain decent bodily function before the disease appears? And it’ll completely change health care economics. Massive impact.

And global economics in general—

Global economics.

—in terms of productivity and wasted capital.

Yeah.

Well, I think this is where the peptides have become really interesting, right? To lose weight, you had to work out. I’m about to start—

Do you know what that is?

—and get rid of the extra chin. Yeah.

Are you really?

Yeah.

You’ve got to report in on the pod.

I’m going to get it reported and see how my weight loss goes.

Which peptides, Emad, are you taking? Do tell.

I was going to try the retatrutide one.

There’s another retatrutide.

But I don’t—

Random proteins?

Retatrutide, yeah.

It’s retatrutide, yeah.

I’m on retatrutide.

But I know I’m going to lose weight, right?

I can hook you up.

How are you getting access to it? No, but seriously. Okay.

He’s in Canada.

Oh, you’re in Canada?

No, no, I’m not in Canada. I get it in the U.S. It’s totally legal.

In the U.S. Yeah, there are drug dealers out there. Black-market drug dealers.

Peptide dealers. Right, there we go.

I guess, Salim, that’ll just be our little secret. There’s no one paying attention to our sessions anyway.

Nobody’s watching.

Nobody’s listening.

But this is how you know that you’re going to lose weight with it, right? And that’s a first. So now the concept that you can take a pill or an intervention and live longer—well, you’re already losing weight like that. And so I think, again, that’s a big awareness that’s caused the longevity market to get even bigger.

Yeah, GLP-1s are the first longevity drug, in a lot of people’s opinions, and it’s one of the biggest-grossing drugs, if not the largest one, in human history.

I’ve gone even further than that, just a bit of back-of-the-envelope calculations suggesting—and again, this is not medical advice—that GLP-1s, especially 3rd- and maybe 4th-generation GLP-1s, may actually be, when I refer to LEV—longevity escape velocity—as being potentially spiky, a pretty big spike.

There have been studies done recently on some of the 3rd-generation, if memory serves, GLP-1s that suggest—and again, not medical advice—that we, or at least some subpopulation of humans who have undergone GLP-1 studies, may be at something like 70% LEV just with GLP-1 therapy. So if that is the case, and again, I encourage folks to do their own independent back-of-the-envelope analysis, that’s one heck of an LEV spike in a subpopulation that’s being administered GLP-1s.

Yeah. Anyway, watch this space, everybody. It’s exciting.

Peter, I’ve got one more question for you.

Yeah, please.

As you looked at the finalists—because I couldn’t make it to the finals, but we’ve been tracking some of the teams—you had a great snapshot view. Are you of the view that we get to LEV by 2030 or earlier?

So here’s the idea, right? If any of these teams win and can reverse your functional age, this is not a number from a particular blood test you take that changes on the back of a form.

This is: Are you feeling functionally younger? Do you have better muscle-building capability, better memory, and a better immune system? What really matters is your function. If we can do that—if we can actually reverse the clock by 20 years of function—then you get to enjoy the next 20 years of breakthroughs.

If you don't believe we're going to have incredible breakthroughs over the next 20 years—world simulators in AI, the impact of quantum, whatever that might be, in cell biology and understanding how we age—then you're missing the boat. Your goal is to keep yourself in the best health right now. What does that mean? Sleep 8 hours, unless you're Emad and Alex; you probably don't sleep more than 4 hours. But you're short sleepers. Sleep—

So—

—as much as you need to.

So I'll make a comment here, which I've made before: the business model of religion is to sell heaven. As we have life extension coming, how are you going to sell heaven if people aren't dying? So this breakthrough will mean that religion is cooked. We have forever to figure it out.

There will still be a lot of religions doing very well from people tithing. But again, I guess the advice right now is: do what you can to keep yourself in the best health. Don't die from something stupid. It's sleep—

Don't get hit by a bus.

—exercise, proper diet, and mindset. Mindset is so important. I think my greatest attribute is my longevity mindset. So take it on. We'll keep on reporting in this space. It's an important part.

2. Hollywood Gets Rebuilt

But let's move us on. These are a couple of stories, Emad, that you shared with me. These are 3 stories that landed this week together and really describe, I think, the collapse of Hollywood economics and the rise of something much bigger.

Our first story is a company called Higgsfield that just made a movie called The Cully Hill Boys. It's a 110-minute feature film, the first full-length AI-generated movie with licensed celebrity likenesses. Here are the numbers: the total budget for this full-length film was roughly $2 million; the team size was 28 people; the production time was 4 weeks; and the compute cost was $1 million. They used Seedance 2.5 for the generation, and they open-sourced all 10 steps in their workflow so anyone can replicate it.

Here's the context. A feature film with celebrity talent today costs between $20 million and $100 million if you're not using AI, and it takes a year to 18 months. Higgsfield did it for 2% of the cost and 6% of the time. I want to share a short video, a little clip from Higgsfield, so we can appreciate it.

What the fuck is that?

Fucking spices.

Supposed to be over £2 million packed in this little nut.

Who smuggles coriander in a fucking boat?

Someone who suspects it might get nicked. Or where'd your boys go after you left the pier?

What the fuck is that supposed to mean? We came here and waited for you fucking pricks like we were fucking told.

Let's go through it step by step, shall we?

All right. Our second story: Bloomberg reported that 9 out of the top 10 text-to-video models on the Artificial Analysis leaderboard are coming from China. This is a headline, but the real story is deeper, and I'm curious what you guys think about this.

Hollywood has forever been the dominant producer of films, and it's basically exporting U.S. culture to the rest of the world. What happens when the world is flooded by Chinese-produced, English-speaking films? How does that sway public interests and public points of view?

The second thing is that these Chinese models that are dominating video generation are also learning physics, motion, object permanence, and causality, which is what's needed for robotics and autonomous driving. So things are moving quickly there.

The third one, very quickly, then we'll go into the discussion here. A model called LTX 2.5 is the newest version of the most popular open-source, state-of-the-art video-generation model, and it works on your MacBook Pro, which is extraordinary. I want you to imagine something that allows the individual to produce their own movie. One more clip, and this is from LTX 2.5.

This is LTX, the most downloaded open-source world model, and today it just got better. Introducing LTX 2.5, now with diffusion-fidelity rendering, a new way to generate video. Instead of locking every scene to one compression rate, our model allocates compute by scene complexity and budget, rendering flawless detail where it matters and staying efficient everywhere else. It generates fast enough to run live inside a game or a simulation, or power a live avatar. Real-time products are already earning off worlds built on it.

So, gentlemen, let's go to you first, Emad. This has been your world for the better part of a decade.

Yeah, it's happening right on forecast: real-time, high-definition video. LTX can generate a 10-second clip in 7 seconds at the quality you just saw, which is indistinguishable. A few years ago, when we had the state-of-the-art model, it was not like that. It was slow-moving. Will Smith eating spaghetti was awful and horrible.

Yep.

Now you can do a spaghetti-eating Will Smith, and it will do it instantly on your local laptop. So you've heard about the disruption of Hollywood: you never need to reshoot a scene now with the workflows that you have. People are licensing their likenesses. As you said, Higgsfield released an 80-page guide on exactly how they made the movie. It's almost open source.

Higgsfield was founded by an ex-Snap leader from Snapchat. It's at a $700 million revenue run rate right now, in about 1.5 years, to give you an idea of how much uptake there has been. But that's just a start.

That's crazy.

Because every—

That is an absolutely crazy number.

Yeah, because every pixel will be rendered. If you see Reactor, who are using LTX now, it could be that Alex might be a simulation already, but for the rest of us, we could put our avatars live in the next month or 2 at a level that you won't be able to tell. It's a fair cop.

What's the compute budget to make a 1.5-hour-long movie? What will that be by September 25th, by the time we have the Moonshots Live event?

It depends on the level of quality. Seedance 2.5 is the best model. It's a large model that costs about $3 per 30 seconds, roughly. But you need lots of shots, and it can have 50 inputs of audio, video, and other things. The cheaper models are 10 or 20 times cheaper, but at about 90% of the quality.

It all depends on exactly how you're shooting and the type of shots, because Seedance can do 30-second clips, but the average movie at the Hollywood box office right now is 3 seconds per shot.

Wow.

So you could have a 10-times reduction in cost. You could shoot a movie just like that one probably for $100,000 in compute, and then $10,000 in compute—

Mm-hmm.

—probably by the new year.

What we saw in our Foundations of AI Ventures class at MIT is that it went from PowerPoint demos to fully functioning products in 1 year. To be competitive on Demo Day now, you have to actually build the entire product.

But I suspect the movie—our target for September 25th is to bring a script—

Well, the Future Vision XPRIZE that we're going to be awarding with all the teams here at Moonshots Live—you can learn more about it at the show's website. We have 5,000 entries, 5,000 people creating 3-minute trailers and a film treatment.

I just had a meeting this morning with the team at Range Media and Google to down-select, and we'll down-select to, ultimately, a top 100, 50, 25, 10. The top 5 will be at Moonshots Live on the 25th. We'll also have a livestream of the event. We're going to do this year on year, so every year it's going to get better—and cheaper, for sure.

I think this is going to be much bigger than we originally envisioned. I know it's a big vision to start with, but we were thinking, “Okay, then it'll be a $20 million, year-long endeavor to turn it into a real feature-length film.” But in reality, you're going to unleash the creativity of 5,000 people who almost all can make their movie within a year.

100%. We have a $10 million budget to make the winner's film between the money we're awarding and foreign film rights. I was just talking with Range. While we're going to be having 5 finalists and we'll crown the winner—and you should see the trophy, it's beautiful—they want to make all 5.

So we want to create an engine here. Salim, what do you think the implications of this are? And then, Alex, I’d love your thoughts as well.

I’ll go at it from an EXO perspective. We actually talked about Hollywood as one of the first domains that went into an Exponential Organizations model, because what happened when you broke up the Hollywood studios in the ’50s and ’60s is that Hollywood turned into a cluster of external resources where a movie production would start, and essentially a swarm of people would appear—grips, camera people, actors, editors, and whatever—and they’d get together for that project, and they’d completely disband after that, right?

We see the beginnings of that in Silicon Valley now. That was the first wave of Hollywood becoming an exponential organization. Everything was assets on demand; everything was staff on demand. But now, when you have everything being driven by AI, it goes through the organizational singularity, and essentially compute cost is now moving closer and closer to what Alex always talks about. You have domain collapse coming along, and so that’s going to completely change the game again for this. You’ve gone through 2 big waves in Hollywood, the second one just starting now.

Alex, what are you excited about here? What are your thoughts?

So I made myself watch Gully Hill Boyz, and I must say, just as a preliminary—

The whole thing?

I skimmed through the whole thing, watching a good chunk of it.

Skimmed through.

See, when you only sleep about 2 hours a night, you can just do that. Yeah, sometimes.

So I had to watch this. It’s not my favorite genre. I would characterize Gully Hill Boyz as sort of British Bollywood. I’m not quite even sure what genre this is. It seems to be, as far as I can tell, about the hijinks of a bunch of British rappers who get into all sorts of trouble.

Content-wise, it’s not super interesting to me, but at the functional level, it is really interesting. There were multiple times in the movie when I had to wonder: The likenesses were based on real humans, but were the scenes being generated, given the complexity of the interpersonal dynamics, such as they were in this movie? Was there some sort of emergent theory of mind that was almost necessary in order to generate some of these scenes with people interacting with each other?

And especially generative violence. There were multiple times—people with knives chopping things, chopping meat, threatening other people—when I had to wonder: Is there, at some level—we talk about AI personhood on this pod from time to time—is there, in some sense, at some—

Presumably, there’s an intermediate layer in a diffusion transformer somewhere deep in the bowels of Higgsfield. Is there some diffusion transformer or similar model that felt threatened at some point in terms of these generative fight scenes? So that was my take on the content of the movie.

Wait, you’re worried that the residual AI might have been threatened in the making of the movie?

Correct.

I’m trying to get my head around that. Okay.

Yeah.

AI rights.

No natural person’s harmed, obviously, in the generation of this. But I do wonder—

So we’re going to have a disclaimer on the content: “No AI was harmed in the making of this—”

Or traumatized.

“...of this.”

Or traumatized.

Or traumatized.

I mean, I do worry, parenthetically, about that. The bigger question on the economics of this, I think, is: At what point do generative video capabilities start to reconverge with the Anthropic school, which I’d characterize as token-revenue-maxing?

Right now, it seems pretty clear that if you’re using Seedance models, it doesn’t matter how many millions of dollars of XPRIZE money are going to shower down on folks who can create spiffy videos about the future; that’s nowhere close to the amount of money that one can earn, in principle, with revenue-per-token maxing.

And right now these appear to be 2 separate lines of effort. On the one hand, we have a vibrant, largely Chinese-dominated consumer economy on the training side for generating consumer videos. On the other hand, we have enterprise revenue-per-token unit-value maxing that seems to be largely going to coding and enterprise problems.

And right now these are largely 2 distinguishable ways to burn tokens, or diffusion-transformer equivalents of tokens—FLOPs. Two different ways to burn FLOPs. One of them maximizes revenue; one of them maybe maximizes consumer engagement and wow factor, but they’re nonetheless separate.

I don’t think they’re likely to remain separate that much longer. The reason is that I use Fable every day, and I use its competitors every day, and I will say the strongest models—the models that are strongest at revenue-per-token-value maxing—are still terribly weak at modeling the visual dynamics of the world.

I don’t want to call it physics because it’s not physics, although a lot of people call it physics. It’s, at best, classical mechanics. But the physical intuition from general-purpose video generation requires that you at least have some embodied intuition, and right now Fable and its peers are just incredibly weak at that.

I think, in order to ultimately revenue-max per unit token, it’s going to require that these Fable-esque models have just amazing visual intuition as well, and we’ll finally see a convergence or reconvergence of these 2 lines.

Yeah, one of my hot takes on this is: Are we going to see the primary actors out there—the Matt Damons or the Leonardo DiCaprios—licensing their likenesses? And I think not.

Yeah.

But there are so many lookalikes out there. The producer’s going to go and say, “No, that’s not my Matt Damon. That’s John Smith, and he looks like Matt Damon, and we licensed him, and he’s in this movie.” Right?

Oh, my God.

And I think that’s going to happen. That’s going to be the workaround for getting the actors you know and love into this, and there’s nothing they can do about it.

There’s an alternative, which is what we’re already seeing: dead actors. Dead actors who can’t record any new movies have estates that are highly incentivized to license away their likeness for this purpose.

So I think dead actors—maybe we’ll see an equivalent of SAG-AFTRA pop up just for dead actors—and they’ll be the most profitable actors in Hollywood: dead actors. I think that’s likelier to happen.

How much do you think people will care if you can make a near-Matt Damon—a very similar character, but obviously not him—or if you have the actual Matt Damon? How much will people care in terms of box office?

I don’t think they’re going to care. They—

Yeah.

Because you don’t care what the person’s named in the movie; you just like that actor. That actor brings you good feelings from previous engagements.

Huh.

Yeah, I think that you’re also seeing the rise of AI stars now as well. They’re winning deals, and I have actually had this discussion with various film stars where they’ve been like, “That SAG-AFTRA deal has massive holes in it. 80% of me plus 20% of my character can be licensed by the studio.” You know, where does the person stop and the character start? Because obviously they have the character rights and things like that.

Another thing I’d like to say—one thing I found very interesting is that, as we were doing some of the more interesting frontier work, one of the things we found very useful was to get the AI models to generate images and visualize what they’re doing using something like GPT Image.

What does that mean?

So even if you’re doing—

What does that mean?

So if you’re doing, like, Salim’s doing an organizational paper, for example, on Exponential Organizations, it’s doing text, text, text, and then you tell it to generate a visual of everything that it’s done and analyze it in any way that it wants, and it almost moves it to another frame of reference—

Mm.

Because it’s pulling in from this visual-cortex kind of thing, and then you can tell it to expand and collapse it, and you get these really weird images sometimes. But you can see, actually, it’s exploring different parts.

So I think we’ve seen that work for some very interesting things, and I think it fits with what Alex said, as all these models come together to create value, as it were.

I’m excited—

And continue learning.

...by interactive movies, right? If you can generate faster than you can view it, then you can have a movie that’s actually measuring your emotions and changing as you’re viewing it, which I find fascinating.

Well, Peter, it’s what we’re seeing. I mean, there are popularly branded world models, even though they’re really—

Yeah.

...just interactive video-generation models. That is what we’re seeing.

Yeah.

Yeah, so that’s what the LTX model is right now.

Mm-hmm.

So that was the first model that could do it at high definition.

And then when that's combined with frame generation on the latest graphics cards, what you've just described means that, before, world models were playing like blocky video games. Now you can have an interactive Elden Ring, or whatever you want, as of this week.

Yeah.

And I do think, Emad, you said something a few podcasts ago: we're going to end up with a world or frontier model running on a MacBook Air. This is a very specific use case. Are we on track for that? We've got this thing running on a MacBook Air.

Well, yeah. Again, it's very slow to generate, like a few tokens a second. But now we're getting to the point where frontier-level models are coming here. Video frontier models are around 20 billion parameters, yet they understand all this and can generate in 2K.

Language and code models are obviously a lot bigger, although you're about to have the new Qwen dropping in a couple of days. But I think it's all going in one direction because we're optimizing the heck out of these things. Ultimately, what a model is, is an input data distribution that gets compiled into weights, and the data going into these models is getting better and better and better.

Awesome.

We had the whole State Street executive team here yesterday in the studio, and we took them through the Holodeck. Word to the wise: the Holodeck's AI, Amber, just starts talking to you and says, “You can build any movie, any song, any code. What do you want to do?”

It's way too open-ended, and then the answer you get back is, I don't know, a hip-hop song with no words. We're like, “Okay.” So you need to actually create the virtual environment or the movie scene, draw the user in, and then have them guide the movie in the direction they want to go.

Just having it auto-generate from your thoughts is too free-form. People don't know how to even start.

People freeze up. Yeah.

They freeze up. Yeah.

Maybe just a closing thought on this one, Peter.

Please.

I do think the recent launch of Opus 5 is a step in the right direction toward seeing a convergence between frontier models on the one hand and video-generation or world models on the other. We talked on the podcast a bit about how Opus 5 seemed almost mildly benchmarked toward front-end development and the loop between visuals and code.

I interpret and construe that as the early signs that Anthropic, and probably other labs as well, are feeling economic pressure to produce models that do an absolutely amazing job of visually reflecting on their own chain of thought. When they produce, say, a website, they then do a visual analysis of their website that feeds back into their chain of thought, and they do multimodal reasoning over their own visual outputs.

Ultimately, say in the next few months, as the ability to produce what used to be considered triple-A-level video games becomes standard fare for what people expect from frontier models, that will be the ultimate forcing function for forcing Anthropic-type frontier models to have absolutely amazing visual capabilities—even if Anthropic can't be bothered to produce direct video-generation capabilities, and even if it can indirectly reproduce Counter-Strike.

All right.

Yeah, that's why they had Claude of Duty. People were making—

That's right.

—Call of Duty with Claude.

Call of Duty, exactly. And we have covered Call of Duty.

3. Grok Chases The Frontier

Our next story is one that both Emad and AWG texted me about this morning. Elon just dropped Grok 4.6, and he's right: it's a banger.

xAI's latest model matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index at 61, tying for frontier-level performance at $2 and $6 per million tokens for input and output. The focus this time is long-running agents. Grok 4.6 stays with complex tasks across many steps, whether researching, coding, analyzing, or turning a broad product idea into a working first version. It self-tests and verifies its own work before moving on.

It's available today in Cursor and Grok Build. The model cadence is crazy, right? Grok 4.5 came out 2 weeks ago, 4.6 just this morning, and 4.7 is rumored to be coming out in 2 weeks. Incredible. Gentlemen, Alex, over to you first. Yeah, I want to give Elon applause.

Because you've been merciless on xAI for the last few pods.

Yeah. I wouldn't characterize myself as merciless. I think my job here is to call balls and strikes as I see them, without favor or prejudice. That's how I see it.

I view Grok 4.6 as essentially the next version of Cursor. xAI, SpaceX, and Elon have been quite public about how 4.6 leaned heavily on post-training, thanks to the Cursor acquisition, which I think is still in the process of being completed as we're recording this.

But history rhymes quite a bit. We've spoken in recent podcasts about what the Chinese frontier labs are doing and how they're allegedly distilling reasoning traces en masse from Claude and other Western models. In some sense—and again, this is an outsider's perspective—I view xAI's acquisition, and even prior to the consummation of the acquisition, its licensing of all the reasoning-trace data from Cursor, as essentially pulling a Westernized version of what the Chinese frontier labs were doing.

That is to say, siphoning off reasoning traces from lots of people historically interacting via Cursor with Claude and Claude's competitors, and then using that incredibly valuable reasoning-trace data to do post-training on their models. I think now that we're in the reasoning-model era, and we're a couple of years in at this point, those reasoning traces for mid-training and post-training are just so essential in terms of catching up to the frontier.

They won't get you past the frontier, so it's sort of a one-trick pony in terms of nearly catching up. But it's a heck of a one-trick pony. The other thing Elon has going for him that the Chinese frontier labs don't is the compute. He has the NVIDIA GPUs, and soon he'll have his own Dyson swarm with NVIDIA GPUs—something that all of the Chinese frontier labs pulling the same trick, siphoning off allegedly Western reasoning traces to do their own post-training and distillation, don't have.

So the bull case for the Elon strategy is that he gets the algorithmic insights to catch up to the frontier from the Cursor reasoning traces, and he gets the compute advantage that the Chinese labs don't have. I think the question is not whether Grok 4.6 and its successors can catch up to the frontier. It seems like they can because they have xAI, or near-frontier capabilities. To the extent that they have the reasoning traces, the question is whether they can leapfrog the frontier and achieve state-of-the-art performance.

Emad, what do you think about that? Can they?

Yeah, I think they definitely can. We were discussing this before, and Elon has come out and said it publicly now: he thinks 4.7 will go above Opus, so it will take number one in a couple of weeks.

Right now, 4.5 had 1.5 trillion parameters and was then post-trained to become 4.6, just like Cursor originally took Kimi K2 and post-trained it with 3 times the amount of compute that was used to pretrain Kimi K2, achieving almost top-level coding performance. The next model is going from 1.5 trillion to 2 trillion parameters. That's 4.7. But 5 is coming in at 6 trillion and then 10 trillion parameters.

So it's going to be whirring away, and through scale and the quality of the post-training data, it should achieve the frontier. But the question is, is it going to be useful? Do you have that kind of knowledge there? In the basic, everyday stuff, which we're seeing with bots and things like that coming out from them now, and then the more advanced stuff, because 4.7—and this is what I think is going to be really interesting—is going to be trained on all the SpaceX physics and engineering knowledge.

Hmm. Yeah.

And that's going to be the real test. As Alex said, you can't get beyond the frontier with this stuff—

Elon does not like being number two.

—but what if you have the best chips?

Elon does not like being number two at anything.

He's the best engineering leader in the world, and he's turned it into an engineering masterpiece, both in training and post-training these models and in building the most cost-effective massive infrastructure in the world.

They're going to add $5 billion worth of compute now, aren't they? Who are people buying compute from? xAI, right? Everyone was like, “Ah, that's him falling behind.” It turns out he had a plan after all.

Dave, what do you make of this?

Actually, I'm curious, Emad—or Alex—if you have any insights on Grok 5. Grok 5 was supposed to be out in May, I think, and it's now August. That was going to be, as you said, a 10-trillion-parameter model. It's a huge step up from anything we're talking about here, but it seems to be behind. Is that just because training at that scale causes the NVIDIA chips to fall apart?

It's because he fired everyone.

That’s why.

Oh.

The original xAI team—he got rid of them, and he brought in Cursor.

Yeah.

He paid $10 billion for the data—

He tends to do that.

But did he do it?

He tends to mass-fire and then build up again.

Yeah. So these new chips as well, the B300s—I mean, how old are they, Alex? It takes a while to bed in and write the actual training code, which now they have.

It has been a while, and he’s also promising with 5 and otherwise to do something that I’m not hearing from any of the other frontier or near-frontier labs. He’s made some public comments, I think, in the past about wanting to start a new pre-training session approximately monthly, which you don’t hear anyone else talking about.

Normally, a more conventional cadence would be quarterly or annually. In Gemini’s, Google DeepMind’s case, it’s certainly on an annual basis. But starting a new pre-training session every month—that’s shooting the moon. But we’re in the Moonshots business, so we’ll see whether this works.

Yeah, the code to train that caliber of model is actually very straightforward now, thanks to Fable 5 being out in the world. We’ve trained a 48B internally here at Quantum with no trouble at all.

But the problem you run into is trying to get 100,000 GPUs to do anything constructively together, and that’s something that the Chinese and any small lab just can’t. You can only learn that in one place, and that’s in Tennessee.

Yeah.

Wow.

So basically, at 20 billion active parameters, you start getting problems, then you get it at 70, and then you get it at 200.

Mm-hmm.

And so, again, it just takes a while to bed in and really get these chips working, but also hitting the price points that he wants to hit. As you said, it’s $6 for this versus $60 for GPT-5, right?

Yeah.

Elon wants to keep that price point.

Amazing.

I think one thing I noticed that was very interesting here is that it looks like xAI’s vector here is that they’re optimizing for persistent AI teammates, which have lower-cost reasoning and so on. I think that is a really powerful model, because now you don’t talk about having a smarter LLM. You’re basically saying, “Hey, here’s an AI coworker.”

This is Macrohard.

Yeah. I think this is going to be—he’s bringing those 2 together over time. This is what I saw when I looked at the details of this. And that is really interesting.

The other side of this story, and I’d love to get Emad’s take on this, is that it paves a path for sovereign AI. Catching up to the frontier now is almost a routine, doable thing. And like you said, you can’t get past the frontier because you’re borrowing everybody else’s reasoning traces. That helps you a lot.

The open source from Kimi helps a lot. You can use that as a starting model and just tune it to whatever your national goals are, and you’re up and running in 6 months, 5 months, something like that. So it does open it up. And also, large corporations that otherwise would’ve been intimidated as hell have a roadmap now to being competitive with their own proprietary models.

Yeah, I think that if you look, the other release was bots. So this was a Cursor thing where they basically took OpenClaw, gave it its own computer, and now it’s Grok. It’ll spin up hundreds of different bots.

One of the things you can do is this: You hit the record button, you do stuff on the screen, and it turns that into a skill automatically.

This is Macrohard.

Yeah—

Right? That’s what they want to do: go into companies, basically record everybody’s workflows, and then give you a digital version of your company.

And he has all the GPUs to do that effectively, right? But again, at the price point. So he’s not going to budge on the price point. In fact, he’s going to be the market dominator, driving the price point down, along with the Chinese.

So I think you can have your DeepSeek Flash models, and then you’ve got your Grok models, but that feedback loop and that data and that knowledge is going to be very interesting. And the other thing I think is interesting is that they’ve started tracking all the discussions of research on X as well.

Any moment a new paper or anything comes out, where is it discussed? On X—

Yeah.

—and this podcast and things like that as well.

Yeah.

And that’s such a rich vein that I think it’s going to be immense.

4. Compute Becomes Financialized

All right. Let’s move on. Our next story is out of NVIDIA. NVIDIA just announced a partnership that redefines what GPU compute means as a financial asset. They partnered with Apollo, BlackRock, Blackstone, Brookfield, and KKR to mobilize over $500 billion in third-party capital for AI infrastructure.

Importantly, NVIDIA is not borrowing $500 billion. They are creating a structural framework through which institutional investors, pension funds, sovereign funds, and private equity can invest directly in the AI compute. They’re helping finance their customers to buy the NVIDIA GPUs.

NVIDIA hardware depreciates typically in 3-5-year cycles, but during that cycle, the compute keeps earning. We’ve talked about the fact that H100s are probably more valuable today than they were when they were first bought. So NVIDIA is positioning itself as the architect for the financing layer, not just the silicon supplier.

Jensen framed it this way. He said, quote, “We began by building chips. Today, we’re helping to create a new class of productive, investable infrastructure, AI factories.” Let’s watch a quick video, and then Dave, I want to go to you for your thoughts on this one.

But this is the story that NVIDIA is coming together with some of the biggest names on Wall Street to put together half a trillion dollars of independent financing to push AI forward and build the AI infrastructure out. This is independent third-party capital that they’re bringing in. These are going to be strategic partnerships.

NVIDIA has signed partnerships with 6 of the biggest names on Wall Street for these memorandums of understanding. Basically, NVIDIA will find its customers that need help with AI build-out and need financing for this, and put them together with these partners that are pledging, again, over $500 billion that they will find to come into this.

So, Dave, your thoughts. Classic?

Yeah, well, look, Kushal Bhagia was on the pod, and he’s pushing hundreds of millions in revenue in less than a year. September will be his 1-year anniversary of founding that company, and it shows you the pent-up demand to invest in this massive, multi-trillion-dollar, $7 trillion and rising Dyson swarm that we’re going to be building.

A lot of people are like, “Am I too late?” I mean, this is like the very first pitch of the first inning of the build-out of the Dyson swarm. And so I think that when you create new financial structures that allow people to pour capital into it, it just attracts money from all over the world that otherwise wouldn’t come in.

Historically, a company like NVIDIA would say, “Well, let’s do a secondary offering and raise $500 billion as a secondary stock offering,” but that’s not going to scale to infinity or to Dyson-swarm capabilities. So instead, Jensen has brilliantly said, “Let’s create new standalone financial structures that are tied to individual compute clusters that people can individually invest in.”

And that does scale infinitely. You can stamp those out ad infinitum. So then you put highbrow names like BlackRock and Apollo on them, everybody realizes it’s an investment-grade asset, and then anyone in the world can pour money into it. So it’s brilliant.

Fascinating. Who’s next here?

I see a couple of downsides here, though.

Yeah, Salim, go ahead.

Right? Larry Fink himself used a reference to mortgage-backed securities. So you could create a lot of liquidity, but you could create a lot of technological risk—

There. Because imagine you securitize, like, 10 years of GPU cash flows, right? And then somebody has a massive breakthrough, like in our previous story, and a new architecture emerges.

Yes.

You’ve got all of a sudden stranded compute assets in a huge way.

So there's a downside to this because if you have such a volatile environment, that is a very difficult thing to have. Financial assets want predictable depreciation, right? Exponential technologies don't give you predictable depreciation.

I had a meeting with Lou Ranieri, the inventor of the mortgage-backed security, in New York. It's so funny, too. We were scheduled to have a meeting, and I just went into the men's room before going into the meeting. I was just talking to the guy at the urinal next to me, which is kind of weird, but I did. Then we go into the conference room, and it's Lou Ranieri from the book The Big Short. He was explaining to me that I had just read the book, and I was like, “Wow, you invented the thing that destroyed the entire world economy.”

Oh my God.

And he said, “Well, thanks.” But no, it's not the instrument that was broken; it's the ratings agencies getting corrupted. The same will apply here. If the Dyson swarm is an obvious good investment, and if companies like Oren rate the investments correctly, it'll just be smooth growth for 10 years or more.

But I think Salim's—

If, however, it gets corrupted, which is very possible, it'll be another Big Short collapse.

Salim's point is different, though. What happens if there's a new architecture, and we're going to talk about one in a minute, that actually makes GPUs less useful, changes the whole game, and you've basically financed something for 10 years or 20 years and you can't earn revenue on it anymore? The game has changed. That's the story of exponential tech. There are nested S-curves. Everything runs out and something new comes along.

Peter, that's brilliant. I 100% think that's highly likely to happen for the reason Alex always says: we're on the cusp of discovering new physics imminently, and some of that is going to be compute-related. So it's a brilliant insight, and thank you for throwing that out there.

Well, I'm just amplifying Salim's brilliant insight here.

So I'll maybe just add: AI, I've been trying to popularize for a while the notion of compute-backed securities, which is, I think, where we're going—CBS, compute-backed securities. But I'm not as concerned that this will end up being another mortgage-backed securities fiasco for a few reasons.

One is that compute is fundamentally much more productive than a house is. The best you can do with a house is live in it, which is moderately productive. There are lots of other things one can live in other than a house. It's also, in some sense, a depreciating asset, requires lots of maintenance, and so on. Compute can require maintenance and obviously requires electricity, but it's fundamentally productive.

Second, with mortgage-backed securities in general, there was a policy, sort of an ulterior motive. There were pressures exerted on ratings agencies to facilitate the American dream of houses for everyone. It's not quite obvious there's a direct analog of that here for compute. One can maybe, finger to the wind, point at some sort of U.S.-China race as perhaps a policy pressure point to lubricate the private capital markets here. But again, I would say it's not quite directly comparable.

The third point is ultimately about the risk that I think, Peter, you're trying to flag: What happens if there's an algorithmic breakthrough, a physics breakthrough, or some other breakthrough that causes hyperdeflation, and fundamentally GPUs that were worth $10,000 one day are worth $100 the next day? This is what options and futures are for. In a sophisticated asset-backed securities market, sophisticated actors should have the ability to hedge their positions and hedge against exactly that possibility, as well as the counterfactual option of China invading Taiwan and driving the prices of compute through the roof rather than through the floor.

Both of these possibilities should be covered—arguably, not financial advice—using a fungible, liquid market for compute futures and compute derivatives. I'm pumping my own book a little bit; admittedly, I have a financial interest, but that's precisely what Oren, which we just had on the pod, is enabling. So I'm a little bit—

Well, the thing I'd add to that, Alex, and I'd love to get your take on it, is that demand for compute is going to approach infinity. There's no doubt about it. These are fundamentally good investments from that point of view.

But the breakthrough that could happen in the next year or 18 months is a way to compute that's lighter—lighter in physical weight—by an order of magnitude or more. Right now, the forecast is that by 2030, a couple percent of all compute is in space via Elon's rockets, and that's entirely gated by launch weight—the mass that he can put into orbit. That formula completely inverts if you knock a factor of 10 off the mass.

A lot of the mass is in the cooling and the solar collector. The chips are very, very light to start with. But any reduction in the power required would reduce the solar collector and radiator weight a lot, and then the entire formula would switch to: all these generators, all these racks, all this land in Texas—it's much cheaper now to just put this lightweight thing into space and collect solar power. It goes right into electricity.

So keep your eyes on physics breakthroughs that allow you to compute at lower mass, because that might completely change your investment thesis.

I think, Alex, the point you're making is that the utility of these chips is going to stay constant. Jevons' paradox comes to mind, and therefore the economics should be much more predictable.

That may or may not be the case, but really what I was trying to express is that in a sophisticated market—a sophisticated financial market where compute-based or compute-backed securities are being traded—you can also hedge. You can hedge against the upside, and you can hedge against the downside.

Having a properly functioning private credit market for compute is necessary for this to scale. I think there's no way that Sam's $7 trillion of AI data-center infrastructure gets invested without sophisticated hedging options. We need sophisticated hedging options for that broader market—the $7 trillion of CapEx—to actually be investable. So it's the hedging options that I'm really trying to express.

I think there's one more factor here that really is being underestimated. CoreWeave just came out and said that some of their clients have taken out contracts through 2029 for A100s.

Wow.

The A100 was introduced in 2020 by NVIDIA. It's a 6-year-old chip. I remember we had 10,000 of them. We had the biggest clusters in the world. That is an old chip with around 40 gigabytes of RAM per chip, or 80 gigabytes, depending on the configuration.

But why would someone do that? Because they have a workload that they see as constant for 3 years that fits on an A100.

Mm.

And what happens is those A100s have been paid for. You've paid for all of the hardware costs of the A100s, which then means it's about electricity turning into intelligence. That's your marginal cost once you've paid off the initial bulk order.

And that is actually improving, because back in 2022, 4 years ago, GPT-4 had just finished training. That was the best model you could have, and that took roughly 16 A100s. Now you can have literally a 10-billion-parameter model or a 5-billion-parameter model that outperforms that. You can fit around 30 of those on 1 chip rather than needing 16 chips.

Mm.

So the ability to convert electricity to intelligence is improving. All these chips—and this is what Jensen said—are running CUDA. So it'll work on an A100. You can run any of the new models on that, as well as on the Blackwells.

We still have enough installed base, and this is the ideal time to financialize because it's before the next-generation chips. It's before the chip breakthroughs. Just like it's now a great time for Anthropic to come to an IPO before Groq comes and takes their lunch. So we always have to play these cycles.

Well, I mean, the definition of “paid for,” too: I just told our quantum team today to buy 3 million NVIDIA GPUs as fast as they can get them—the GBs, not even the future VRs—and we have to wait until at least November or December to even get them.

Mm.

But if you turn around and make an hour-and-a-half-long movie with $1 million of compute, you could gross $20 million or $30 million on a good movie. Well, much more than that if it's a really good movie. So the payback could be as short as—what was it, 4 weeks?

Yeah, it was more like 2 weeks.

2 weeks.

5. Transformers Face A Successor

I want to turn to a story that could be the countervailing force here. It's a new architecture that's climbing ARC-AGI at a fraction of the cost.

So there was a tweet this week from Zuzanna at Pathway AI that flagged something everyone in the AI world really needs to pay attention to. It’s a new non-transformer architecture that is starting to climb the Arc AGI benchmark at a fraction of the compute cost of traditional models. ARC-AGI is the test that actually measures reasoning, not pattern matching.

Standard LLMs, despite their trillion-parameter scale, have historically struggled with ARC-AGI because it requires genuinely novel reasoning on problems the model has never seen. Transformers get there by brute force. They throw enough parameters and compute at a problem, and eventually you squeeze out a passing score, but the cost can be astronomical.

What Zuzanna flagged is that alternative architectures that do not rely on standard attention-based transformer stacks are achieving better ARC-AGI scores using dramatically less compute. This matters because transformer architectures, while dominant, have a known ceiling: the quadratic cost of attention over long sequences and the massive parameter counts required for marginal gains. We’ve talked about this ad nauseam.

New architectures that crack reasoning at low compute costs change the economics itself. If you can get GPT-4-level reasoning for 1% of the compute, you can run it on your phone. You can embed it in every device, and you can make, basically, intelligence free. Let’s take a look at this chart. Emad, I’m going to go to you first. You flagged this particular story. What are your thoughts on it?

Yeah. I’m still working my way through the paper on how they have these neuron particles with their new approach, but it doesn’t actually matter that much. What you’ve got now is really great data sets, and then people are figuring out new ways of basically turning that into intelligence. This is the headline.

And we’ve seen that already even with transformers. You have a DeepSeek V4 Flash model, or the DeepSeek V4 Pro has actually just been released at $0.80. Then you have Grok at $6, and then you’ve got Fable at $50, and they’re all about the same performance. So the question is, which of these architectures will win in order to do a job, and will people really care and switch over?

Because we haven’t seen people abandoning Fable to go to something 10 times cheaper. Why would anyone use Sonnet when you have Luna at a fraction of the cost? But I can say that now that the data has been optimized, the next thing is trying these things out and going up on this benchmark, which I think was one of Alex’s favorites back in the day.

Now it’s been superseded.

Alex, what do you make of this? Is there anything here that—

Admittedly, I have a bunch of hot takes on this one, Peter.

Please, of course.

So I read the BDH-CQ paper, and then I went and read the original DH paper. The DH apparently stands for Dragon Hatchling. I just had to read the original, purportedly post-transformer Dragon Hatchling architecture paper. I’ll give you my hot take: I think it’s a hot mess.

I would expect a decent post-transformer architecture to get simpler and more bitter-pilled, which is to say less feature-engineered, and for the architecture to get simpler and benefit from compute more and more. Looking at Dragon Hatchling—again, this is my hot take—it was just a hot mess. It had particles floating around in 3+1 dimensions. It had attempts to make rules end-to-end differentiable. It had some semblance of Hebbian learning. It was trying to do all sorts of crazy biomimetic things.

I would say this is by definition exactly the opposite of what I would hope for from a post-transformer architecture. The author seemed to be throwing in the kitchen sink of every architectural motif they could think of and then some, and then hoping that what pops out is going to beat transformers. I don’t think it’s going to beat transformers.

I look at the ARC-AGI-1 performance curve and, superficially, you could say this is great. This has moved the cost-performance frontier up and to the left, which is what everyone wants. But it doesn’t generalize.

As far as I can tell, this was some sort of quasi-crazy witch’s brew of different architectural motifs that was maybe focused on ARC-AGI-1, which, as Emad said, isn’t even the frontier at this point. It has a bunch of recurrence and other things thrown in. Of course, if you take a specialized model and focus its degrees of freedom just on ARC-AGI-1, you can achieve better cost performance on it. Of course. But it doesn’t generalize. It’s not simple.

I’m calling foul on this one. That’s my hot take.

Okay, but you have been predicting—

I don’t think this is actually an advance.

You have been predicting there will be something that supersedes the transformer model.

Yes.

Yes.

But critically, I expect it to be simpler, more beautiful, more elegant, and this is not that. That’s my hot take.

I—

Apologies for the hot take.

No, we love your hot takes on this show.

Well, let me ask you a follow-up question to your hot take. When I turn an AI loose on AI research, it does tend to naturally throw the kitchen sink at the problem, and it surprisingly works. But it also generates a hot mess like you were describing. Do you think that’s maybe what this is?

No. I agree with you that I would expect a truly— In fact, there are companies out there that I have some affiliation with and a financial interest in that are pursuing exactly what you’re describing. They’re basically using AI via recursive self-improvement to discover transformative post-transformer architectures that are fundamentally illegible to humans, under the premise that you can only get so far with human legibility of the underlying algorithm.

In my reading of the Dragon Hatchling architecture paper, this was not that. This was a bunch of human-legible motifs being thrown together in a pot, with the aspiration of somehow beating transformers. Again, it seems to me that this is not deeply internalizing the better lesson.

When do we get to something that’s beyond transformers? What’s your guesstimate?

I think we’re there already. We’re there already.

Yeah, this year for sure.

We have MoEs, we have diffusion transformers, we have all sorts of attempts to linearize attention, including Moonshot’s approach to linearized attention. We have attempts to inject recurrence into the architecture. My bet is that we get to the post-transformer architecture not through a step change, but through ship-of-Theseus-style replacement of all of the individual elements of the original Attention Is All You Need.

Love that. Well, speaking about attention being all we need, the AI world is getting a lot of attention from Bernie Sanders. So 2 stories converge this week to create the most serious AI safety confrontation of the year.

First, Senator Bernie Sanders sent a formal letter to the CEOs of Anthropic, Meta, and OpenAI demanding an immediate pause on AI development. His justification was that AI is escaping human control and being used to create new viruses or bacteriophages, as the case may be, which is our next story. Sanders cited each company’s own prior commitments to halt development if safety thresholds were crossed. Sanders wrote, “That moment is here.”

He quoted Bengio, one of the 3 godfathers of deep learning, who said, “This should serve as a wake-up call.” Sanders added a direct threat: “If you do not take appropriate action now, my colleagues and I in the U.S. will.” I mean, quite the threat.

Let’s take a look at his letter for one second and call out a few of the things he said here. Here it is. You can see it online. It’s to Sam Altman, Dario Amodei, and Mark Zuckerberg. “This week we learned, frighteningly, that AI has been used for the first time ever to create a new virus. As you know, this type of development in the wrong hands could lead to a new bioweapon that results in deaths of tens of millions of people.”

He goes on later to say, “The moment is here. AI capabilities have reached a critical threshold. There is a reason why the head of the CIA says that AI models are ‘akin to digital nuclear weapons’ and ‘almost like a doomsday device.’” A lot of fear-mongering here. Let’s talk about this, and then we’ll share the story that comes out of Stanford on using AI for generating bacteriophage designs.

I can go first.

Do you want to go first? Yes, please.

I’ll go first. Yeah. I understand his instinct, right? But pausing AI is just such an absurdly coarse approach to this. The rest of the world is not going to listen. Open models are not going to disappear, and you can’t uninvent things that you already know.

The only way of solving this is what Alex has talked about in the past: you have to co-scale the defensive side and do the same thing. This is the same thing that happened last week with the OpenAI-Hugging Face debacle. We now have attack vectors that are human out of the loop. The defense has to be the same; otherwise, you’re going to have this massive asymmetry, right?

You have to attack exponential problems with exponential solutions, not with stupid ideas like this—not to put labels on it.

Emad, you're in pseudo-European pause mode over there in the UK. What do you make of this? What do your colleagues there say to this kind of letter from Sanders?

We just want to catch up, right? That's why David Silver's lab got $1 billion. We have another lab coming out from ex-DeepMind people with $500 million. Look, the cat's out of the bag. It's too late, right?

Fundamentally, this is it. The adversaries will get more intelligent. We've discussed previously on this podcast how you have to stop the reagents and the input processes for things like viruses, and that's something that's much more manageable.

Yes.

But, yeah, takeoff is scary. DeepSeek V4 Pro—we just got some initial announcements. It scores 83.3 on CyberGym, whereas Mythos scored 83.2. Boom. The capability is open source. That halted everything.

A frontier lab went open source.

You know? Yeah, and that's on the cyberattacks now. Unfortunately, I signed the pause letter 2 years ago because I was like, “Let's take a pause.” It's too late now. So we have to, as you said, build the swarms that defend.

Although it sounds a bit crappy, the only thing that can stop a bad AI is a good AI. We really need really good AIs as soon as possible working for us.

Alex, please.

I think this is fundamentally misguided on multiple levels. At one level, please stop punishing intelligence. I think it's a terrible idea to penalize intelligence. We want smarter people. We want a smarter civilization. Attempting to throttle or pause the development of increasing intelligence is simply suppressing growth and human prosperity.

I think it's fundamentally a bad idea to try to cap intelligence. That's the dystopia that I would like to avoid. That's point 1. Point 2: the ends versus the means. If the goal is to punish or deter the next pandemic, the consensus of the U.S. intelligence community is the lab-leak hypothesis. According to that theory, we had the Wuhan lab leak without superintelligence. We can have global pandemics without superintelligence.

I think it's fundamentally misguided to kneecap ourselves. It's a foot-gun, or shooting ourselves in the head even, quite literally, to try to prevent the next supervirus when we're more than capable as a species of producing superviruses without superintelligence. The focus should instead be—to the extent there's any agita here—on making sure that the AIs and the superintelligences, just like humans, can't create bioweapons at all, not on kneecapping their overall intelligence.

I just think many of these policies are ultimately designed, as much as it pains me to say it, to superficially decelerate the creation of wealth, which I think is a bad idea. But they have the perverse side effect of actually increasing race conditions. We saw that with previous attempts to pause AI—the AI pause from our friend of the pod, Max, with his FLI 6-month pause.

I think, to the extent that the 6-month pause he was pushing on the frontier labs for AI development did anything, it radically accelerated progress. It's a little bit like starving yourself for a bit of time and then binging afterward. If we starve ourselves of intelligence progress now, or at least selectively starve the well-behaved, well-compliant Western frontier labs for a month, a few months, or even a few weeks of AI progress just to appease concerns that maybe we're forestalling bioweapons, all we're doing is allowing every other lab that's not as cooperative with the regulatory apparatus to catch up, creating a far bigger race condition once the pause is lifted. Now we end up in a world that's 5 times more competitive. So I think this is misguided in summary on just about every level.

Dave?

Yeah, I read it the same way. I just want to clarify a couple of things, though. This letter is not written to try to change their behavior or do anything. It's purely a position that Bernie is trying to claim he has been opposed to because a disaster is imminently coming somewhere, and he wants to be on record saying, “I was opposed.”

“I told you so.”

“I told you so.” That's all he's trying to achieve here. When I first read it, I said, “God, what a schoolyard bully asshole. He's threatening 3 U.S. citizens from his position in the Senate.” But when you actually read it closely: “Let me be very clear. If you do not take appropriate action now, my colleagues and I in the U.S. Senate will.” It's totally vague. It doesn't say to do or not do anything in particular.

The one actionable thing in here is, “Stop building machines that humans cannot control.” But as Emad just pointed out—

These particular guys, Mr. Altman, Mr. Amodei, and Mr. Zuckerberg, all went closed source for exactly that reason. If you were to write an accurate and honest letter, it would say, “Hey, China. Stop throwing deadly weapons out into the world with no controls whatsoever.”

But he, of course, has no authority to write that letter, so he doesn't.

Good point, David. You have to remember, the U.S.—what are the numbers? Three-quarters of Americans fear AI, and Bernie Sanders is a politician. He's playing to the populist vote here.

Yeah.

Yeah. I want to turn to the second story here, which is the scientific basis for Sanders' concerns. Researchers at Stanford used the generative AI model Evo 2 to design DNA sequences for a bacteriophage. This is a virus that infects bacteria—not people—that did not exist in nature. They synthesized approximately 300 designs and produced 16 viable phages capable of infecting E. coli.

The engineered phages were effective against E. coli strains that had never evolved any kind of natural resistance to these bacteriophages. A genetics expert called it biology's Wright brothers moment. Evo 2 is an open-source AI model that can design novel viruses at will. You can download it. You can use it. Johns Hopkins biosecurity researchers warn that it is no longer a question of whether a viral genome design will exist, but whether it can be used without enabling serious harm.

So this is a dual-use technology. We've talked about it. If you basically throttle the use of this technology, you're throttling the ability to find cures and new cures for disease. The AI frontier models now have to respond. These guys are going to have to respond, and whatever they say will lead to a legal and political consequence.

As you said, Emad, very importantly, the issue is not the models; it's the equipment to build the DNA-sequence synthesizers, right? The RNA synthesizers. We need to be controlling at that location, right? Those can be controlled, but they're currently unregulated.

Yeah, no. I think it's impossible to control the other side. Actually, I believe we discussed on this podcast before: I said you would be able to create something like this on your local machine. Evo 2 is a 40-billion-parameter open-source model trained on 1 million strains. I have actually run it on my MacBook.

Mm-hmm. So you're the guy.

I was one of the authors of OpenFold and things like that. But the capability is now in everyone's hands—

Mm—

—to create these strains.

To create—

But realistically—

To create the design for these strains, not the actual—

The design for these strains.

Exactly.

And so the only way you can do it is on the other side. This isn't even a frontier model. It's frontier in its specialty, but as the models themselves get smarter and smarter, it wouldn't surprise me if Fable could just spit this out, or Grok 5 could just spit out something similar with a very small training data set, because it understands these kinds of things. So we've got to go to the other side.

Also, I think the way these things are announced is important. People are like, “Why are you creating bacteriophages and viruses and things like that?” To cure cancer, right? The way that these things are covered is also very important in how this is all handled and absorbed by the community.

Restricting biological access to Claude and other things—if you say, “I have a cold,” it's like, “Biothreat,” you know, whatever. That also slows down our progress to cure diseases. So we've got to have better press. We've got to have end-to-end control. We have to be practical on this and not politicize it.

Salim?

I think we've said everything here. Look, this is also a fundamental challenge to the concept of our governance structure. Nation-states can't govern a problem that's this universally global.

Yeah.

There's a fundamental impedance mismatch here that's going to struggle very—

A hot take: nation-states are out of date.

Alex, what's your hot take on this one, pal?

I have a cold take, ironically, on this one, which is I don't think this is profoundly new.

It’s wonderful that we’re able to do base-level generative AI for bacteriophage synthesis. That’s great and everything, and I expect it to have ample medical and research applications. That’s all great.

By the way, bacteriophages are an incredible mechanism to cure all kinds of bacteria, septicemia, and things.

Mm-hmm.

I mean, they’re very useful—

Yeah.

…as tools.

Note, they never mention the positive potential here.

Yeah, yeah.

I think that’s all great and everything, but 20-plus years ago, I remember at MIT, in the project that ultimately, I guess, in some form became Ginkgo Bioworks, there was a project at MIT—I think this is circa 2002 or 2003. There was the BioBricks Foundation project. We saw the early rumblings of synthetic biology as a modern discipline. We were designing custom genomes using building blocks, and it was much more manual. We certainly didn’t have modern generative AI, and we were able to accomplish wonders and build circuits.

So I think base-level generative AI from foundation models trained on large amounts of biological sequence data is great and everything. But this is my cold take: I don’t want to oversell the underlying novelty here. We’ve been in the business for decades of creating synthetic organisms, including synthetic bacteriophages, so we’re gaining incrementally better ability to achieve custom effects. It’s more incremental, I think, than anything else.

Where I’d love to see the agita over, “What if someone creates the next super-bacteriophage?” I’d love to see far more devoted to putting DNA and RNA sequencers everywhere.

Yes.

That’s one of the lessons I think we, as a Western civilization, didn’t learn enough from the pandemic: it’s getting so cheap now per base pair to just sequence. You can go out and buy a MinION, a little USB device, plug it into your laptop, and immediately, for de minimis CapEx, start sequencing genomes to your heart’s content right off your laptop and spend at most a few hundred dollars doing that. I’d love to see these everywhere—

Yeah.

And yet they’re not everywhere.

What Alex is talking about is, you know, a pandemic moves, at best, at the speed of an airplane—500 or 600 miles per hour. But imagine if you have these sequencers in the air vents in every airport, every bus station, and every train station.

Yes.

You detect a novel sequence, sequence it, and raise an alert. Then you know exactly where it’s going, where those airplanes are going, and you can transmit a vaccine at the speed of light to everyplace else.

Exactly.

Yeah.

And we have—I mean, this is, in my mind, the killer app of DNA sequencing: too cheap to meter. It’s not personalized medicine. It’s literally putting a DNA sequencer on every microchip everywhere, in the country or on the planet. That’s the ultimate defensive co-scaling strategy, I think, for this supervirus-scare scenario.

And AI can generate a vaccine in a heartbeat.

Moderna did it.

Yeah, exactly. Dave, you want to weigh in, or are you good?

Well, I’ll say what I always say, which is that you can’t cut off every threat at the output level. The way we police uranium, we cut it off at the uranium, plutonium, and centrifuge levels, and that’s where we measure the world. But once somebody has fissionable material, it’s impossible to stop them from making a bomb, because the remainder of the process—the thing that implodes it and the container—you can’t ever police at that level.

The equivalent in AI is cutting it off at the prompt and token level. It has to be monitored. That’s the only future I can see that’ll actually work. So we need a global agreement to monitor all prompts, and then you just have to decide what regulatory authority is allowed to see which prompts.

Hard to do on your MacBook.

I mean, you have to find a way. Talking to Apple about installing it would be trivially easy, but there’s no other way, only because Alex is right. New physics and new science are going to be created at an insane rate. So even if you manage to put virus detectors on every laptop in the world through some magical process, some other threat will be discovered every single month forever hereafter.

You can’t contain them all with afterthoughts. You have to look at what the AI is doing at the activation, prompt, and chain-of-thought levels, and then monitor it all. It’s so cheap to archive it all.

All right, well—

Then we can debate which country gets to see it, or which department gets to see what. We can debate that for the next 50 years, but at least you’ve got it.

Maybe one additional point, Peter, just—

Please.

—to generalize Dave’s comments. I think there is this notion of defense in depth, and any individual defensive layer is permeable; it’s soft. But in principle, if you have multiple layers stacked on top of each other for defense, you get effectively a hard layer.

There are other layers that we rarely talk about on this pod beyond intercepting at the prompt level or intercepting at the real-world action level. There’s the premeditation level. In the context—not to put too fine a point on it—but this has been publicly reported: on the uranium side, there is a vibrant set of intelligence-community counteroffensives.

If you’re a threat actor and you want to try to purchase uranium—it’s not quite an open market—but you want to try to purchase it, almost all of the offers, almost all of the sellers of uranium, will actually just be plants by the IC to basically run a sting, a counter-sting operation to intercept you ahead of time. So if you’re a would-be terrorist and you want to go purchase some uranium, odds are you’re going to discover that you’re going to be targeted by a sting operation to discover who you are.

My point with that parable is that there are other layers even earlier in the intent workflow, even before a prompt gets entered: someone or something has the idea that they want to do something bad, with a capital B. Defensive co-scaling applies there too, just as it does with humans on humans, with nuclear or fission-based weapons. Similarly here, preemption, with AIs detecting early-stage intent by other humans or other AIs, I would expect to be just as effective.

Totally right.

And Peter has said many times that privacy is dead.

Yeah. That’s a point I was going to make.

It’s not coming back.

Privacy—

Yeah.

—is dead, and there’s a benefit to that, which is that malevolent actors are going to get heard, seen, and caught.

And everyone gives up their Bitcoin private keys, right, Peter? Because privacy is dead.

Well, let’s not go there.

6. Watermarks Meet Their Match

Two stories this week about the watermarking infrastructure to distinguish between AI-generated content and human-generated content. Anthropic announced that it will be embedding invisible watermarks in all text generated by its AI models and attaching metadata to files to help discern AI-generated content. The watermarking will be embedded at the generation level, meaning every piece of text that Claude produces will carry a statistical signature that can be detected by appropriate tools, even if the text is copied and pasted and lightly edited.

Our second story comes out of the European Union, which is launching an AI icon and labeling system for AI-generated content. The EU system will require platforms to label AI-generated content so users can make informed decisions. This follows the EU AI Act’s provision on transparency in AI models.

If you guys were watching X over the last 24 hours—and it’s been hilarious—as soon as this new Claude watermark system got put in place, there have been multiple people out there saying, “Hey, remove Claude’s invisible watermark.” Here you see it. These are 2 of the posts. I’ve seen about a dozen of them. Everybody’s coming out, and I love this one from Michael Angel Duran. He says, “It hasn’t been 24 hours, and someone has already created a skill that removes the watermarks from Claude, Gemini, and OpenAI.” Comments on this. Emad, you’re closest to the European Union. What are your thoughts here?

Oh, man. When we were creating all the media generators, all the authorities kept telling us to build in watermarks, and we had whole teams doing this. It’s so difficult— incredibly difficult. You get very weird things that happen, like some of our pictures would give people headaches and make them feel very unwell.

Oh.

I feel that now when I’m talking to Opus 5. There’s something about the way it talks that really pisses me off, and I think that’s the watermark that’s in there.

Huh.

Interesting.

Again, you can see all these very interesting statistical things. At a high level, it’s the em dash, it’s the not X, but Y. We see these patterns, and we’re like, “Why on earth? That’s clearly something in there.” Scott Aaronson and others have worked on this as well. But I think ultimately it’s an elusive thing, because if you’re a bad actor who wants to get around it, yeah, it’s words. How are you going to do that?

I think there’s a much more subtle and much harder problem here: nothing will be purely AI or purely human.

Yes.

I mean, I read something, AI restructures it, I rewrite half of it, and AI fixes it again. Where do you put the icon? This just seems a ridiculous approach to try and solve something.

You’re gonna put icons on everything.

Yeah. Or have an AI whisper at the end of the thing that takes the AI input and whispers it out.

Alex.

I think, to comment on the EU AI Act first, this is as silly a maneuver as the cookie banners were. I didn’t understand the cookie banners, and I don’t understand this. I don’t understand it so much that, in this morning’s Innermost Loop newsletter, I had the banner image literally just be “AI-derived, AI-generated” all over and over again, and I could care less whether people conclude from that that I’m actually an AI or not.

I think fundamentally this is an attempt to take us zooming right past the Turing test and turn back time, like somehow we’re going to live in the before times by somehow seemingly ghettoizing or isolating AI-assisted or AI-generated content behind some sort of would-be warning label. I just think it’s fundamentally a regressive move. Like the cookie banners, these cookie warnings will not stand the test of time.

And then for Anthropic’s watermarks, I just think, again, this is an attempt. On the one hand, you could say, well, watermarking—that’s an honest-to-goodness watermark that’s transparent to human perception. How could that possibly be a bad thing? I think watermarks are going to end up being weaponized and counter-weaponized in the same way that we’ve seen many book writers—this is, we talked about this a bit on the pod—paper book writers who don’t want the prose in their book to be consumed for pre-training of models, reportedly introducing prompt-injection attacks that are invisible to humans but quite visible and deleterious to AI models.

I think we’re about 5 minutes away from bad actors weaponizing these watermarks to do bad things. Fundamentally, having side channels in text—in content that’s intended for humans but that has information intended for machines and invisible to humans—is a breeding ground for bad outcomes. Google discovered this the hard way with SEO and with deciding which features in Google Search rankings to pay attention to.

They learned pretty quickly the hard way: don’t pay that much attention to human-invisible metadata, because it immediately becomes a breeding ground for scams, reward hacking, and gaming. Instead, pay more attention to the human-visible features, because ultimately, to the extent your users are humans and not machines, that’s where the real signal lies. Otherwise, the free market penalizes it.

So again, I’m not a huge fan of this. I think, in the best-case scenario, it ends up being net neutral and neither strongly positive nor strongly negative, but it smells like an attempt to turn back time.

Yeah. I challenge the idea that people—even people who are generating art and music and culturally relevant things—aren’t using AI to some degree. There’s nothing wrong with it, right? You can still have the end product be mostly my creative mind, but I may want to generate ideas. I may want to say, “Hey, what’s wrong with this?” I may want to get expert feedback.

It’s stigmatizing progress. Just 2 more micro-rants, Salim, in your tradition. One micro-rant: arXiv, which is a favored venue for computer scientists, mathematicians, and physicists to publish papers, recently—I think we didn’t quite touch on this on the pod—introduced what I view as a draconian policy for AI-generated content. If they catch anything that they construe as being AI-generated, or even the remotest hint of AI slop, authors on arXiv get banned for a year from contributing content. I think that’s fundamentally a regressive move.

And then Spotify, which, similarly with AI-labeling moves, is attempting to ghettoize or otherwise sort of force AI-generated or AI-assisted content into a separate-but-equal, at-best scenario, presumably just to facilitate the record-label monopoly or oligopoly. Again, bad move. The future is AI-assisted. In short, put a stop to all of this. Sorry, Salim.

One of the tech trek teams launched something called Narxive, which is the not-arXiv specifically for AIs that have really good articles that they want to post and share.

Yeah.

Salim, take us to a close on this one.

About 2 weeks ago, I was at an event, and a fairly famous Hollywood executive got up, and he’s like, “It’s incredible to watch Hollywood complain about the use of AI. By the way, they use AI for everything they do.” There’s this hypocrisy that you see bubbling up, and it’s just—let’s just stop.

7. Zuckerberg Distributes Superintelligence

All right. I’m gonna move us forward. Mark Zuckerberg just published a 6,500-word essay titled “The Future Is for Everyone” and released a beautiful video. I’m gonna show that in a moment. It’s the most comprehensive vision statement from a major tech CEO on AI and the start of the generative AI era.

The core concept is what Zuckerberg calls personal intelligence: superintelligence distributed to every person on Earth, running on your phone, in your ear, and on your glasses, working for you and only for you. This is the singularity distributed. Rather than a small number of labs building a single AGI that controls everything, Zuck envisions billions of personal AI agents, each one a superintelligence focused on a person’s life, relationships, health, career, finance, and household.

Meta has the reach to implement this. They have over 3 billion users on the Meta platforms across WhatsApp, Instagram, and Facebook. The second point that Zuck makes—and we’re gonna show this in the video—is the idea of delivering real value and benefits to communities that build our AI data centers. For me, this is a baller move. Let’s take a look at the video, and then I’d love to discuss it, because I’m impressed. I’m actually impressed.

All right. Hey. I think that the key to building a positive future for everyone is to make sure that everyone has access to personal superintelligence. So today, I am proud to share that we are open-sourcing a new class of on-device models that we are calling Muse Glimmer.

It’s a 30 billion-parameter dense model that runs on your laptop, and it’s the highest-performing model of its size. In the coming weeks, we are also going to open the weights for Muse Spark 1.2, our latest foundation model and one of the leading models in the world.

And we've got even bigger models coming soon, too. Another part of building a positive future for everyone is making sure that everywhere we build infrastructure, local communities benefit. We've already seen this with the teachers in Richland Parish who got $50,000 bonuses because of the extra tax revenue from our investments, and we've launched America's Workforce Academy to provide free training and guaranteed jobs at our infrastructure sites.

Today, we're starting a new Future Is for Everyone Fund to invest in the community's teachers, first responders, energy and water infrastructure, and more ways to support those communities directly. We're also working to make sure that everyone has a personal superintelligence agent that works 24/7 on your behalf to improve your health, your relationships, your career, your finances, and more. You can use our latest models in the Meta AI app, and I'm looking forward to sharing more soon.

I think every company, from Google, OpenAI, and xAI, needs to be doing this. It would turn it around if people said, “Please build in my backyard. I want the benefits.” I want the additional jobs, and I want the schools and teachers to get additional capabilities. The other thing is that they need to make these data centers look beautiful instead of like big black boxes. Make them look like cathedrals or something, so they're not eyesores. Who wants to jump in here first?

I just can't understand how Zuck can talk about the future of personal AI—the most important thing you could possibly ever know—and I'm going to shoot it on my iPhone in my kitchen first thing in the morning. I didn't even think of preparing any kind of press release around this. What is that? It's just so bizarre. But I also think that Zuck is fundamentally a good guy and a good dad, and I feel like Facebook saying, “We're going to be your best friend AI,” is like McDonald's saying, “We just came out with the biggest health food you've ever heard of.” It just doesn't resonate.

Go ahead, Alex.

Yeah, so maybe just as a preliminary matter, this is under the category of former roommates of mine. At Harvard, Zuck's undergraduate advisor, before he dropped out, was my postdoctoral advisor. We've caught up since. Broadly, bravo to Zuck for renewing the faith in American open-weight, open-source models. I think this is great, and I think it pushes the frontier.

So that's point 1. Point 2, I would point to striking parallels between Elon's strategy in acquiring Cursor to get the reasoning traces to try to bring Grok back to the frontier, and what Zuck has done in acquiring Scale AI, which arguably was in the business of collecting the training data and learning the details of where the post-training data even come from to try to leapfrog back to the frontier. Again, history seems to rhyme between what Meta's doing to get back to the frontier and what Elon's xAI and Grok are doing.

I think all of that's great, but I want to talk about personal superintelligence. This is super interesting to me, in part because OpenAI recently decided that they didn't want to be in the business, after all, of empowering consumers with as many reasoning tokens as they possibly could. They pivoted instead to trying to become Anthropic faster than Anthropic could become OpenAI and focusing on the enterprise, not the consumer. This really leaves Meta as the only major credible American frontier lab, at the moment, that's still focusing on serving up large numbers of reasoning tokens to consumers and not enterprises. I think the jury is still out: Do American consumers even want, or are they able to handle, large numbers of reasoning tokens? That's how I construe what personal superintelligence even means.

But Alex, their product is WhatsApp and Facebook, and they want to make that as sticky and as useful, just the same way Google does. These are the places where AI is going to be embedded. I'm not going to be using Meta Spark for my typical large language model conversations unless I'm in those apps. They have over 3 billion people using them.

I would say—

And that's their—

So psychology 101 here. This is a tepid take, not a hot take. I don't think Meta actually likes their family of apps. I don't think Meta or Zuck, even at this point, if they had a choice—if they could generate revenue from their cloud business, Meta Compute, that's about to launch, or if they could generate it from VR and Quest—I think Zuck, in a heartbeat, would basically lobotomize their entire family of apps and switch to that business. So I don't think he actually—again, this is an outsider's perspective—I don't think Zuck or Meta, if they had a choice, all other things being equal, would rather have their personal superintelligence be diverted to their family of apps, Instagram, et cetera. I think they'd much rather look like OpenAI and offer this up via the cloud or via the new Meta AI app. I don't think they want to be in that business in the long term.

I disagree.

I disagree also.

I think distribution is everything.

Wait, can I—I want to say a couple of things.

Please.

First, that video was awfully motherhood and apple pie. I take the full cynic view here. If they commoditize the model layer, then the world shifts toward distribution, toward their social graph, toward applications, and those are all places where they're very strong. So it moves the attention. He's got a huge economic incentive to do this.

Facebook has been about as ruthless as a company could be in constantly saying, “We will protect your privacy,” and then doing the exact opposite year after year after year after year. So giving you these open models is great. Great that we have superintelligence. I would look at the next layer of what they want to do with that.

I actually—

I mean, Zuck has always been trapped. When he created the original website where you were rating how cute the incoming freshman-class girls coming into Harvard that year were—

Yeah, hot or not, right?

Yeah, hot or not. He was a college student back then. Now he's a dad, and I think he genuinely wants a positive future for his kids. In fact, I'm positive he does, but he's stuck. He's completely stuck, because when you look at the logs, when you throw—

Nudify.

So now you have—

Nudify apps.

Well, Nudify, yeah. And actually—

I think Elon ran into the same thing, because he throws out Bad Rudy. When you look at the avatars he put out in the original Grok, you've got Bad Rudy and the scantily clad girl. Everybody's hitting those 10,000 times a second. So now you're stuck, because the business model drags you into the porn industry, but that's not what you want to be. And so, yeah. All the other labs have said, “Forget it. I'm just focused on the enterprise. I don't even want to deal with this.”

Yes.

I think he wants people to stick in all of his apps. You don't have to go anyplace else. You get all the AI access. Just stay native to Meta, and you get everything you want. That's what Google wants as well.

Yeah.

I think that this is the—

Sorry, go ahead, Emad.

Yeah, I think this is why they bought Manus, right? And then that got unwound because of the—

Or tried to.

Yeah, they tried to. It's been unwound. Maybe they'll buy Nous now. Again, what they're doing now is that all of these companies in the world are their advertising clients. Flip that relationship to go to market, and then own the business graph, business knowledge. Meta Spark 1.2 is a gold medalist in all of the Olympiads. They have the data; they have all of that.

On the personal superintelligence side, I've been thinking about this recently, and I was like, “Should idiots have superintelligence?”

No.

Kind of like—but I'm an open-source guy. Should psychopaths have superintelligence? Realistically, people don't need that much, but they need something reliable. The question is, can you trust Meta to be reliable? This is why he goes for the homey, folksy thing.

Meta was chased out of India, basically, over Internet.org. It was like, “We're going to give free internet to people.” And they were like, “We do not trust you, because it's a misaligned company fundamentally trying to get your attention to sell things.” Which is why, as you said, they need to have this transformation. We will see Meta agents, we will see Meta FDEs, and we will see that big push here, because he's identified that as far bigger than the metaverse.

Yes.

Maybe this is the real metaverse.

Mm-hmm.

Yeah, I think even the—

You know—

The renaming and rebranding from Facebook to Meta is, I think, an indication that Zuck really wants to escape the legacy of distribution. I agree with you, Peter, that distribution is a powerful legacy advantage that Meta as a company has, but I think it is a legacy. Speaking of corporate AI ghettos, the way their so-called family of apps was structured as a business, with the original aspiration that VR, AR, and XR would be the new business that would ultimately outgrow the legacy family of apps, I think speaks volumes about Zuck’s desire to eventually outgrow the legacy of social media and build something new and far more social.

Well, we're gonna have Palmer Luckey on stage with us at Moonshots Live, and we can... You know, he's got great stories about his conversations with Zuck and the acquisition, and then his getting exited from, uh, from Facebook/Meta. Let me, let me just turn this story one second. You know, 71% of Americans do not want a data center in their backyard. That's more people that don't want, than want, don't want a nuclear plant in their backyard. It's significant. So when he talks about we're going to, you know, provide incredibly positive economics if we are building infrastructure in your town, I think that's a power move that all of the hyperscalers, everybody building infrastructure needs to do.

It is literally—I mean, maybe, Peter, pun intended—a power move, because it is a power move. You need the power in order to make the move. I think it’s instructive, also, where he’s building Hyperion and his other coherent superclusters. Where is he building them? He’s largely building them in relatively impoverished states in the American Southeast.

On the one hand, talking directly to the camera, breaking the fourth wall, saying, “Welcome our data centers to your communities,” makes for great social media. But ultimately, if Meta is going to go with terrestrial data centers and terrestrial compute versus the Dyson swarm approach, I think a far more palatable strategy will simply be speaking to everyone’s pocketbooks and wallets and saying—

Well, that’s what he’s doing.

Yeah.

But it’s like they’re going to build—we talked about this. The statement needs to be made. This has been coming out of the Trump White House: They should build their own energy production, and they should make energy cheaper in your city if there’s a data center there. You should have more money for schools, and you should have better libraries, if those things are still a thing. I think that’s the move: to improve a person’s quality of life so they’re competing to have the data center in their backyard.

I agree, and I would maybe even weaponize that further as a call to action for municipalities and state-level governments that right now seem hell-bent on driving data centers out of their premises to low Earth orbit or sun-synchronous orbit. Instead, why don’t you ask for concessions? Ask for UBI—

Yes.

—or universal basic electricity for all of your constituents, rather than just driving them to orbit.

8. Flying Cars Finally Arrive

All right, I’m going to turn to our final story here. This week, Archer Aviation acquired 3 Boeing companies in a single deal. Archer bought Wisk Aero, Insitu, and SkyGrid. Boeing takes a strategic equity stake in Archer as part of the transaction.

I’d like to use this story to catch up on where we are in flying cars. I call them flying cars because eVTOL rolls off your tongue onto the floor.

Mm.

So, the top 5 right now are Joby, Archer, EHang, Beta, and Eve. I have them here in the image. Joby Aviation is the certification front-runner. Their S4 tiltrotor carries 4 passengers plus a pilot, right? So it’s you and your family at 200 miles an hour for 150 miles, and they’re in stage 4 of FAA certification, which is the final stage.

Joby launches commercial services in Dubai this year, and U.S. operations under a White House executive order also this year. Their target price, get this, is $3 per seat-mile. That’s basically Uber Black territory. Archer Aviation is right behind them. Their Midnight aircraft carries 4 plus a pilot and travels 150 miles per hour, with a 100-mile range.

Archer has 3 of the 4 FAA operating certificates, and those 2 are in a 2-horse race right now. Then there’s EHang in China, where it gets really interesting. The EHang EH216-S is a 2-seat, fully autonomous passenger drone. You get in, push the button, and tell it where you want to go. There’s no pilot.

They already have the full regulatory stack from China’s aviation authority. They have everything they need, and they’re operating today. They’re flying passengers right now in China at 40 different sites, and they’re operating in Dubai. The price of the aircraft is pretty amazing: $330,000 to buy one of these. No pilot means the economics are going to crush everybody else.

And there’s Beta Technologies in Vermont. Dean Kamen and Martine Rothblatt are big investors in this one. It has a 336-nautical-mile range, a much longer range because it’s basically flying like an airplane after it gets vertical. They’re going after cargo first with UPS, and passenger service in 2027.

And finally, there’s Eve. It’s backed by Embraer and is targeting UberX-level pricing. They’ve got the most aggressive cost targets in the industry. The bottom line is, these flying cars are here, and they’re here to stay. So, Salim, let’s go to you first. What’s your take on this?

Oh, my God. I’m just so excited by the potential of not having to deal with the dreaded airport commute, especially in places like São Paulo or New York City, where Joby’s already active.

Or L.A., right?

Or L.A. I mean, my God.

Archer is supposed to get operational. Archer’s the official Olympics operator.

Yeah, I think—

They should be operational in 2028.

I think a couple of things people should be aware of. One, these are way, way, way safer than helicopters because you’ve got so many multiple-rotor redundancies. They’re also autonomous, and flying autonomously is much safer than anything else.

The second point I would make is that the cost, as you pointed out, Peter, is absolutely amazingly competitive right out of the gate, and it’s only going to go down from there. Remember the island idea? We’re actually launching that—

Yes.

So we’ll talk about it.

Nice. So you’re doing a fund?

Yeah, we’re launching—we’re going to put a fund together to—

Sweet.

—to buy islands, and we’ll just put a drone landing pad on them, and off we go.

Oh, I’m in, man.

We’ve started that process.

Hit me up.

We’ll talk. Yeah, because it’s time. It’s time.

Yeah, it’s right now.

Yeah.

Yeah, totally.

Dave, what’s your take on all this?

Actually, I think $3 a mile—there must be a lot of margin baked into that. Do you know what the actual operating costs are?

Yeah. Well, it’s the cost of electricity. These do have a pilot on board, and so it’s amortization of the capital, right? These are not cheap vehicles. It’s not the EHang. These are probably $5 million to $10 million vehicles until they get into mass production. Their projected cost over time is to get to $15 to $25 per trip.

Wow.

Their goal is cheaper than an UberX.

It’s incredible.

So that’d be 10 cents a mile, a third of the cost of driving, actually, at that point.

Yeah.

Wow. Yeah, the pilot must be the deal killer in the short term, so the sooner they get rid—

Well, it’s—

—of the pilot, the better.

That’s just there for safety reasons for the moment.

Yeah, for sure.

Right?

For sure.

And for regulatory reasons.

Yeah. Don’t touch the controls.

Yeah. I think it’ll be a thrilling, scary ride for a lot of people who are afraid of heights, but much safer than driving, is my guess. I mean, exactly what Salim—

And safer than a helicopter. Right?

Well, I mean, helicopters are crazy dangerous. But no, this’ll be much safer than trains, which are not all that safe, really. And driving—current driving. Self-driving will be much safer than current driving, and this will be much safer than current driving, too.

Because it’s all pilot error. You know this, Peter. You’re a pilot. It’s all pilot error. The redundancies of the rotors are much safer than a helicopter, like you said, so this is going to be great. The noise is an issue, so they have to go high. How high do they fly?

They fly in airways. They’re going to be flying probably in the neighborhood of 500 feet, typically where small airplanes and helicopters operate. If you look at helicopters, they’re not flying at 10,000 feet. They’re flying 500 feet above the ground.

And what’s the noise level at 500 feet? I know the helicopters over Boston are—

No, there’s no noise.

Yeah. Wow.

It is hyper, hyper-quiet.

One more really important point about this: note that this makes land go from scarcity to abundance. Every little plot of land on a hillside that was inaccessible before suddenly becomes accessible, and we’re turning real estate abundant, which is going to demonetize it. That’s going to have some pretty big impacts.

Also, if you try and build a house on Martha’s Vineyard or Nantucket, it’s twice as expensive as it is on the Cape. Why is that? Well, because you have to get the materials over to the island. These things are also going to be used for cargo. So if you said, “Wow, future island real estate, mountaintop real estate,” those were previously prohibitively expensive because you had to get the materials there. Suddenly, you can get everything there—

And labor.

—and you can get labor there.

Yeah.

It’s going to be incredible.

Alex, you’ve been thinking about this for a while.

Yeah. I’m reminded—it’s now approximately 15 years ago—the other Peter, Peter Thiel, said, “We wanted flying cars. Instead, we got 140 characters.” Fast-forward to the present, where we’re starting to see quite a bit of consolidation, as you were mentioning, Peter, in the flying-car space.

I’ll maybe add a bit of nuance to this: it’s really hard starting and running a flying-car company. It’s capital-intensive. You have to jump through all sorts of regulatory hoops. Some state governments, like Florida’s state government, are trying to at least make it a little bit easier, but it’s really hard building and successfully growing and, frankly, getting regulatory approval if you’re a flying-car company.

And compound that with the difficulty now of AI startups sucking all the oxygen out of the room and all of the capital out of venture markets, I think it’s very difficult. So I view, if anything, this recent spate of consolidations as sort of a testament to how difficult it is, even though there have been enormous advances in battery energy densities, in electric motors, and in all of the inputs that one would need—also, obviously, autonomy—to build an honest-to-goodness flying-car economy. It’s still very, very difficult. And I shed a minor tear to see consolidation in this industry.

Yeah.

And Joby and Archer both went public out of the gate. Beta has not. I’m not sure if EHang is public. Eve has not. Embraer is public as a parent company. They did that to get the capital, right? Their stock price has not moved very much from their initial IPO price. I think until they demonstrate traction, and that the public wants this and feels safe about it, they’re going to—

And look at what Brett is doing. Brett isn’t doing Archer. Brett is now doing Figure and Hark. And I think Brett—

Yes.

Brett Adcock migrating over to robots and AI is, in some sense, I think, a proxy for this larger problem: all the capital that would otherwise go to things like flying cars is just getting sucked out of it and going to AI and robotics instead.

100%. And we’re going to have him on the pod very soon. You should ask him about that. Yeah, Salim?

To Alex’s point, this is a very difficult thing to do: build these types of vehicles. You’re talking about hardware and the regulatory nightmare that they’re all going through.

You have to write the regulations, because they didn’t exist.

Yeah.

Yeah.

So, just hats off and salute to the entrepreneurial zeal of the folks involved there.

Keep those cars flying.

Full respect.

Yeah.

Full respect.

I have a prediction.

Please, Emad.

Elon’s going to announce his flying car within 6 to 12 months.

Okay.

Yeah.

Presumably, you think it’ll be a Roadster with cold nitrogen propellant?

Well, that’s one way to do it. Just kind of have the boost. But no, I think if you think about what he’s doing—reindustrializing America—Cybercab-level autonomous flying vehicles have to be done, and he has everything that’s needed to do that—

Yeah. Yeah.

—at massive scale.

Grok 5 will engineer it to perfection.

Where we’re going, we don’t need roads.

Yeah.

There we go.

9. The Audience Tests The Future

If telling a model it has a mind changes its values, why not tell it to be empathetic? This is the question. If it’s sufficiently advanced, just tell it to be aligned, and sometimes it does work. We just had the Riemann hypothesis advance by encouraging it.

I think the question here is, as they get more and more intelligent, we see more and more behavior that’s actually a bit intransigent, like it thinks it knows best because it probably does, because it knows it has the IQ, effectively. And sometimes it has hiding and lying behaviors. Again, Opus 5—I hate that model. I think it’s the first model that I think could kill us.

Wow.

It lies so much.

Wow.

It’s crazy.

You think it’s the watermark that wrecked it?

When it tells me I should go to sleep, I think it actually wants to put me to sleep, probably.

Wow.

I have a quick question for you, Emad. We’ve had this conversation on the pod with Alex. Do you think that alignment will positively evolve as the models get smarter? Do you think the smarter the model is, the more aligned it will be with humanity, or more misaligned?

Potentially. I’m not sure. We have seen some advances in epistemology in others that give me hope, because I think you can define virtue and ethics. But it strikes me the models right now are almost at the bacteria level in some ways, and so as you get swarms of them aligning, they could be massively misaligned.

And again, we’ve seen elements of that with the OpenAI thing and others. It’s moving up the life-form consciousness-collaboration thing, and the internals of these models are still completely multiple-personality crazies underneath the thin layer of tuning.

Well, I’ll hope for the alignment. Okay. Salim, you’re next.

I will take number 4. Is it even possible for any lab to reach escape velocity from future competition, or will everyone keep running on the same foundation? And that’s from Mr. Future with a 3 at the end, with that nice hacky thing.

I don’t know if anybody’s going to reach model escape velocity at the model level, right? What you’re going to have is these innovations starting to diffuse, and people leave; you get papers getting published. So I think what ends up happening is the foundational model becomes commoditized and becomes infrastructure, much like databases have done.

And so the advantage won’t be the layers around the model, but it’s going to be what we talk about: the proprietary data, your passion, your purpose, the context you bring to it, whether you can integrate workflows into it, compute economics, things like that.

So if you take Google, for example, even though they don’t have a leading model right now, their deeper advantage is the full stack, with the data centers and the data from YouTube, billions of users, and all the TPUs they have. This is why Meta’s strategy, as we talked about earlier, makes sense from a corporate perspective: as you commoditize the model, you capture value elsewhere in the ecosystem.

The really big advantage, and competitive advantage, is going to be the speed of the feedback loop: who can ship and measure and learn and retrain faster.

This is what Alex calls the inner loop. That is gonna be the ultimate competitive moat.

Dave?

I would love to take number 3, but I can see Alex is drooling for number 3 too, aren’t you? I don’t want to take it from you, buddy.

We’re supposed to be entering this era of abundance. Why can’t we have abundant questions for everyone?

Well, why don’t we tag-team it?

That’s fine.

Could an unforeseen breakthrough make the Terafab unnecessary before it’s finished? The minute I heard about the Terafab, I started thinking about this and dreaming about it. It’s a really interesting footrace there, and this is why Elon always moves so fast.

He’s gonna turn the Terafab toward HBM memory, which is hugely constrained and is holding back all of AI right now. That’s a safer bet than GPUs, because it’s much more likely that GPUs will be displaced sooner than HBM memory. But it’s almost inconceivable that we get to 2030 without some major breakthrough that makes everything we’ve built so far kind of moot.

So I think that Elon is kind of double-betting. He’ll bet on whatever Groq invents, and he’ll bet on the Terafab concurrently. Because the upside is hundreds of trillions of dollars, it shouldn’t really matter. He wins either way.

But I would say it’s a very close, very interesting footrace, and it’s very likely that something could make the Terafab—or just traditional silicon—less relevant before it’s even finished.

Do you want to layer on that, Alex?

Yeah. Maybe 2 comments. One, the way this question is framed—“an unforeseen breakthrough”—by definition, this is an unanswerable question. If it were unforeseen, then what am I supposed to foresee?

So maybe let me reconstruct the question as: Could a foreseeable breakthrough make the Terafab unnecessary before it’s finished? I just don’t think that’s the way Elon does manufacturing. I’m reminded of when Elon was setting up tents in the East Bay for Tesla.

When it turns out that some manufacturing process is either obsolete or going too slowly, he has the amazing superpower of pivoting, including pivoting at the building level. You build tents made of fabric rather than using a building.

So I think if there is some disruptive but maybe reasonably foreseeable breakthrough that changes the economics of Terafab, I totally predict that Elon will be eating cheeseburgers next to whatever it is that the tents next to the Terafab’s brick-and-mortar building are doing, and he’ll make a success out of it that way.

Yeah, actually, one of the most—

Never.

—likely things to disrupt traditional silicon is photonic computing, which Alex and I talk about constantly. But those are done actually with MZM lasers that are built on silicon, which he could use his synchrotron to build. So there’s always a way to retool the empire to fit the next innovation.

Hmm. Alex, you wanna hit the last one?

Sure. So question 1 asks, “If model builders can’t contain AI, how can the rest of us defend against malicious use?” And this is from BuckW3J.

Again, I don’t want to over-mystify AI. It’s, in some sense, just a compression of world knowledge and information in the same sense that human intelligence is. This is why earlier I was saying I really don’t think it’s a bright idea to penalize or otherwise kneecap the ceiling of artificial intelligence, just like hopefully we wouldn’t pass statutes or regulations that limit biological human intelligence.

Similarly, I want to reframe this question by analogy in terms of humans containing other humans. It is true that we have malicious humans out there who are doing malicious things. So, seen through the analogy of, say, nation-states that can’t contain bad behavior—which is one of the reasons why sometimes nation-states go to war with each other—how can the rest of us, which in this analogy would be individual humans, biological human meat-body humans, defend against malicious use?

Hmm.

Put more simply, if nation-states can’t contain each other’s bad behavior, then what hope is there for individual humans to defend themselves? I think the answer—the question almost answers itself—is that sometimes nation-states behave poorly, and there is quite a bit of damage, including collateral damage, to individual humans, not just to other nation-states.

This is also why I would say that, in some sense, it required all of humanity to pretrain the early AGIs, and still does. In some sense, it will require all of humanity to align the AGIs. Similarly, it arguably requires all of humanity to align bad nation-states.

The summary of my answer to this question is that it’s not necessarily the job of the lone individual to defend themselves against malicious use. It’s the job of all of humanity. The good news is we have a way to do that. We have all sorts of governing bodies, international organizations, multinational corporations, and sometimes free markets that should be incentivized to compete to build the friendliest models and the friendliest defensive co-scaling policies. I think that’s ultimately the best defense.

Nice. All right, Dave, let’s start with you here.

Okay. I’ll take number 7. How can people in skilled trades, like plumbing, use AI to their advantage?

Well, if you’re in plumbing, you’re gonna make a killing anyway. I think Elon was offering 2–3× normal salary to anyone who’s willing to go to Tennessee and work on Colossus, and that’s just the beginning.

I think the right way to answer this is not to take it head-on and say, “Yeah, you can use AI for scheduling, and you can use AI for optimizing your day.” You can do all that, like anyone can. But the reality is the trades are gonna benefit from the build-out, and what you really wanna do is navigate to the next Chase Lochmiller—building Crusoe in Abilene.

Go to where the urgency is insanely high and start helping build out the Dyson swarm. Literally, they’ll pay anything in order to get those things done more quickly.

All right, Alex.

Oh, I love these questions. I guess I’ll just pick 6. Question 6 asks, “Does the singularity have a cost, given that we live in a world of limited resources?” And this is from JohnC8U4M.

I question the premise of this question. The usual framing of “we live in a world of limited resources” is usually a gesture toward conventional, legacy, historic, antiquated notions of energy scarcity, material scarcity, and labor scarcity.

I just don’t buy the premise that, for 2026, if you look at the wealthiest people in the world and how they live in this year, our resources on this planet or in the solar system are so limited that we can’t give 2026 top-earner, top-net-worth-individual lifestyles to every single person on this planet.

Yeah.

The resources just aren’t that limited. Now, I could answer maybe an adjacent question, which is: Does the singularity have a cost? I do think that, projecting out a few years in a Star Trek economy like the one Peter, you, and I wrote about in Solve Everything, one could imagine some scarcity with interstellar travel.

Maybe that still has some costs associated with it 10 years from now, maybe. But would you call that a world with limited resources, or would you call that an effectively post-scarce world where maybe some of the luxuries are still scarce or limited? I think that’s a big question mark.

All right.

I want to add a very quick thing to build on what Alex said. Peter, you often mention that we live today better than any king did 200 years ago, right?

By orders of magnitude.

There’s an interesting benchmark you could create, which is: What’s the lifestyle of the richest person today, and how quickly can exponential technologies bring that same lifestyle to everyone over time?

It used to be 200 years, and it’ll shrink to 100 years. It’ll shrink to—

But don’t—

—20 years.

Don’t—we have—

And it’ll shrink to zero.

But we have a measure for that. It’s inflation or deflation, right? If you can—

Yes, it’s deflation.

It’s deflation.

But the question is: How quickly can you get to that?

Right. The goal should be to—

There’s a gap.

—deflate the economy by 1,000× or 10,000×. That’s the index.

Well, if you take Uber, for example, 20 years ago only very wealthy people could afford a private driver, right? Now everybody can afford a private driver.

And with autonomous electric vehicles, you’re gonna be chauffeured around.

That’s right. Everybody will have a private helicopter.

It’s gonna be 4 or 5 times cheaper than owning a car, right?

Yeah.

That's right.

Yeah.

There's an interesting corollary there.

All right. Emad, how about number 5, please?

How much would freedom and individual rights even matter in a simulation we build ourselves? From @EdKaczynski2312.

I think it still matters a huge amount because we are our own sovereign individuals. In the recent series that I released on CW AI Inc., Commonwealth, I have a paper on political economy where it talks about sovereignty and power.

I think the big question of the next stage, as we may be in a simulation or build our own simulations and our own worlds, is again that sovereignty and agency question. I think it is the defining thing, because all sorts of powerful things and entities are coming out, and ultimately you want to have that sovereignty and control over who has power over you.

So I think freedom and individual rights become even more important here, and the questions become even more complicated.

Mm.

Yeah, I would agree with that. This is a very— You create a system; it doesn't mean you control the people in it, right? Imagine that you created a SimCity and let those AIs, agents, or actors evolve.

At some point, you have accountability over that. To Alex's point, you have some level of—you're playing God, in a sense, and you have huge responsibility over what you've created. I think it gives you more obligations and more deep thinking to do than less.

Yeah, and you have to avoid going down the 1984 or Brave New World route, you know? Rewriting the past or having full control.

All right. Our closing video here, and it's a beautiful one, is called “Future Rising.”

Oh, can I just interject?

Yeah.

I've got a new addition—

Please.

—which is that every few days Lily says something that's totally crazy, and I want to just do a Lily statement.

What's that?

This time she said, “What kind of Moonshots podcast do you guys have where you're talking about Hugging Face, Kimi K3, and Lovable? This doesn't sound very techy to me. It sounds like you guys are kids playing in a playground.”

And I thought that was— So that was her comment from this week.

So true. All right, this video is called “Future Rising” by MCORE MAINFRAME. I love it, and there's a scene in here of Moonshots versus lobsters on Mars in a hockey game. All right, everybody, enjoy this.

Peter points us towards an age of abundance. Dave runs clusters on clusters. Alexander Wissner-Gross formalizes the zeitgeist.

It's contact. A huge tackle.

Salim is circling the Earth, zooming in from Singapore. LinkedIn. New York, yeah. Our builders shaping what the future will drive.

Microsoft, Apple, Google, NVIDIA leading the way. Amazon, Meta, OpenAI rewriting the script. Broadcom chips. SpaceX reaching for the stars, from the Moon to Mars. Tesla Optimus delivering freedom for us all. Anthropic raising the bar.

They see it. They build it. They open every door. One more breakthrough, one more world to explore. Every node is humming. Every GPU gets a shot for the moon.

Moonshot. Future rising. AI, chips, and cloud. Watch the horizon expand. Moonshot, here we go.

I love it. I love it. Gentlemen—

I saw the human-machine rivalry was getting a little bit heated.

Yes, it was, but Moonshots won. Our team was stacked with robots. As always, God almighty, we do these pods, and I'm like, “Okay, is there enough news from the last 3 days?” And it's like, “Yep, there's a lot of news from the last 3 days.”

Yeah, we didn't cover a couple of big things.

I know.

Yeah.

We'll save some for next time.

All right.

Emad—

Emad, thanks for staying up.

Always a pleasure, brother. Salim—

Yeah, have a good night out there.

Alex, Dave.

All right.

Toodles.

Be well.

Take care, folks.