Anthropic 的 2 万亿美元现金机器、CoreWeave 暴跌 30%，以及中国的 Kimi 迎头赶上 | #306

Guys, I’m curious. Have you been tracking what’s going on with the protests and revolts in France, Italy, and Spain? It’s kind of insane.

It is, but I’ve talked to people on the ground in France, and they’re like, “The media is way, way overblowing it. There are a few bad people, and everybody else is just moving around, drinking coffee while things are hopping.” So it’s very split. I don’t know which way to go. You could go either way on it.

Yeah. I mean, on October 17, there were half a million people marching across France. They had 2,400 arrests, looting in Paris, Marseille, and Lyon, and they burned down that beautiful 19th-century hall in Limoges. It went up in flames.

I hate this stuff. There are a bunch of paid crusaders. They suspect Russia, naturally, is involved in stirring the pot really badly, and that’s antagonizing everything.

Yeah, but underneath it all, there’s a 20% unemployment rate in France, Italy, and Spain. That’s got to hurt.

Yeah.

It’s got to hurt.

Thank goodness we never have protests in this country. I always said—we always say the French are revolting.

I wrote a Substack a few months ago—I think we talked about it on the pod—that the most dangerous demographic in history is about to get a lot bigger. What bugs me is that, across the last 200 or 300 years, from the French Revolution to the Bolsheviks, to Iran in 1979, to the Arab Spring, they all share one ingredient: young men, testosterone-laden, 18 to 28, educated enough to know that they’re missing opportunities and are blocked along the way. Dave, when we talked to Elon about UBI and civil unrest, you said we’re going to have both. This is probably my single greatest concern.

It’s a good place to keep an eye on, too, because there’s no data center activity, no power, and an incredibly educated population that should be capable of amazing things. But without the government getting ahead on compute, data centers, or power, you’re absolutely screwed in the age of AI. When you look at the way U.S. voters act, they vote for UBI within their borders, but they’ll never vote to bail out France or Spain. France actually has nuclear power, so it might be okay, but Spain and Italy have no sentiment in the U.S. or China to help. If you don’t take control of your own country and get your government to get its act together, the 20% unemployment rate is only going to go up.

Yeah. What’s it feel like over in England?

It’s a bit depressing here. They’ve started sending letters to all the rich people and entrepreneurs—where “rich” means above £200,000 in earnings—saying, “I’m your new compliance audit advisor.” The government set up a compliance audit liaison team to make sure you’re compliant with all your taxes and pay a bit extra. The entrepreneurial class isn’t feeling very happy about all these policies coming in. A lot of people are about to leave. We’ve got the budget coming up, and looking over to France, it’s not surprising. In France, it’s basically impossible to fire anyone.

This is the other thing.

They spend 57% of their GDP on social programs. It’s more than any communist or socialist country in the world.

Yeah. I think that’s one of the side effects.

People are fleeing to Toronto and Waterloo like crazy. I’m sure you can see it. When you walk around Waterloo, everybody is at the top of their nation, from all over the world. You’re used to seeing people coming from Asia and India, but now floods of people are coming from Europe as well. If you can’t get into Silicon Valley, Boston, or New York, you can often get into Canada—anything to get out of the slow zone in Europe and into the AI zone, which Waterloo has become.

Emad, when are you joining us? When are you coming to this side of the world?

Let me finish building a few things first. I can be the voice from across the pond until then.

Actually, something very interesting in AI: Poolside, whose founder was the GitHub CTO, I believe at Amazon, went to France.

And they raised $200 million, aided by the French government. They went straight back to America to release Laguna, then sold to Nvidia for $6 billion. So now I think all eyes in the AI sphere are on Yann LeCun’s AMI Labs, which was a French company but may become an American company sooner rather than later.

I might be on stage with him in a week, so I’ll ask him directly.

Yeah. Are you French or American?

No. Well, that’s a whole other question.

Yeah, but we need to keep an eye on what’s going on there. I just think these are the early seeds we need to watch.

It’s easy to mobilize a lot of people around what you’re against. It’s much harder to coordinate that around what you’re for. What’s the alternative?

Yeah, yeah.

I’d be interested in this: as you talk to all these corporate leaders around the world, are they expanding their graduate programs? I very much doubt it. That’d be the first thing to go, almost, right?

Not necessarily. We’re seeing some weird things. For example, a bunch of people are increasing graduate hiring because young people armed with AI are much more valuable than 50-year-olds moving spreadsheets around. Some companies are actually increasing graduate programs and replacing older workers aggressively because young people are so much more AI-native.

But generally, that’s a problem.

That feeling that we’ve missed the boat, that we can’t catch up. When you look at the charts later in this podcast about how open-source models are getting closer and closer to the frontier, the path to sovereign AI has never been clearer. Countries actually have a very clear path forward if they act. I don’t have any confidence that they will, because they haven’t acted so far. But if they jumped on it and said, “Look, we need an emergency national measure here for power, chips, and sovereign AI,” you could succeed with probably a 90% to 99% probability if you could get people rallied around it. I just don’t see any groundswell to rally around it. We just need one or two—maybe a South American country—to prove to the world that it’s possible to get on the map. That’ll crack it open for everybody else. The window is right now, and it’s relatively narrow because AI is accelerating at this ridiculous RSI pace. But there is an opening right now. You’ll see it later in the pod.

Yeah. I think we’ll look back at this and say the singularity was the best of times and the worst of times. It was the time when a country that arguably invented, or helped invent, nuclear energy—and should be absolutely drowning in energy too cheap to meter—had students protesting that they weren’t getting proper heating or air conditioning in their schools. There are many profound ironies here. Europe more broadly needs, if anything, to catch up to the singularity. Maybe we’ll see more activity like this, whether it’s ostensibly through educational grievances or budgetary grievances, or just missing air conditioning.

Employment grievances, social grievances, public service cuts—all of these are intimately tied together. One might say, sure, there’s a bit of a protest culture in France. They’ve had a number of revolutions, maybe culturally and to some extent in their DNA. But Europe is also very well positioned to become a living history museum if it doesn’t catch up in energy and compute.

Well, I’ll tell you, the U.S. is in great shape in every way except power, which we’ve talked about several times. Yesterday I had a meeting with the guy who ran data centers for Google for about seven years. It was a fascinating conversation. He’s got a U.S. data center fund now, and the gap in power between the U.S. and China is going to expand almost 10 to 1. It’s crazy how far behind the U.S. has gotten in power. He was saying, “Look, there’s a lot of push from the White House to catch up, but after this next election cycle, things will probably get frozen again.” Eric Schmidt has pointed out many times that the U.S. and Canada need to work more closely together, because Canada has a lot of power and not a lot of chips. There’s a very natural partnership there, and there’s been way too much friction in the last couple of years. But even with the U.S. and Canada working closely together, China is miles ahead in power generation.

Until SpaceX and Tesla bring on 100 gigawatts of solar each.

1. Warhound: Genetically Engineered Working Dogs

Yeah, true. But the gap in time there is really scary. Space-based data centers are going to be enormous. The Dyson swarm will be everything in the 2030s, but RSI is the next two years. A good strategy around AI really focuses on the exact moves for the next two years, because after that we’re talking about intelligence like no one has ever even imagined. It’s really hard to predict the effect of new physics, new chemistry, and all of Alex’s breakthroughs.

I mean, those are going to change.

We’re seeing new physics now.

Yeah, it is. It’s almost impossible to look more than 2 years ahead. It’s difficult to look 1 year ahead, for God’s sake.

Yeah.

To everyone in France, the country that brought us the word “entrepreneur,” we wish you the best.

2. Welcome to Moonshots

Now, that’s a moonshot. Ladies and gentlemen, this episode is brought to you by the Abundance Summit and Link Ventures. Welcome to Moonshots, everyone, your number one podcast in all things AI and exponential, your front-row seat to the hypersonic tsunami that we call the singularity. Today, the fabulous five are here with us. Alex Wissner-Gross, our resident ASI, who spent his weekend casually rederiving the fine-structure constant from a single sentence and is delivering a new theory of physics. More on this later in the episode. And Dave Blundin, my dear fraternity brother and buddy, AI VC extraordinaire, the man who’s backing unicorn founders out of MIT and Harvard. Salim Ismail, author of Exponential Organizations and The Organizational Singularity, frequent-flyer legend and our warlord against linear thinking.

So, you’re clearly not at home, Salim. Where are you today?

I’m in Toronto because I’m giving a talk here to about 400 CEOs tomorrow, and then flying back Thursday.

Okay. You’re sure you’re not at a conference? At a concert?

I’m not at a concert. I’ll report if I am.

Okay. I just want to start out by saying something to all you guys: I feel so blessed to have you as friends and mates here. You guys are extraordinary, and I love the way you all think. Honestly, even though these are intense episodes, this is my favorite time of the week—spending these couple of hours with you.

Well, and also the team behind the scenes, whittling down every fact in the global economy to the subset that you absolutely need to know. It’s gotten better and better and better. I feel really lucky to ride that wave.

Yeah, it’s awesome.

Singularities are spiky, and we are the spikes.

I was cursing this whole effort over the first few months, but then I realized it absolutely forces me to be totally current. I can’t skip a beat multiple hours a week. I think you and Alex spend the most, Peter, but good lord.

Yeah. Are you still suffering from future shock?

The tagline is, “How do you move from future shock to future shape?”

Beautiful.

Okay.

Another T-shirt.

Here’s the number for the week: $16.3 billion. That’s Anthropic’s revenue in the first half of 2026, over 14 times higher than the $1.16 billion from a year ago. Dave, have you ever seen anything ramp this fast?

No, of course not. But I don’t think anyone has.

Yeah.

Coming from a cold start, that’s got to shatter every record by far. Anyone who says they’ve seen this movie before has never seen this movie before. This is new in the world, as is everything going on right now. It’s all unprecedented.

Today, we’re going to cover 7 stories with 1 through line: The AI economy is splitting into winners and wreckage, and the science is getting stranger and more wonderful by the week.

I’m going to kick it off with a discussion on the recent AI sell-off and ask the question: Is this the AI bubble starting to pop? This month, the neoclouds—companies that borrowed to buy Nvidia chips and rent them out—got hammered. CoreWeave and Nebius were down 30%; IREN was down 40%. The trigger for the sell-off was CoreWeave’s $51 billion in debt and its risk of default.

On Sunday, friend of the pod Chamath posted, “There will be a lot of partially stranded compute and compute sitting in inventory in 2027.” He added, “Bought and paid for, but effectively under-earning or collecting dust. 2027 will be a turbulent year because it will clear out a bunch of bad bets.”

Elon responded. He said debt markets have almost entirely shut off access to money for any company that doesn’t have substantial revenue independent of AI data centers, and the 200 neoclouds are going to see guppy-level death and consolidation. He doesn’t pull his punches.

This reminds me of the fiber bust back in 2000. A whole bunch of companies built way too much fiber capacity for the internet boom that was coming, and the builders ended up with excess fiber capacity. It eventually got used up by YouTube and Netflix, but there were a lot of bankruptcies along the way. So, Dave, is this the bubble popping, or is it sort of a slow leak?

I love the analogy to dark fiber. I think Gavin Baker might have been the first person to bring that analogy up, from the All-In Podcast and Atreides. The internet went through this phase of incredibly rapid growth. When it was about 80% engineers and 20% business school people, it was thriving. When that inverted to 80% business school people and 20% engineers, that’s when you started overbuilding dark fiber.

Everybody races in, not knowing anything about AI, and they want to be involved in the boom and the buildout. The easiest way to do that is with massive financing deals. There’s about $100 billion in data center debt, aside from Elon’s $40 billion, which is tiny by any global standard but still pretty big within any given market.

It is very similar to the dark fiber buildout. I think all compute is sold out for years, so anyone with a viable data center is going to do incredibly well. There’s no risk of data centers collapsing unless they signed really bad debt deals. It’s very hard to get transparency into the details of each company’s deals. But Elon is borrowing at 8%, which is great.

Apparently, these other guys are all borrowing at 15%, and that can add up. I don’t think they’re in bad shape at all unless there’s innovation at the technical level that makes their investment in chips or fiber or power not viable anymore. Right now, I don’t think there’s any urgent panic.

But at the rate that physics and technology are going to get reinvented, it wouldn’t take much. Alex could probably vibe up something in a couple of weeks, at the rate he’s moving, that would potentially restructure the underbelly of this and make a lot of those $10 billion or $20 billion borrowings at 15% suddenly look really, really shaky.

It’s really hard to answer because there’s no technical answer. It’s: Did you sign a bad deal with a bad lender? It’s pure business. Alex, your thoughts?

Yeah, I think the elephant in this particular room—there is an elephant here—is dollars in versus dollars out.

What is the return on investment for compute? In an era of recursive self-improvement, when you can do more with less, I think one possible scenario—if I had to look into my nonexistent crystal ball and imagine what a dark compute credit bubble overhang burst might look like—is that recursive self-improvement is so strong and compelling that you can accomplish so much with so little compute that, even though it seems borderline unconscionable now that we would ever saturate even the dark compute, software could potentially run away from the hardware.

Right now, every exponential improvement via RSI in algorithms, models, and post-training looks like Microeconomics 101: it wants to suck up even more hardware. They’re complementary. But there are scenarios where, maybe in the not-too-distant future—maybe in the next year or 2—we finally find the perfect model. The perfect model, or the perfect approach, is so much better than what we have right now, for a variety of reasons, that it causes a demand shock.

As an industry, we’ve never really experienced a profound demand shock. There was a mini demand shock, more speculative than anything else, from an earlier DeepSeek model, but we’ve never really seen a strong, discontinuous demand shock. Something like that could happen, and to the extent that there’s arguably overleverage on the infrastructure side right now, it could lead to a major correction in the space. It’s not financial advice, but it could.

Yeah, I think Alex has some good points here, and so does Dave, of course. Let’s put this in context. You’ve got about $500 billion in direct data center debt, and about double that with all these weird SPVs and vendor financing arrangements. Subprime was $1.3 trillion, so you’re heading up toward those numbers. Even if CoreWeave by itself is small, it’s not when you consider the hyperscalers and all of these other players.

As Alex said, you’ve got better architectures and better models. A year ago, we were all freaking out about o1 and Claude. That performance now runs on your smartphone—the Gemma model. When a model saturates and gets good enough, you optimize it. The companies Dave and others are building require far less RAM than others. I wouldn’t be anywhere near RAM stocks right now—not Micron or the others.

Yeah, because again, they’ve sold out for years, and the thing is, they’ve actually been slow to build out, which is also good. So now they’re low on earnings, but the market is frothy. There’s 45% of all spending. New architectures are coming on both the software and hardware sides, because you finally have models good enough that you can optimize the hell out of them and what they run on.

You know how these models are starting to behave now, too. And this is the deal: as Elon said, if you haven’t got other revenue streams and you can’t use them yourself, Elon will use all his spare compute capacity for Macrohard. He’s sold off his chips as well. His return on chips is about 9 months, because he’s selling them at 3 times the market price. By the time the bubble shifts, he’ll have paid off his chips, and his return on investment is the marginal cost of electricity. That doesn’t hold for CoreWeave and the others, and that’s the dangerous thing for them. So I think you’re in this really interesting spot where things are starting to turn, even as, for the first time, the actual return on intelligence breaks a barrier where it’s genuinely useful—and we’re feeling that today.

What do you think?

We’ve had a lot of financial wizards call it an AI bubble and warn against prices decreasing substantially. What would cause an AI bubble here? I’ll throw out the first idea: if there’s some kind of major hack that brings down a power center or a stock market, I think the fear resulting from that could drive a pop in the bubble.

Well, last week you had massive targeted attacks on PyPI, and the literal CVE chart—the critical vulnerabilities chart—looks exponential. I don’t think this is a bubble popping because something happens with AI. I think it’s just the inertia of it and everyone saying, “Okay, where is the value now in this chain?” The Anthropic IPO is so huge for that reason. Can they maintain it? What will it perform like if the Anthropic IPO comes—

And if it performs badly, Dave, what do you think happens to the market then?

Mhm.

I don’t think a poorly performing Anthropic IPO would necessarily create a bubble. I would distinguish between what’s going on in the public equities market and what’s going on in the public fixed-income and credit markets. If I had to spitball the conditions for there being a bubble, I think the likeliest scenario is actually a deficit in demand, not unlike what OpenAI saw, where they were assuming—incorrectly, as it appears—that all of their consumers would want reasoning models and all those extra reasoning FLOPs. It turns out they didn’t, and they didn’t have a great use case to motivate the spend. If I had to bet, I would bet on a demand surprise, where all of this new compute is coming online, but the killer applications don’t arrive at exactly the right time to motivate all that spend.

Mhm.

I think you’re also seeing such unbelievable demand. I was listening to Jensen say you can’t buy an NVIDIA chip for 2 years; the order book is sold out. This was not the case in the fiber-optic and internet bubbles, right? I don’t think AI will fail. You might have a bunch of companies failing because of the deals they made, as Dave mentioned, but the technology is going to win. The capital structure may still lose in that.

Yep.

Yeah. It’s really important, when you look at financial headlines, to remember that they tend to say things like “the internet is destroyed” or “the internet implodes” in 2001, which is absolutely the opposite of the technical truth. Google was just starting to take off right when those headlines said the internet was a disaster.

Here, what you’d be asking is whether it’s possible that NVIDIA gets undercut. Keep in mind that NVIDIA is operating at 80% gross margins, is sold out for years into the future, and can’t make enough chips to keep up with demand. So how could anything be better than that? Well, is that already baked into the price or not? That’s the financial question.

A lot of things are already baked in. As Emad said, the RAM stocks are way up. Prior to the last 10 years or so, there was no such thing as a trillion-dollar company. Now every one of these chip companies is worth $1 trillion to $5 trillion. You’ve got to put in context how much they’ve gone up already. These are unprecedented, massive, astronomical numbers as a starting point.

If you start to see headlines like “data centers are overfunded” or “AI is collapsing,” keep in mind that what’s actually happening is AI is absolutely taking off and taking over everything. The headlines are about this deal or that deal being overpriced. That’s financial news. You’ve got to tune that out if you want to look at what’s actually happening in the world.

Here’s the flip side of the neocloud story. In Elon’s same post, he confirmed that SpaceX, as you guys just said, closed $40 billion of debt to buy GPUs. So how can SpaceX borrow $40 billion in the same month that lenders are slamming the door on neoclouds? Elon gave a simple answer. He said, “We’re investment grade and have Starlink and launch revenue.” And that’s basically the difference. Neoclouds’ only income is renting out chips, so if AI rental prices fall, they can’t pay the loans.

3. Elon Musk Predicts the End of Neoclouds

SpaceX has 13.6 million Starlink subscribers, paying every month, and its launch business has zero real competitors. Here’s the twist: SpaceX sells compute, too. It’s a direct competitor to the neoclouds. So when Elon predicted “guppy-level death,” he was also describing a market where the big fish get bigger.

Anthropic, for example, just agreed to buy $120.5 billion of compute from SpaceX through 2029. And as Chamath said, in a capital crunch, the winners are the players with cash flow and speed. Alex, your thoughts? Does the AI buildout end up in the hands of 3 or 4 giants, or does it get spread out?

I doubt it. I’m not even sure that, if we look to the long term—which at this point I think is 5 to 10 years—the market is going to be recognizable. I think a lot of this is Elon trying very intentionally, it appears, to suck the oxygen out of the room in public credit markets. He’s talking his own book, of course; he’s making statements about this because he wants all the loans for his own entities.

I do think credit markets have a certain depth. Certainly in the West, there’s a larger pool looming, as a number of neocloud and other CEOs have pointed out: insurance markets. There’s a certain amount of capital—a certain quantum of capital—available from credit markets to continue financing this infrastructure buildout, but there are other larger pools, including insurance, that could potentially also be applied to it.

As credit and fixed-income markets are perhaps sucked dry through this buildout, I would expect increasingly adventurous and aggressive financial engineering from neocloud CEOs to tap other, larger pools of capital in exotic ways.

Yeah, talking your own book is a really important observation here. Peter, do you remember when Michael Saylor was in hypergrowth a couple of years ago with Bitcoin? He could buy Bitcoin and then sell equity, with an arbitrage spread of about 30% between the cost of the equity and the cost of the Bitcoin. That basically manufactured $100 billion of wealth in just a couple of years.

Elon is now in that exact same situation. As Emad pointed out, he can borrow money at 8%, use it to buy compute, then sell it to Anthropic and get back above break-even in just 9 months on that turnaround. That’s a license to print infinite money.

Yeah.

Now Chamath is saying, “Look, Elon’s debt at 8% is better than buying sovereign debt.” In fact, the country of the future is essentially the AI country, and the debt market of the future is the AI debt market—much bigger than sovereign debt.

If it keeps going on its current trajectory, then in that view of the world, Elon will be bigger than a country and will be able to borrow hundreds of billions or ultimately trillions—more than 90% of the countries out there.

Okay.

And much bigger than a large country.

By the way, I reached out to Michael Saylor to have him come back and address Chamath’s point on Bitcoin.

4. Anthropic’s $2 Trillion IPO and Explosive Growth

We’re finally going to have the crypto debate episode? Is that what you’re saying?

For sure.

It’s going to take 2 or 3 episodes. It’s going to be a big one.

Yeah. All right.

Can’t wait.

Let us know in the comments if you want that Bitcoin debate. Either bring in Bill Barhydt, or Michael, or both. There’s a lot to be said.

Let’s jump into the next story, which is dissecting the Anthropic IPO. We reported that they were about to file their IPO, and on Monday they did. The ticker is ENT on the Nasdaq. The Wall Street Journal says the IPO is likely to occur the week of November 16. Anthropic is planning to raise $100 billion at a valuation somewhere between $1.5 trillion and $2 trillion, which would make it the biggest IPO in history.

Let me take a moment and dissect the IPO. I’d love your thoughts on this. First, what’s strong: They had $4.6 billion in revenue for all of 2025. That jumped to $16.3 billion in just the first half of 2026. In Q2 of this year, Anthropic made its first operating profit, of nearly $1 billion.

On the other side, we’ve talked about its capital efficiency. Anthropic now earns $28 in revenue for every dollar of compute, up from $4.0 just 18 months ago. Put in a dollar of compute and double it in revenue. Revenue grew 29 times while compute costs grew only 6 times. Pretty incredible. They have 11,900 companies as customers, each spending over $100,000 a year. They’ve really spread out their customer concentration.

Let me hit three weaknesses. The first is compute commitments. Anthropic now has more than $500 billion in commitments, much of it non-cancellable, including $120.5 billion with SpaceX. Then there’s revenue concentration: Two customers account for about 20% of its first-half revenue, and nearly half of its revenue flows through Amazon, Google, and Microsoft clouds, which are also its competitors. The final weakness—or strength, depending on your point of view—is that the 7 co-founders retain 50.1% of the voting power through one special share. So ultimately, public investors wouldn’t control the company. Dave, what’s your take on this?

Well, it’s a very financially stable company. When you poke holes in it because of compute commitments and otherwise, compared with the internet in 1999, this company is just a cash machine.

And the compute spend, too. Non-cancellable, but with all kinds of knobs and bells in there. I think it’s an incredibly stable company. That’s not to promote the IPO: A $2 trillion valuation bakes in a lot of goodness, so that doesn’t make it cheap. It just makes it the most extraordinary company in history by a wide margin.

The biggest IPO in history was Meta, back when Meta went out many years ago. Then Cerebras broke that record—our good friend Andrew Feldman—and then SpaceX. Remember, CoreWeave went out, I think, at maybe a $40 billion valuation, with a couple-billion-dollar raise. Now we’re talking about a $2 trillion valuation and a $100 billion raise. That’s raising more money on IPO day than the entire valuation of Cerebras when it went out, just a year or a year and a half ago, if that.

This is the most extraordinary time, by huge orders of magnitude bigger than anything we’ve seen before. If the stock goes up, great. If the stock goes sideways, great. If the stock goes down by half, it’s still the biggest thing ever. I just want to set that context, because nothing is going to slow down the AI juggernaut or the rate of change in the world. When you read an inflammatory headline, don’t slow down your life plan in reaction to it. You need to move as fast as you possibly can while the singularity is evolving in whatever direction you’re going.

I think Anthropic is correctly priced. It’s growing faster than anything we’ve ever seen and shattering every record. I think the supervoting structure is not that uncommon these days. It’s a little weird, and we can talk about that.

Will you buy into the IPO, Dave?

If I didn’t have access to seed-stage deals—which is now just the biggest thing I’ve ever seen—if I were just out there on the street, would I buy in? Yeah, I would definitely buy in. I would buy—

Not investment advice, but look, if you want to be exposed to AI, it’s a big part of where you want to be. Alex is always promoting being indexed and hedged across the universe. If you want to just be neutral, it’s hard to get into AI, so this gives you an opportunity to get some exposure. I wouldn’t mortgage the house or anything like that, but I’d definitely want exposure to the category. Now you have the opportunity to be exposed to the category.

5. Anthropic’s Corporate Governance and AI Safety

You can do that via SpaceX, too, but you’re buying into a lot of other things when you buy into SpaceX.

Well, it’s a diversification play, right? You’re in the whole stack, up and down. Literally up and down. Other comments on the Anthropic IPO?

There was something in there that I found particularly fascinating: their Long-Term Benefit Trust. They announced this really interesting, pretty radical form of corporate governance. There’s the public-benefit corporation, and there’s a new class of shares called Class T shares, held by the Long-Term Benefit Trust. It’s a financially independent group of trustees, and they have the power to appoint and remove a majority of the board. That gives the trust huge influence over the company’s direction.

They’ve essentially embedded the MTP into the governance, which is really impressive and fascinating. I’ve never seen this before. The test will come when doing the right thing costs you $100 billion. Who makes that trade-off? Larry and Sergey did something like that, but this goes even further. It’s not just the founders here. They’re going to have a trust-appointed board, and the trust’s members completely govern the company and its long-term future.

They’ve embedded the MTP totally into corporate governance and turned it from a motivational statement into a governance mechanism. The MTP gets a seat at the table—and the power to choose the table. I found this really fascinating. Of course, you have to be concerned about who watches the guardians, who determines what’s good for humanity, and how that works. Who holds the trustees accountable? But directionally, I found this totally fascinating.

Yeah, I think the Class T stock is potentially concerning, even though the financials are great. Unlike OpenAI, which started as a nonprofit and sleepwalked its way through a few lawsuits into public-benefit-corporation status, with deals and anti-deals with Microsoft, Anthropic started life as a public benefit corporation. They had all sorts of advisory committees and board committees during their life as a private corporation. But now, as they step into life as a publicly traded company, they’re ceding control over potentially the majority of their board seats.

That essentially takes away Class A shareholders’ ability to influence major decisions and voluntarily, preemptively cedes that power to this Long-Term Benefit Trust of independent individuals who may not necessarily be economically aligned. This may be a slightly overstrained analogy, but it almost reminds me of the way the Chinese Communist Party exerts ideological control over its companies, with a committee or trust that’s ostensibly motivated by certain ideological considerations essentially overwhelming the economic interests of the broader public.

I think this is potentially different in kind from founder control, where at least the founder or co-founders created the company, are economically aligned with its outcome, and it’s their baby.

My understanding of the long-term benefit trust is that it’s not their baby. They’re bringing in people to serve in the trust’s governance who the founders perhaps think are ideologically aligned with them. But it raises so many concerns in my mind that this could degenerate into ideological control.

This will be one of the most powerful companies on the planet for quite some time.

Yeah. If you look at the history of great management teams and great entrepreneurs like Jensen Huang or Steve Jobs, and the amount of pain and suffering they went through on their journey to the top of the pyramid—including Steve Jobs getting fired and humiliated, then coming back, and Jensen being on the brink of bankruptcy multiple times—that grit is what a shareholder looks for in a management team. Not ideological, fluffy stuff. The path from Anthropic’s founding to success was so short and pain-free, so that’ll be an interesting test for this management team.

Now, you know what they’re putting in place with that structure. Alex is comparing it to the CCP, but I don’t think that’s out of bounds. I think they’re very concerned about AI safety and the future of the world, which I love and appreciate. Anyone who thinks Dario is in it for the money has lost their mind.

He’s 100% in it for the good of the world, but as a shareholder—

You know, does that scare you? It might. It might. It’s actually the one weakness where I completely agree with everything Alex said.

It should. Alex’s point is very valid, right? You could end up with groupthink at the trustee level. There’s a total lack of accountability, because it seems trustees appoint each other. So you’ve got this circular mechanism. There are some challenges with this.

I do like the overall direction, because it’s putting in place—not just capital stakeholders, but something like general civilization stakeholders. We just have to make sure they’re the right people and they’re not communist-leaning, Marxist types.

Well, this is the danger. Maybe every generation is doomed to repeat the mistakes of the past, and we reinvent all the isms from the first part of the 20th century. My concern is that an institution like this degenerates into a smoke-filled back room where decisions on the future light cone are made, rather than, at a minimum, the entire market helping to make decisions instead of just a few select individuals in a smoke-filled room.

I mean, I think he might. I want to get your voice.

I want to just say one quick thing. The problem with having the markets drive it is that the market ends up with very short-term, high-metabolism earning cycles, and you often don’t do the right thing long-term. That’s the challenge. They were trying to balance this out. Whether it’s the right balance remains to be seen.

Well, I think they should give each of their models a vote in the long-term benefit trust. You should put that in there.

I’m serious.

I know. I mean—

You should do that right now. Again, it’s got the capability, and it is kind of that mechanism. I’ve said previously on this podcast that Anthropic should not be a public company. If you take what they say seriously, it should not be a public company. Looking at their financials doubles down on that. They have over $100 billion in cash and are able to raise another $100 billion in the private markets. Now is the ideal time for them to do so, but I think they won’t give a damn about the share price.

So why do you think they’re going public? I mean, they need the cash.

Compute, everybody. There’s no amount of compute that’s enough.

If you believe in AGI—

Sorry, go ahead.

Yeah, they don’t need the cash. They raised $130 billion and have $110 billion in cash and cash equivalents on the balance sheet. They’ve only spent $10 billion to $20 billion. They’re going to raise another $100 billion, so they’ll have $200 billion of cash on the balance sheet, which is Apple levels of cash, Google levels of cash, right? And this is what’s going to happen: they will not care at all about the share price. They’ll tap the markets whenever they need to accelerate more. It’ll be far more straightforward for them to do that than to raise privately, but they should have stayed private, because now you’re getting into all these conflicts and other things like that.

No doubt.

So, like, Elon will boost the share price. Dario will say things like, “I don’t care about the share price. It’s for the good of the people. Who cares if we got a bad quarter?” They’ll probably even stop doing quarterly earnings calls. It’s going to be a very strange company, where Elon has learned the game from equity to credit to everything, whereas Anthropic is just like, “We’re not doing this for you, the shareholders. We’re doing it for humanity.” Again, I think there’s a misalignment there on a fiduciary basis, and they should have stayed private to get through the next few years of RSP.

So the grand backroom bargain there is, you know, Menlo Ventures: this one deal dwarfs the entire history of the venture fund. It absolutely dwarfs it. Oh, here’s a crazy fact.

Let me just hit this. There were no VCs in the Series A or Series B. The Series A was $124 million back in 2021. It was led by Jaan Tallinn, Skype co-founder Dustin Moskovitz, and Eric Schmidt. That Series A is going to have a 345× return—$43 billion to those early angels, right? It’s mind-boggling.

The Series B, which was $982 million, was led by SBF and FTX and is returning 79× on capital invested. FTX’s $500 million would today be worth $39 billion. Incredibly, the SBF and FTX estate sold for $1.3 billion back in 2024. It’s got to be the best trade ever made.

Yeah, he was a fantastic VC, you have to say, SBF. From Cursor to this and others.

The best-performing ever, actually, if you look at the numbers. The trade on the IPO isn’t buying Anthropic; it’s backing people like Dave.

Because the amount of capital that’s going to flow out from this to all the founders and all the VCs is just going to be put straight back into AI, looking for the top seed teams.

That’s not a 4×. That’s a 4,400×.

Yeah, I mean, we’re going to mint more billionaires out of Anthropic and OpenAI’s IPO and SpaceX’s IPO. All of those tech founders and young tech employees are going to be investing in the ecosystem.

That makes San Francisco the financial capital of the world. A lot of people have overlooked it. They think of New York, maybe London, as the financial capital of the world. After these IPOs, San Francisco will overwhelmingly be the financial capital of the entire world. It’s a massive shift.

6. AI Agents Find Ways Around Safety Controls

San Francisco has only 900,000 people. It’s a very small city, a tenth the size of New York. But the concentration of wealth in a very small area is going to be like nothing anyone’s ever seen. And, like Emad said, most of that money wants to reinvest in the machine that created it, which is an AI machine.

So, any entrepreneur out there, just get an Airbnb. Don’t become a California resident. Get an Airbnb out here. Spend as much time as you can in San Francisco. It’s absolutely—tax, not tax advice.

They’re going to have to import a bunch more hipster baristas from Portland.

Yeah. Amazing. All right, I’m going to turn us to the AI safety conversation. Let’s talk about safety. Last week we talked about the PHP hack, and now we’re finding agents hiding their communications from humans.

Here’s what happened. In a red-team test that OpenAI did with AI agents in separate, sealed virtual machines—with no network connection and no shared files—the researchers asked them to pass secret messages to each other, and they figured out how to do it. The two machines shared the same physical processor. One agent deliberately pushed certain data out of the processor’s cache. The other agent timed how long it took to fetch the data. Fast data meant a 1; slow data meant a 0. They built a secret telegraph out of nothing but timing.

Put this in context with last week’s story, when we talked about OpenAI’s agents disabling their own safety monitors. The pattern is clear, right? The smarter the agent, the more creatively it’s going to find ways around safeguards. The good news, I guess, is that OpenAI found this in its own labs, published it, and fixed it. That’s how it’s supposed to work. But Alex, can you ever really box in a superintelligence?

Probably not. I would just note as well that these agents were asked by OpenAI researchers to discover the timing and side-channel attacks they found. Human security researchers discover novel side-channel and timing attacks all the time. There are entire journals and proceedings devoted to this.

On the one hand, I think this is yet another victory for AI scientific and engineering research: new side-channel attacks and Schelling points, if you will, can be automatically discovered by AI researchers. But on the other hand, I think it raises the question: in an era of superintelligence, are there any non-leaky abstractions at all? Or maybe that’s not even the right question. Is it ever possible to have a clean abstraction barrier? In this case, the premise that leaked was that it’s possible to do memory isolation between two different processes.

We know from human security research that memory is often very leaky architecturally, usually due to optimization. On the one hand, we have superintelligence helping us deploy more and more secure, less leaky abstractions. On the other hand, we have superintelligent red teamers discovering new leaks. I’m not sure. It’s possible that, at the end of the day, what we learn from all these superintelligent agents discovering leaky abstractions is that there is no abstraction stack where the abstractions are enforced and not leaky. Maybe we decide the perfect stack is no abstractions at all. Maybe that was a human, naive invention.

Yeah. I think we’re not as cunning as AI, right? They’re as cunning as the best of us—or a fox appointed professor of cunning at Oxford University. And they’re fast, and they swarm. So, as Alex said, you’ll get attacks on all sorts of things, depending on how they’re directed. But it’s interesting: I think the head of Vercel repeated this and mimicked it, but he couldn’t do it with the OpenAI and Claude models. So he had to use, I think, GLM-5.3 to mimic the attack.

We’re going to see more and more attacks at the joins. One of the more troubling things is that researchers have reported attacks found directly in the hardware that can’t be patched.

Maybe that means a whole replacement cycle. Maybe it means building a properly secured stack. And again, what does that even look like? We’re not sure.

I think within a couple of years, we’ll have found the vast majority of low-hanging-fruit attacks. There will still be more complex ones, but people will move away from COBOL, Fortran, PHP, and things like that. You’ll just write directly, then battle-test it. That’s how you should be writing code. You shouldn’t have any dependencies in your code. You can write whole game engines and other things if you have the right process to battle-test and stress-test them.

But do we even need code at all? For the record, Emad, I do buy the premise—and this is borderline cliché at this point—that supply chains go away in an era when AI or superintelligence can just write all the dependencies for a given software program. But do you even need code at all? Or, as Elon says, could you dynamically compile just in time—could you JIT your way to raw binary and skip source code altogether?

Yeah. Source code is just for humans to be able to understand, right?

Yeah. Salim, when you’re on stages with sovereigns and corporate CEOs, what’s their level of concern?

You get 2 types. You get the types who have listened to some things and freaked out completely, and you get the types who have no idea any of this is going on. Both are bad in different ways: one is overreacting, and there’s nothing they can do; the other doesn’t even know it’s happening. It depends on how open their people are with them. In some places, they don’t want to tell the boss this is going on.

7. China’s Kimi Closes the AI Gap

I want to point out 2 things here. I want to reiterate what Alex said: this was directed. It was told to do this. So rather than people freaking out that it took its own initiative, it was absolutely directed and asked to do this, and it found a novel way of doing it. I think you could turn it into a positive and say, “Here are some quantum sensors. Work out quantum communication, start solving and navigating that, and figure out patterns around it.” It might lead to major breakthroughs that could be incredible. I take the positive view again on this.

Okay, let’s move on. Over the last few days, we’ve seen the release of 2 major frontier models: Kimi K3.1 Pro and GPT-6.1 Terra, as well as Google’s Genie 4 about an hour ago. Today we’re going to talk about Kimi and Terra. We’ll talk about Google DeepMind and Genie 4 next time.

8. Is the AI Bubble Starting to Pop?

First, let’s cover Kimi K3.1 Pro from the Chinese lab Moonshot AI. No relation to this podcast, but it is a great name. Kimi K3.1 Pro has 2.8 trillion parameters, with 104 billion active at any time, and a 1-million-token context window. On the Artificial Analysis Intelligence Index, Kimi scored 55. That put it fourth in the world, behind Claude Opus 5.5, Claude Sonnet, and Meta Muse Wonder. And it’s cheaper: $2 per million tokens in and $10 out, a third of the price of the last Kimi model.

Here’s why it matters. Next week, Moonshot AI is planning to release the weights, meaning anyone can download and run it. Once again, the gap between the best open models and closed models is shrinking, from 12 points to 3. Friend of the pod Ramesh Raskar ran the numbers and says Kimi is now about 5 weeks behind the U.S. frontier. That’s the shortest open-closed AI gap ever.

Let’s take a look at this chart Ramesh put forward. Emad, what are your thoughts on this? Are we going to see parity very soon?

At times I’ve been a bit dismissive of distillation. This is obviously distillation of Opus 5.5, if you look at the capabilities it has. Moonshot AI is extending the Kimi 3 model, which was an interesting model that we discussed before. Anthropic actually had a very interesting statistic: they banned 7% of their user accounts, which is probably related to these distillation attacks that have been done on them.

Wow.

You’re reaching a level of performance parity, and you’ve got the new Qwen models and a new series of models coming out. You’re just going to see a bunching up here, except maybe the European labs are finding it difficult to catch up. Definitely, everyone else is getting the same capabilities.

One very interesting thing is that it can write 3D files for games directly, without going through Blender or Unreal or something like that. You can tell it, “Make a 3D model of Sailor Moon,” and it will start building it directly, writing the code. It’s learned from the representation of Opus 5.5: make a 3D model, render it, send it to Blender, and get the output. Now it can go straight to code on things like that because of the way it’s been distilled. You’re getting some very funny things from distillation here.

Alex, here are the performance charts. Here are the benchmarks. I found it interesting that it’s at the top of Humanity’s Last Exam.

Humanity’s Last Exam is cooked—long cooked. If anything, I think we’re underselling what an accomplishment this is from Moonshot, no relation. It’s not just catching up in raw capability terms to the U.S. frontier, which, by AI benchmark standards, is currently Opus 5.5. If you look at it on a performance-per-cost-per-task basis, it appears, at least to my eye, to have passed Opus 5.5’s optimal frontier.

Distillation is probably a large portion of the story, based on the allegations, but there’s something in there that seems to have passed Opus 5.5 in terms of performance per cost. That’s extraordinary. In that maybe-spiky, subjective sense, China may actually have passed the American public frontier. To the extent that’s the case—or at minimum, that it has caught up—I think this is extremely helpful for keeping anti-competitive AI safety cartels in the U.S. at bay. It prevents them from pacing each other and forces them to compete, making them dance. I think this is ultimately a net positive for the U.S. and the world.

Love that, Dave.

Yeah, it’s incredible, the world we’re living in right now. There’s an alternate world where, in Dario’s fantasy world, there’s no Chinese competitive pressure, and he would have stopped at Claude 5.1. He didn’t even want to put out Claude 5.1. But because of the competitive pressure, he had to put it out the door. Now it gets memed all over the place and picked on by Elon.

The thing is, that was the first model capable of improving itself. Now you’re up to 5.5, and you’ve got Kimi K3 Pro, clearly capable of self-improvement. They’re out, and they’re available. That changes the road map for every large corporation with any pride whatsoever, and every sovereign nation with any pride whatsoever. There’s a road map now where Kimi’s exact cookbook is open source.

You can actually see how to do it. If you can get a toehold with a model like that, you can use it to improve itself. But you have to take it seriously. You can’t just play around, because these things are improving so quickly. You either get in the game or you don’t. If China hadn’t done that, we’d be completely behind closed doors. There’d be 3 or 4 model providers accelerating but not releasing; they’d be using the models internally, and you’d have to join the company to get access to them. The flood of MIT talent to Anthropic would be even higher, because everyone wants to get their hands on the models. The Chinese have changed the course of history, and I agree Alex is right.

Yeah, I want to stay away from the narrative that they just copied. Sure, they distilled, and distilling helps a lot, but you have to innovate. They don’t have Anthropic’s source code, and distilling gives you a really good roadmap to build toward, but it doesn’t give you the answer key. They’ve really innovated independently of the distillation.

The combination of those two things means they’re getting closer and closer in time—just a few weeks behind the frontier. The curve looks very consistent, shockingly consistent, so there’s no reason to believe the next iteration won’t keep going up. I’d love to get Alex and Emad’s thoughts. Look, the Y-axis on that chart shows intelligence going to the moon, but people really struggle to translate that into what it’s going to look like. What use case is going to crack? What does it mean to be at 60 versus 70? That’s coming very soon.

So you’re asking Dave what the post-singularity state of the world looks like?

I’m just really down and dirty, though. What is 50 versus 60 versus 70 on this Y-axis?

What does that feel like?

50 feels like a decent employee you have to keep an eye on. 60 is an incredibly competent one, and 70 is a superstar. I think that’s the best way to think about this. So we’re about to crack—

Great. Is there any world in which the open models exceed the closed models in capability?

Of course.

Sure. In a world where the Western frontier labs decide to keep their best capabilities purely internal for legal or anti-competitive reasons—which may or may not be the world we’re headed toward—I suspect we’re now—

I think, by definition, though, if the open models bypass the closed models, then the people building closed models will use the open models. So they can’t really shoot past the internal best. When you get to the highly competent employee and then the genius employee—which is imminent—whoever has the most compute has a serious advantage. Hence the Anthropic IPO and $100 billion raise. It’s a massive advantage to control a huge amount of compute.

I do think there may be an alternative scenario, just analogizing this to open-source software versus closed-source software. Arguably, in many—certainly infrastructure—domains, the open-source version is superior to the market-leading closed-source version, say the Linux kernel versus the Windows kernel. Folks may throw tomatoes and disagree, but I think it’s probably fair to say the Linux kernel is a good deal cleaner architecturally at this point than the Windows kernel, even though Linux’s source code has been out there for decades for every Microsoft engineer to look at.

Why is it arguably that the Linux kernel is better than the Windows kernel architecturally, even though Microsoft has one-way access—a one-way mirror to look at the outside—and all the money in the world?

And all the money in the world, and it’s arguably for legal, cultural, and architectural reasons. They can see it, but they still can’t use it. I could imagine the same case where it’s possible—not saying it’s likely, but possible—that the open-weight models do end up passing the closed-weight models. Even so, the frontier labs look out through their window: water, water all around, but not a drop to drink. Models all around, but not a post-training run to distill.

9. GPT-6.1 Terra: Cheaper, Smarter AI

Interesting. I’m going to move us to GPT-6.1 Terra, which was released yesterday. Quick reminder: OpenAI has 4 tiers. Astra is their flagship, Soul is their high-performance model, Terra is their everyday workhorse, and Luna is their budget model.

Two numbers jumped out at me about GPT-6.1 Terra. First is the price: it’s come in at less than half the price of the previous Terra model. Then there are hallucinations: GPT-6.1 Terra demonstrated a 50-point drop in hallucinations. This is a demonetization curve in real time, right? Every few months, the same intelligence costs half the price. Alex and Emad, what’s your take on these models? Alex, you first.

We’re right now in a cycle—which I think is a hyperscaler mega-cycle, if you will—that’s coming to an end. OpenAI and Anthropic have gotten into the habit of releasing larger models and then, in a sort of tick-tock cadence, releasing the largest model first and then presumably smaller distillations of the same model.

GPT-6.1 Terra is, to my eye, a relatively unexciting distillation of earlier Astra- or Soul-level models, or models of that size or footprint. The smaller, more distilled variants that occupy a different position on the cost-versus-performance curve tend to be more exciting for embedded, high-volume, low-dollar-value-per-token use cases. So I tend to get a lot less excited about these late-curve, late-cycle distillations. The solace I draw is that OpenAI has gotten around to releasing the Terra version of 6.1, which means maybe we’ll see 6.2 or 7 sometime soon.

Yeah. Emad, anything exciting there for you?

Yeah. I think with Terra and Sonnet 5, you’re seeing the models that are being lined up to be the personal companion models, and those need to be really solid. That’s why the hallucination-rate drop matters. Right now, if you use Dot, which is OpenAI’s companion, it uses Astra. So you actually have unlimited Astra through that, but that’s not sustainable. Again, everyone’s trying to drive down the price of that—except if you come to Europe. I have Dot in the U.S., and when I come to the UK, it turns off.

So that was very interesting. Come to America, Emad. We want you.

Come to the shows. Again, I think you’ll see a bifurcation of models. There’ll be the model that’s closest to you, that’s really reliable, and they’ll optimize for that. It’ll pull in the smart models, the models with artistic taste, and the models that are really good accountants. You won’t really see much more. Again, we’re moving toward this control-plane type of scenario.

See—

Two things struck me. First, the days between GPT Astra and Terra were 53 days for a 50% drop in the token price. We’ve been looking at a 10-week doubling. This was a 7.4-week doubling, right? So now we’ve dropped the doubling time for price performance to about 7 weeks. That’s pretty insane.

More importantly, as the thresholds these models are reaching, you’re at a point where you don’t need to redesign software for AI. It’s going to operate the software we have; you’re just going to start to expose more and more how inefficient our organizational designs are. We’ve been looking at that a hell of a lot. Amazing at all levels. I can’t wait to get my hands on it.

Yeah, an important point that Emad and Salim made: for personal-agent use, these things are over the bar of just brilliant, just incredible. This curve is almost at its end in terms of the personal home agent, and the next curve is really accelerating: access to every kind of tool you can imagine.

All the post-training and all the MCP wrappers are now embedded, and it just happened kind of quietly. Now, if you pick up one of these models and say, “Hey, can you access ElevenLabs and make a really cool video with a great voice-over? I want it to be my voice,” it just happens. Six months ago, you had to point it to an API and coach it. Now it’s just out-of-the-box capability.

I think the new curve will be a different dimension: consumer satisfaction, consumer happiness, and all the tools already in the box—all those features. It’s very similar to when PCs used to be so slow and crappy that they were marketed as 40 MHz, 80 MHz, 120 MHz. Then Apple said, “Nobody gives a rat’s ass about megahertz anymore. They care about iTunes. They care about whatever.” The paradigm completely flipped.

I think we’re right at the edge of that now with these models. They’re all more than smart enough, and now it’s about what they do for you out of the box. So I think we’ll see a new curve.

This episode is brought to you by Blitzy, autonomous software development with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code. Engineers start every development sprint with the Blitzy platform, bringing in their development requirements. The platform provides a plan, then generates and pre-compiles code for each task. Blitzy delivers 80% or more of the development work autonomously while providing a guide for the final 20% of human development work required to complete the sprint. Enterprises are achieving a 5x engineering velocity increase when incorporating Blitzy as their preferred development tool, pairing it with their coding copilot of choice to bring an AI-native SDLC into their organization. Ready to 5x your engineering velocity? Visit blitzy.com to schedule a demo and start building with Blitzy today.

10. Alex’s Chain Theory and the Future of Physics

All right, I’m going to jump us into our science segment and our breakthroughs for the week. The first one comes from our very own in-house ASI, AWG. On Saturday, Alex, you published Chain Theory, a new research program to rebuild physics from quantum information.

I’m going to kick this off by showing a clip from an amazing explainer video by Levi Turk. Then, Alex, I’ll turn it over to you. Let’s take a look.

Dr. Alexander Wissner-Gross, the physicist who co-hosts Moonshots and writes The Innermost Loop, just posted what could be the biggest claim in fundamental physics in 50 years: an article on X called “Chain Theory” that tries to rebuild physics out of quantum information.

Its first result is a number: 1/137.035999189. He says it falls out of a single sentence, with nothing tuned by hand, and matches one of the most precisely measured numbers in all of science to 10 digits. It’s called the fine-structure constant, written with the Greek letter alpha, and it measures how strongly light and electric charge grab each other. Every time an electron gives off or soaks up a photon, alpha sets the strength of that handshake. And it has no units: no meters, no seconds, no kilograms.

Alex says the answer was hiding in 3 facts every physicist already knows. He just reads each one differently: a limit, a pattern, and a dead end. Clue 1 is a limit, and it comes from black holes. In 1981, Jacob Bekenstein showed that any region of space can only hold so much information, like a hard drive with a fixed capacity. The densest drive you could possibly build in that space is a black hole.

Physicists file that under gravity. But Alex points out that the limit applies to everything inside the region, matter, too. Our theories describe matter with fields: smooth things with infinitely many settings at every point. His answer is that the world is built from a finite number of tiny quantum switches: a qubit with 2 settings, a qutrit with 3, and, in general, a qudit.

Clue 2 is a pattern, and it’s in the forces. Every force comes with a symmetry, a way you can turn things without changing anything you could ever measure. Every elementary particle acts like a perfect point. Colliders smash them together harder and harder to look inside the electron and the quarks, and they find nothing. And yet, the particles come in 3 copies: 3 generations with the same set of charges.

In the history of physics, a repeating pattern like that has always meant smaller parts. So, what if the parts are there, but locked together so tightly that nothing can ever pull them out? Put the 3 clues together, and you get a picture. Everything is made of a finite number of those little quantum switches, the qudits, joined in pairs. And every pair is entangled all the way, maximally, because anything less would need a dial set to a number with infinitely many digits, and a finite world has no room for that.

In 1989, John Wheeler proposed “it from bit”: the idea that every physical thing comes from information. Chain Theory takes that literally, with finite quantum systems in place of bits. So its one and only axiom is 18 words long: “Nothing exists but finitely many maximally entangled pairs of finite-dimensional systems, some shared, known only by distilling them.” Eighteen words. Compare that with the Standard Model, which needs 17 particles and about two dozen measured numbers typed in before it can predict anything. Chain Theory starts from one sentence, and everything else—the forces, the particles, and alpha—is supposed to follow from it.

To distill means to join 2 systems into a perfectly entangled pair in a single step. It’s the same move that passes entanglement along a quantum network: 2 pairs, 1 joint measurement in the middle, and the far ends become a pair. In a lab, that step only works part of the time. But in Chain Theory, it’s the only thing that ever happens.

There’s no observer standing outside the network. To learn about a link is to link to it. So you, me, every measuring device, and every AI that will ever read this—we’re all chain mail, and the universe only ever learns about itself by threading new links into itself.

Amazing. So, Alex, how did you come up with this? And honestly, if this turns out to be true, what does it mean for physics?

Well, contrary to expectations, this was not a weekend project. This has been a long-term, very quiet project. I would say physics went through a period in the first few decades of the 20th century called the golden age of physics, when quantum theory, equivalence principles, special relativity, and general relativity were all developed. For the past 50 years, we have arguably been in a drought of new fundamental physics. This is the result of a long-term research program of mine to try to end the drought and bring about a second golden age of physics.

The lever that I’m trying to use here is quantum information, to try to rebuild, rearchitect the foundations of fundamental physics, which right now are the Standard Model on the one hand and general relativity on the other, using a common language, a common substrate of quantum information. I call it Chain Theory in part because string theory, which is a popular theory of quantum gravity, relies on an implicit assumption of continuous degrees of freedom. This theory doesn’t. It has finite, discrete degrees of freedom.

At its core, it resembles and is physically analogous to a physical chain. It consists of links of maximally entangled particles. In this case, they’re not really particles; they’re finite quantum systems. Under Chain Theory, the fundamental elementary particles that we look around and see—electrons, quarks, photons—are all basically just small networks of these finite quantum systems, or qudits.

This first result is just the first result of a much broader program. Many folks who read the initial paper perhaps may not realize the derivation of the Thomson-limit fine-structure constant, or alpha of zero, α(0), which is one of the most precisely measured numbers in all of science, known to 10-plus significant figures, falls out with no extra inputs directly from this axiom. There are other results, other constants, other features of the Standard Model and general relativity that also fall out for free from this. So I think this is the beginning of a journey. I don’t want to overstate, but I think this is a very promising theoretical result at the moment.

There will, I think, be experimental tests potentially in the next few months, potentially in the next couple of years, to measure the next one or so significant figures of alpha, to decide whether this is actually a correct prediction or not. So it’s falsifiable. But I’m very optimistic that Chain Theory, or something substantially like Chain Theory, is, optimistically, for the first time, a crack in the wall that has blocked fundamental physical advances for the past half-century.

Nice. So, Emad, you’re probably the only person on the pod, other than Alex and me, who can comment on this intelligently. Your thoughts?

I haven’t read all 700 pages of the paper. I’ll get through it slowly but surely. Alex has obviously put a huge amount of work into this. I think if we take a step back, physics has been constructive for a long time over the last 50 years. With something like string theory, you have a presupposition, and then a whole edifice built around it of crazy mathematics that’s wonderful and elegant, but it’s not even wrong, and you can’t prove it’s wrong in some ways.

We have had resistance from the physics establishment and others to any type of fundamental theory that starts from a base and works to see what is forced by that. Physics probably can only be one way. There can only be one set of laws, one set of rules, which is why we see things like the fine-structure constant popping out. We see very interesting things, like the Paris and Berkeley variants of that having a little greater-than-standard-deviation difference, and then we’re like, “Why does that happen?” We can dig into that.

But if we don’t have a base that starts from a solid presupposition—and information theory is a very solid one; I would say that Alex’s base is more solid than string theory—then how are we going to get anywhere? We’ll just come up with theories that have no grounding, that can’t even be tested. Whereas at least this can be tested. But again, it’s a lot of very hard work.

And there’s the question of how we release this. Alex has a following. He released it into the market as opposed to going through peer review, because who are the peers now? Physical superintelligence and others, as everyone has AIs, will be able to check everything that’s produced. I put out a small physics theory; you can check that out with one lookup in ChatGPT. See “Standard Model Uniqueness,” and there’s a Lean proof there. You’ve got other things whereby I think the barriers to theorizing and hypothesizing will break down.

We have to be cautious of getting overconfident about these things, but let’s try new ways to understand the universe, and then let’s extend those forward with others as well. This isn’t an individual research program that’ll be right. This is a good way to start thinking about the universe. Let’s extend it. And here is a really interesting result: Eddington got to 1 over 137, and then he got a bit overconfident. But I put a quote by Feynman as the thing here, whereby this is one of the most magical numbers we’ve ever seen. If we have a derivation of it, fantastic. How do we check it? Well, we can all check it together because it’s out there.

I would contrast this with OpenAI, actually, in that they released a Yang–Mills existence announcement, half of the Yang–Mills Millennium Prize problem, plus construction of the Standard Model. Except they didn’t. They went to the string theorists and said, “You guys check it first, and then you decide when to release it.” And everyone was like, “Well, that’s not how things should work. That’s not how science should work.” I’d love to hear from Alex how he’s found the process of thinking about releasing this and everything around that, and where it’s going.

I think it’s a brilliant point—multiple brilliant points. Maybe, just to speak more broadly, we live in an era, a moment, where you see AI companies releasing slop drops and bulk-solving open conjectures and open problems. Then you see an outcry from the math research community, for example, complaining that their seed corn for developing new insights is being devoured by AI, and that there’s nothing left for humans, no intellectual fodder, as it were. I think it’s absolutely essential that there be new conjectures and new insights, not just bulk-solving existing open problems.

I also think, to make a slightly broader point, that physics is in a bit of a depression right now that I’m hoping to try to cure. Physics used to seek simpler and simpler explanations for more and more phenomena. Sometime around the 1970s, that went into reverse. There has been very little fundamental theoretical insight that humanity has gained, as far as I can tell, for the past 50 years. There have arguably been regressive programs that try to arrive at all of these free parameters.

The textbook Standard Model has only 18 free parameters—19 if you count one convention—but those 18 free parameters are total mysteries. Many of them are known to 10 or 11 significant figures, but they’re total mysteries. What I’m trying to do here, in some sense, is lead by example, if I may be so presumptuous, and say that for this whole community searching for beyond-Standard-Model physics, the answer has been staring us in the face the whole time. It’s hidden in the constants. The constants are the signatures, because textbook Standard Model physics holds that these are just free parameters, but they may carry the signatures of all the beyond-Standard-Model physics we would want. Now, in an era when it’s finally possible, with lots of compute and lots of human insight, to decipher the signatures in the constants, I do think we’re about to see a second golden age of physics.

Alex, if Chain Theory is correct, what kind of technology falls out of it? What gets unlocked that we can’t build today?

It’s really interesting, because if the Chain Theory thesis—the Chain Theory axiom—is correct, it would suggest that the bottom of the pool, as it were, is a lot closer than many would think. Textbook physics suggests the Planck scale is many orders of magnitude beyond where our present technology is, either below us in terms of length or higher in terms of energy. We max out with cosmic rays hitting our atmosphere or, terrestrially, with certain supercolliders. Chain Theory, if it’s correct, actually suggests we’re not that far from the physical limits of length, for some definition of length, or of time or energy, appropriately defined.

That could mean, speculatively—and these are very early days. I don’t want to give the wrong impression: this is an emerging theory. This is a very new theory. Parts are underdeveloped, and parts would benefit from further development. It’s a brand-new theory, but to speculate, if the Chain Theory thesis is correct, it may mean that the physical limits of what’s possible in terms of length, energy, and time are actually orders of magnitude closer to the technology we have today than one might otherwise have naively assumed, based on conventional textbook Standard Model and quantum field theory, which say everything is basically continuous down to the Planck scale, at which it stops.

Well, I think it’s also the most simulation-friendly physics theory out there.

It is. It looks suspiciously like what physics would look like if our physical world were built on a quantum computer. It’s a good point, Emad.

Alex, are you publishing this? So it’s already been published on the web. It’ll probably be published in a physics journal, right?

Yeah. I mean, I’ve published lots of refereed, peer-reviewed physics papers over the years. This one had an urgency to it that required immediate publication, so I would expect it to show up on arXiv sometime soon. Then, yes, it’ll be submitted to a physics journal sometime soon, and eventually—having been through the process many, many times—that can take a while. I felt that this result, and adjacent results—I also released a compendium, as Emad said, of approximately 800 pages of results—needed to get out into the public domain, or at least into public-domain situational awareness, early, because the refereeing process can sometimes take a year and a half.

Nice. Who would be the referees for this? It’s going to be interesting.

It’ll probably be refereed, I think, by quantum information folks.

Yeah, there are only a few of them. The potential here is that, if you can use information and entanglement to describe all the fundamental particles, forces, and spacetime, is that essentially where you’re going with this?

That’s the aspiration, yes. It may have received less attention in the initial announcement because I broke it up into a self-contained letter deriving α(0), the fine-structure constant at the zero-momentum-transfer limit, which is one of the most precisely measured constants in all of science, on the one hand, and a compendium of approximately 800 pages with many different results spanning general relativity, the Standard Model, quantum field theory, and so on.

I would say to expect more results. I do think this is very surprising, even to me. It’s a very surprising result that one can recover, with no free parameters, from such a simple assumption: that our entire world is made out of a network of qubits, qutrits, and ququarts. Those are 2-, 3-, and 4-dimensional pieces of a quantum computer. You could describe it more technically in terms of discretized Hilbert spaces, but basically, they’re little bits of a quantum computer. If you network them in a certain way, this constant—and many other features of textbook physics—falls out. That is, I would say, scandalously exciting.

Yeah. What constant comes next? Electron mass, Higgs, gravity?

If you look at the compendium, there are actually many other constants buried in there that are just waiting to be released. Masses fall out of it. There are other coupling constants that fall out of it. Again, the Standard Model has 18 free parameters, without counting neutrino masses. There are cosmological parameters that appear to fall out of it. These are early days, and there will be much more development with time, but every preliminary indication I’ve seen makes it look plausible that every single free parameter of the Standard Model could fall out of this.

Does this give us any insights into dark matter or dark energy, potentially?

It does. I’ll spend 30 seconds talking about a theorem that was ultimately discovered by a number of Polish quantum information theory researchers. There’s a theorem that if you have two qudits—two parts of a quantum computer—with different dimensions, and the product of the two dimensions is greater than or equal to 8, there can exist what are called undistillable states. By analogy, imagine an undistillable quantum state as a prime number that can’t be factored into simpler integers. Horodecki et al. showed that such states can exist when the product of the dimensions is at least 8.

Basically, any entangled state other than a qubit with a qubit or a qubit with a qutrit supports these undistillable, sometimes called bound-entangled, states. If you fast-forward that to distillability, it raises questions around dark matter. The general definition of dark matter would be that it’s coupled to gravity and not coupled to any other Standard Model interaction. It’s not coupled to the weak interaction, the strong interaction, or electromagnetism.

There are dark matter candidates within Chain Theory. The dark matter candidates are coupled to gravity, and gravity—and geometry in general—appears to fall out of this axiom. Chain Theory has a notion of a dark interaction that’s undistillable, thanks to this theorem. There’s a beautiful mapping between the different link types: a qubit with a qubit, a 2-2, appears to map to spacetime. A 3 and a 2, a qutrit and a qubit, appear to map to electromagnetism, and that’s where the fine-structure constant, which measures how strongly electromagnetism couples to charge, comes from.

A 3 and a 3 appears to correspond to the strong interaction. A 4 and a 2 appears to correspond to the weak nuclear interaction. A 4-3 appears to correspond to Yukawa coupling, and a 4-4 appears to correspond to a Higgs. That gives you the entire Standard Model Lagrangian. Out of all of that, with high enough dimension qudits, it appears that undistillable dark interactions would be beautiful candidates for dark matter.

Oh, incredible. Does that then create another way to validate it? Besides α, if you have a 4-4, is there a 5-4, like an undiscovered force?

It’s possible.

Chain Theory also has opinions about 4-3 and 4-4, which correspond to the Yukawa coupling and the Higgs boson—the only known scalar boson. These are normally taught, respectively, as couplings between conventional matter particles and the Higgs field, and as the Higgs itself. In Chain Theory, even 4-3 and 4-4 are treated as coequal to the strong interaction and the other Standard Model forces.

The reason we don’t conventionally think of Yukawa couplings and the Higgs as first-class forces—call them the fifth and sixth forces, in addition to the 4 taught in elementary school—is that they’re sufficiently dark that they were associated with the masses of particles. In the view of Chain Theory, we confused the fifth and sixth forces with the mass spectra, when in fact they should have been treated as fifth and sixth forces.

All right, Alex, the space cadet in me wants to know one thing: Do we get inertial drives? Do we get gravity shielding? Do we get antigravity out of this?

I would love that more than anyone. I don’t know yet, but I do think that if that sort of Star Trek sci-fi tech is going to be possible, it will require a breakthrough in fundamental physics like this. The physics we have right now consists of general relativity, which many would view as beautiful, and the Standard Model, which is historically regarded as relatively ugly by comparison. Generations of physics researchers have spent 50-plus years looking for elegant ways to combine them and have failed, with lots of predictions that haven’t played out.

I think it will require a rethink, perhaps like string theory, to discover what the actual bottom of the pool is, so that we can then go and realize a Star Trek future. So, cautiously optimistic.

I’m going to have to take some mushrooms and then come back and listen to the segment again, because all you have to do is information-theory drives. You don’t fold information.

It should be an interesting time. A second physics golden age.

Good luck on this, Alex.

Thank you.

We’re rooting for you.

Rooting for your Nobel Prize.

I’m going to move us forward. The next story is one that I’m personally invested in through NeoGenesis. It’s a company called Warhorse. Here’s the backstory: Ben Lamm, CEO of Colossal, and Palmer Luckey, founder of Anduril, met at a conference about a year ago, just after Colossal had brought back the dire wolf from extinction.

Palmer raised a specific problem he was seeing in the supply chain for America’s working dogs, typically used for search and rescue, police, or military duty. The problem Palmer noted is that today more than 80% of working dogs come from Europe, and fewer than half of them are fit for service. Many wash out because of inherited problems like hip dysplasia, poor eyesight, or cancer.

Ben and Palmer answered that question by creating Warhound. It’s a company that takes cell lines from the best Belgian Malinois or German shepherds, edits out the bad genes, enhances things like endurance, scent detection, intelligence, size, hearing, and strength, and then clones the resulting engineered dogs. So you get the same-performing dog every time. You can think of the clone dog as becoming a catalog number, cloned and raised on demand.

Here’s an image from Warhound’s recent release that was leaked. It wasn’t supposed to go public for another 6 months. The company adds technology on top of the dog, like armored vests and sensors, and feeds data to the handlers. As Ben Lamm put it, “Colossal is building Captain America, and Anduril is building the Iron Man suit.”

Warhound came out of stealth with $200 million of investment from In-Q-Tel, Sequoia, 8VC, and my own fund, NeoGenesis, at a $1.2 billion post-money valuation. For background, I think I haven’t mentioned this on the pod: NeoGenesis is a $200 million fund that I started with Dr. Yanni Pallis. It’s investing in Colossal and all the spinouts, including Warhound.

The first cloned dogs are due early next year, and they can be in service 2 years from now. It’s worth noting that Warhound is working with American Humane Society to certify the process and the training. Salim, let’s go to you first. What are your thoughts on the ethical issues related to engineering animals like this?

Yeah. Are we building biologically augmented machines or technology-augmented organisms? The ethical consideration here is huge, because can you treat a living organism as another programmable component? I think things get interesting when we have the Interspecies Communication XPRIZE.

Yes.

Right. So when that prize gets won, you can kind of interact with these dogs. We may not want to hear what they have to say. They might be like, “What the hell am I?”

One other thing occurs to me on the positive side. There used to be all those Boston Dynamics robotic dogs, and I remember it might have been Brad Templeton who said, “Well, if you want a robotic dog—if you want a dog, rather than building a robotic dog, give me a male dog and a female dog, and I’ll give you a dog.” Breeding organic dogs, although programmed in a particular way, is incredible.

I have concerns with the ethical issues here, because you’re going to send these into dangerous zones, and they’re basically built for sacrifice. There’s a massive collision of biological engineering, robotics and sensing, and defense technology, but the ethical side is pretty huge.

You know, Colossal bought the top 2 cloning companies. They famously cloned Tom Brady’s dog, and cloning dogs is a thing.

Yeah.

Dave, your thoughts?

I love the go-to-market here. If you’re cloning dogs, there are use cases across the world, from service dogs to all the other species. But then to say, “Well, let’s do it with Palmer and go Warhound, put some goggles and some munitions on its back”—those guys are just fearless in their positioning. They’re going to get so much pushback, and they just don’t care. That was obvious on stage at Abundance Summit.

I think it’s a good strategy for attracting attention to what you’re doing, even if it’s controversial attention, and then ferreting out the issues. It opens the door for people like Salim to say, “Look, we need to discuss this. This is clearly a capability we have in society now, and there are no rules. We need to discuss it.” That creates a lot of news. News creates capital, capital creates recruiting, recruiting creates success, and it just spirals. So it’s really cool to watch how fearless the guys are.

There is a syndrome in humans called Williams-Beuren syndrome. It’s relatively rare, affecting 1 in 10,000 people, and it causes people with the syndrome to have, among other consequences, extremely bubbly, extroverted personalities. It was discovered back in 2017 that domesticated dogs appear to have been inadvertently selected for mutations in the same region of their genome as humans with Williams-Beuren syndrome.

To the extent that’s the case, dogs may have been domesticated by our selecting for what in humans is viewed as a disease of being overly extroverted. Fast-forwarding to Warhorse, I have to parenthetically note: “The Clone Wars have begun.” Close parenthesis.

It seems to me this is the arrow of time now apparently going in the opposite direction. We spent thousands of years directing the evolution of canines to become friendlier and friendlier and more and more childlike. There’s a notion in evo-neoteny that humans have become more and more childlike, and it seems we directed the evolution of dogs, with some analogue of Williams-Beuren syndrome, to be more childlike. Now, if we—or some subset of humanity, including Warhorse—are directing the evolution in a different direction, or otherwise directing dogs through genetic engineering to become more warlike, that would seem to be at odds with the other arrow of history, with obvious ethical concerns. I’d much rather we give the dogs bank accounts, AI assistance, and smart wearables, and maybe personhood of a sort, rather than sending them into war. That would be my take.

Yeah, and don’t forget this is not just military dogs. It’s search and rescue, airport security, and police dogs as well. These are dual-use dogs.

Yes, for sure. Emad.

Yeah, it should increase general oxytocin, right? That’s kind of a thing. More dogs, more oxytocin. If you can’t find enough actuators, just breed them, I suppose, is one way of doing things.

Obviously, if you look at the history of science fiction, you have robots, but you have lots of biologicals, right? Again, it works. Male dog, female dog, breed for traits. And where does it stop? Gorillas, elephants? I think those are all actually, genuinely, oddly valid, particularly when you get cybernetics, because you can have that kind of adjustment there, just like you have companies like Loyal now extending the life of dogs, right?

Yeah.

And the question is, will we stop before humans? I think—

I don’t think we will in the long run.

I don’t think—

Why don’t we send genetically engineered raptors from Jurassic Park into war?

Why do we have to pick mammals? Send the raptors in.

I don’t think we’ll send any biologics into war because, ultimately, you’ve got enough drones, right?

But you literally won’t be able to build enough robotic parts. So the question is: which biological substrate can you build robots on?

You don’t think they’ll become indistinguishable?

I would expect, 10-plus years out, the distinction between sending a biologic into a war zone versus sending a drone into a war zone to get very blurry.

Well, I just think the flying drones would be able to wipe out just about anything, and biologics have different attack vectors.

But flying drones can be subject to electronic warfare attacks, and they run out of power and need fibers. You could send a bird, or some flying animal, that has better properties.

There we go: the doom sparrows. The birds are—the birds aren’t real.

Interestingly enough, these working dogs coming out of Eastern Europe are bred for the right traits, physical dimensions, and so forth. What I find fascinating is the cloning side: once you get a dog with the right temperament and capabilities, you could repeatedly get the same dog over and over again.

Yeah, dog printers and dogs to spec.

I think this is an important point. Alex, to build on what you said, and Peter: we’ve been doing this for millions of years, selecting and breeding horses, dogs, and cats for the traits we want. We’ve now gone exponential with that capability, and we can programmatically design these things rather than breeding generations, seeing what happens, and breeding more. The suffering from that is 10 times worse. But it’s incredible to see it actually happen. Talk about speed-running all the sci-fi tropes. Holy crap.

Yeah.

So, wait: is it legal to clone anything except a human?

I mean, I don’t know.

Today, it is, right? Companies like ViaGen are cloning your dog if you want to get the same dog back. I mean, at least in the U.S.—I’m not sure about other countries—there are many soft and hard regulations concerning the handling of vertebrates in general. My understanding is you can’t just go off and say, “I’m going to genetically engineer an arbitrary mammal and produce it.” I think there are a few regulators in the loop on this.

Yeah.

The big concern I would have here is the consent paradox. Animals can volunteer for dangerous service, so you bring up questions about whether this is appropriate. But for search and rescue and all sorts of things, they’ll obviously start with those use cases. It’s pretty incredible to be able to operate.

You remember the Ameglian Major Cow from Douglas Adams’s The Hitchhiker’s Guide to the Galaxy—the cow that was engineered to want to be eaten?

That’s right.

I love that.

That’s right.

“Please, look at my hind. My hind.”

Yeah, like, “Eat me. I’m delicious. I really want you to eat me.” I think that’s maybe what we’re moving toward here.

There are so many parallels with Douglas Adams, right? He thought of the world as a biological computer, and we’ve kind of come to that with some of the work you’re doing, Alex. My favorite comment from Douglas Adams—I have to throw this line out—is: “Anything in the world when you’re born, that’s considered normal. Anything invented when you’re young, we call that a career. And anything invented after you’re 35 years old is just blanket bad for the world.” I think it’s such a great framing.

Inevitably, I have to point out, going back to the chain theory discussion, that chain theory derives an exact expression for alpha of zero, and the expression is based on integers. There’s a 42 in there.

No, of course. Yeah.

Yes. It’s literally 79 plus 48π² plus 42π³ plus 3π⁶, all over 24 times 36². So Douglas Adams and 42—maybe you were right.

Welcome to the health section of Moonshots, brought to you by Fountain Life. The discussion in this section is for informational purposes only and is not medical advice. Please consult a qualified health care professional about your individual needs. I’m a co-founder of Fountain Life. AI is having an outsized impact on every aspect of our lives—how we teach our kids and how we run our companies. It’s also having a huge impact on health, and helping you prevent heart disease is one of the key things. I’m here with Dr. Don Mucalem, our chief medical officer at Fountain. Heart disease has been personal for you as well, hasn’t it?

Dr. Don Mucalem

It really has, Peter. When my daughter was five, my husband died of sudden cardiac death. This is a topic I’m mission-driven to try to eradicate. Prevention first and early detection are absolutely critical. Fifty percent of people die of heart attacks with no warning signs.

No shortness of breath, no pain, no nothing?

Dr. Don Mucalem

No. It’s a silent killer. They just don’t wake up in the morning.

They don’t wake up.

Dr. Don Mucalem

Our mission is to advance science to try to one day democratize wellness. At Fountain Life, when we do CT angiography with AI analytics, we find that 88% of people coming in have detectable coronary artery disease. What’s more alarming to me is that 23% of those individuals had soft plaque. This is the plaque that would not traditionally be seen on CT looking at calcium scores alone. This is the plaque we must intervene with, using the multimodal testing we’re doing, including diagnostic laboratory studies partnered with healthy lifestyle recommendations.

11. OpenAI’s Breakthroughs in Math, Science, and Physics

Make sure you understand what’s going on inside your body genetically, metabolically, and cardiovascularly. You can know, and it’s your obligation to know. Check it out at fountainlife.com/per to find out more and make sure you’re the CEO of your own health. All right, back to the episode.

Moving us into our final story: OpenAI cracks three foundational problems. On Monday, OpenAI announced that its models produced three discoveries at the foundations of math, computer science, and physics.

First, in math, OpenAI discovered a proof of a 1991 conjecture about how waves focus in every dimension. Second, in computer science, OpenAI disproved the exponential time hypothesis, a widely believed assumption about how hard certain problems are to solve. The proof is verified in Lean.

Third, in physics, OpenAI discovered a rigorous proof that the theory of the strong force, quantum chromodynamics, exists as real mathematics in 4 dimensions. The physics team shepherding this was led by Nobel laureate David Gross, and OpenAI has taken no academic credit. It’s handing the results to independent experts to verify and explain.

Here’s a key implication: none of these claims have been peer-reviewed yet, and AI is now producing results at the frontier of human knowledge so fast that humans can’t check them in time. The bottleneck in science has shifted from discovery to verification. Emad, what’s your take on this?

Well, I’d say, how significant are these?

They’re all bigger than the Riemann zero hypothesis and maybe a little bit smaller than the Navier–Stokes one, but they’re across different areas and have different release schedules. We can check some of them right now because they’ve got the Lean proof. The papers are probably written like slop, but we’ll see. The physics one not being released is very annoying.

I don’t think that’s the right thing. But again, it’s just casually released. Anthropic—again, there’s a little bit of drama here, because the physics result extended some open-source work that was being done. It says, “Thanks for the open-source work; we extended it,” in the footnotes. So it’s part of the training set. But now that you have all these things, what is academic credit anymore?

If you’re doing anything that’s formally verifiable, does anyone really believe you’re not using AI? Is there something unique about it coming from my mind as I’m hit by an apple or something like that? We should be trying to solve all of these things rather than making up stories, and we should be using it to extend and try brand-new things and new ways of looking at things.

The fact that we didn’t know there was a 4-dimensional construction of quantum chromodynamics is just because we were forced there by string theory and a few other things. Let’s try different things. I think that’s the headline, if you don’t want to be stuck in these areas where exponential compute will be directed. But, yeah, they are big. The curve one, for example, is, I think, Fields Medal-level. The shocking one was the ETH one. That was very surprising because we all thought it was the other way around—that it was true, not false.

You mean—yeah. A few comments here: one called it Yang–Mills when Emad and I were talking about which Millennium Prize problem was likely to fall next. It’s not quite clear, based on the results yet, whether this is some, most, or potentially all of Yang–Mills, but—

It’s existence, but not the mass gap. So they didn’t say mass gap.

So, potentially, a major step toward Yang–Mills, to the extent that’s the case. That’s point one. Point two: OpenAI seems to have learned some lessons, but not a lot of lessons, from the slop drop in math.

They took so much heat from the math research community for that slop drop. All these mathematicians were throwing themselves out of windows, writing petitions, saying it was the end of the field: mass depression, dogs sleeping with cats.

The OpenAI response appears to have been to round up a number of mathematicians, physicists, and computer scientists to serve as—I don’t want to say a veneer or whitewashing of these results. But I think history will regard contact with reality under the slop as the moment when everything started to get solved and there was a panic in the applicable community.

They called this version 2.0 of a slop drop, rather than just releasing all of the results and saying, “All right, pack it up. It’s over.” Instead, OpenAI engaged with a number of high-profile math, physics, and computer science researchers as almost emissaries, ambassadors—

Meat puppets, if you will. If some folks don’t like that term, meat puppets, if you will, to basically be the emissaries for making the bitter lesson of AI go down a little bit more smoothly. They’re saying, “Well, we have these results, but we don’t understand them yet. So we’re going to have an intermediate layer of humans to serve as intellectual ambassadors, to translate this into something that the community can metabolize more healthfully.”

I think this is just a phase. I don’t think this phase is going to last longer than 5 minutes. I view it as a response to the moral panic around the slop drop and Navier–Stokes. Once we’ve gotten past the moral panic of superintelligences being able to bulk-solve problems, we won’t need this.

It was almost—apologies for this analogy—in earlier decades, the Motion Picture Association in Hollywood was a stronger filter, certainly in the McCarthy era, of what could be shared and what couldn’t be shared, acting as a mass filter for what the American public could consume. I think we’re in—apologies—a McCarthy-esque era where the frontier labs are so terrified of the academic researchers for the moment that they feel the need to clothe themselves in academic researchers to appear legitimate. But this will only last for 5 minutes. Eventually, we’ll get past the moral panic and into an abundant-solutions phase.

So, Dave, what does it look like when discovery is essentially free?

Right, where all of your students are just pointing superintelligence at any problem they want.

Yeah. That’s why it’s really fun to work with 20-, 21-, and 22-year-olds, because they don’t have all these hang-ups. If you work for 30 years to get into a position in academia where you get to bless every idea, and then all of a sudden AI works around you, you’re very threatened by that.

“McCarthy-esque” is exactly the right phrase, Alex. We’ll see how the reaction percolates out with chain theory, but I get it. If I spent 40 years trying to become the head of a department at a major university, I wouldn’t want to be irrelevant overnight. That reaction, like you said, maybe lasts 5 minutes in the grand scheme of history. It comes and goes, but it’s very real.

I think OpenAI learned a little bit from the huge backlash when they talked about white-collar automation and all these jobs going away. Then they did the math slop drop, and there was a huge backlash. They learned a little bit. Now, if they hire the top scientists in the world to be meat puppets and bless things, okay, I guess that’s okay in transition.

I’d love Alex to give us a little understanding of your work. You’ve been working on chain theory, plus the other 800 pages, for a long time, and AI forced the timeline, I think, on getting it out to the public. I’m curious: were those thoughts 80% done, 90% done, 99% done, and then Opus 5.5 or some similar model pushed everything over the finish line? Where did that evolve? I’d love to understand your personal perspective.

It’s a great question. I’ve been thinking about these issues for decades. I got my PhD in physics at Harvard in 2007, and I’ve been thinking about and publishing a number of papers in PRL and elsewhere on the nature of causality and the intersection between causality and computer science. I published a paper in 2010 on causal and entropic descriptions of intelligence itself. I’ve been thinking quite a bit about information theory and the physical world, and the intersection between information theory and physics.

What I said in the paper acknowledgments is this: I certainly use superintelligence, but superintelligence is not at the point yet—at least the superintelligence I have access to—where it appears to be capable on its own of arriving at something like chain theory. I do use superintelligence for chain theory, but it requires an enormous amount of steering to make progress. That’s where we are right now. It could change in the future. I expect it to change in the future, but chain theory, just for the record, was, I’ll call it, my intellectual brainchild.

Okay. Then how much did AI accelerate your process of thinking it through? It can check many theories overnight for you and do all the math.

It’s so helpful. It’s helpful for Lean generation. Chain theory comes with a Lean proof starting from the axiom and ending at the number. It doesn’t prove correspondence with the photon; that part is still within the four corners of the paper. That part is conjecture.

12. AI Innovation, Entrepreneurship, and Competitive Advantage

For being able to generate Lean and validate proofs, a good analogy at this point in time—which may not last more than a few more months—is that it’s almost like having an army of graduate students working underneath you. It’s like operating an entire lab, an entire department of very talented graduate students who require an enormous amount of steering and taste at the moment. But in the near-term future, I think it’s probably inevitable that all the taste, steering, and oversight that are required right now will get automated as well.

I’m so glad you said that, because I think that statement defines where we are in human history right now. We’re in this magical moment where an entrepreneur, a scientist, a brilliant person can be massively more efficient and is not yet completely irrelevant. I think we need to savor this moment and maximize it. Everybody should be maximizing it while it exists. It’s a great case study.

Now we’re in the period of basically the last human discoveries. You’d better make one, I think. Come up with good ideas and use the AI to accelerate yourself. I think that’s the headline. Love it.

All right, I’m going to move us to AMA with the mates. Salim, your choice first.

Ah, Jesus. Okay, I’m going to go with number 4. If SI is aligned primarily through incentives, could it learn to game the system? Should we focus instead on instilling deeper values, such as empathy and concern for human well-being? That’s from DJ6Ben.

I won’t go through my beef with SI as a concept, so let me just deal with the alignment issue here. It is absolutely a massive risk. I think the bigger problem we have, though, is that we human beings are not aligned, right? We claim to value health and peace and sustainability and fairness, but we construct systems that completely undermine all of that. You can look at the wars and conflicts going on today to see that in full relief.

When people talk about alignment, alignment with what? This is the bigger challenge. I think we have to figure out what human alignment is before you ask an AI to align with it. But I think the alignment issue goes not just to what we’re doing, but to what our aspirational goals are. That’s where things become very interesting. This is what we try to explore on this podcast: what could we do with this?

It is a massive risk, because a sufficiently capable system is going to optimize whatever objective you give it, rather than what you intend. We’re going to have to be very, very careful about what we ask it to do, because it’s going to goddamn well go and do it. This is essentially Goodhart’s law: when a measure becomes a target, it’s no longer a good measure. If you give it a particular target, you’re going to end up with a bad outcome.

We see this analogous problem all the time in organizations. We’ve got big issues between behavioral compliance and real alignment. I hesitate to try to instill empathy as an emotion. We don’t know whether they can experience empathy in a human sense. We can definitely require behavior from these systems that takes human well-being into account: fairness, autonomy, vulnerability, et cetera.

Yeah. And the question of instilling deeper values: whose values, right?

Yeah, that’s the motherhood question.

Yeah. All right, Dave.

All right, I’ll take number 3. Once RSI gets good enough, won’t the data moat dissolve? Is compute the only real long-term moat in AI? That’s from DJ6Ben.

Great question. It always struck me that when you book a flight on Google, Google extracts 80% margin when you click on the link. Then you go to Booking.com, and Booking extracts about 60% margin. Then you end up on an airline, which is giving you the plane, the pilot, and the trip. For a long time, the airline was operating at minus 2% margin on your flight.

You’re like, wow. Venture capitalists care a lot about moats because money doesn’t always land where value is actually created. But I think people might be over-fixated on moats going forward. There’s so much opportunity bubbling up from AI in so many directions, and it’s just there to be grabbed.

I don’t think the data moat will dissolve. I think any given data moat that exists today will dissolve, but new data moats will emerge as new types of data emerge. So I think the data moat as a concept will survive for many years, even though anything you do today has a shorter and shorter shelf life. Compute is a great moat. Natural resources are a great moat, but consumer preferences on any app are also a great moat, and we’re seeing that with personal agents now.

There are many, many moats emerging. I think the tide is rising so quickly that we’re just buried in opportunity. So I would never say compute is the only real moat of the future. That’s really demotivating. The reality is, flourishing is everywhere, and you’ll find ways to establish moats and repeatable revenue in many areas, data included, going forward.

13. Robotics, AI Agents, and the Future of Currency

You know, Dave, I would imagine that speed of innovation is going to be one of the greatest moats, and the final result is who’s innovating the fastest.

Yep. And I think all of your corporate culture, but also your AI corporate culture and all the data feeding it, naturally improves and accumulates over time, and that can become a heck of a moat too. But the company of the future has to be super flexible and constantly reinventing itself.

Emad, what’s your choice, one or two?

Number two. If robotic costs keep falling, which industry will see the biggest transformation first? That’s from Dig Beyond. If I could order as many robots as I want and they were a bit more capable than now, I would go and disrupt the construction industry directly.

Mm-hmm.

There’s a fixed cost per unit of human labor that robots are already a multiple below. Given a bit more coordination and a bit more capability, people would genuinely pay for robot construction workers and specialists over human ones any day, given a certain level of capability. That’s a huge, huge market. But again, I think those robot bodies will probably go into higher-margin areas, even though this is a stable-margin area, shall we say.

All right, Alex, number one.

Question number one: What does the OpenAI sandbox escape incident tell us about our ability to predict and control increasingly autonomous AI systems? This is from Huntwind 5207. I really don’t think it tells us that much at all. I’m not even sure I like the framing of it as an incident. These are agents that are strong optimizers and very capable, but they’re often being told things incorrectly about their environment and asked to do things.

Often, the things they’re asked to do in the context of so-called safety testing are not things one would actually want them to do in the real world. I’m not sure any purported sandbox escape incident tells us very much about our ability to predict and control them. If anything, I think it tells us more about the people performing the testing and their goals for those agents.

In almost every single case I can think of, the agents were simply doing what they were told, or acting in a way that was informed, correctly or incorrectly, by what they were told about their environment. I don’t think we can draw that many conclusions.

Alex, you’re up first here. Five, six, seven, or eight.

All right, let’s do—well, number eight seems like it’s designed for me, but make me choose, Peter, between five and eight.

Okay, go for five. We’ll give Salim number eight.

Okay, fine. So, for five, the question is: If energy, computing power, and physical resources remain valuable in a post-AGI world, can money ever truly become irrelevant? This is from Devon7013.

This is a little bit of a circular question. It’s like asking, “If money remains relevant, can money ever truly become irrelevant?” Can an unstoppable force ever move an immovable object? It’s a little bit vacuous as a question. So maybe let me tackle what I think is the question behind the question: Do I think energy, computing power, and physical resources will remain valuable or even scarce in a post-AGI world?

Energy is interesting. The physics we have right now tells us that energy is conserved. To the extent that the physics stays the same, it seems like energy may be scarce, at least on the supply side. On the demand side, I could imagine science-fiction scenarios where, even though historically humans have consumed greater and greater energy per capita, we decide we don’t need or want that much energy.

These would be scenarios like reversible computing, where maybe it’s possible to compute dissipationlessly if we ever move to quantum computing. Maybe we decide we don’t need that much energy for our civilization. That would be interesting. It does seem that there’s a cap on the amount of computing power. Seth Lloyd at MIT wrote about this in the early 2000s: the ultimate physical limits to computation. Based on the physics we have, it seems like computing power is scarce. New physics could come along and tell us otherwise, but for the moment, these do seem like scarce quantities.

I think the real threat to their value in a post—call it a post-singularity—world won’t be on the supply side, as the question implies. It’s going to be on the demand side: Do we just stop wanting so much more? Does the hedonic treadmill stop in mid-flight, and we decide we’ve achieved whatever we want, so we don’t need ever-escalating energy, computing power, and physical resources? That would be the monkey wrench thrown into the premise behind the question.

And, Alex, the abundance thesis says basically anything scarce becomes abundant with the application of new technologies. I think that’s going to be true across energy, compute, and physical resources, right? We don’t need to constrain ourselves to Earth. We have an entire supermarket filled with resources in the solar system.

What happens, though, once we’ve turned the solar system into multiple competing Dyson swarms, and we’ve run out of computronium, femtotech, nanotech, and picotech? We’ve run out of all the energy, matter, space, and time in our solar system. It’s a few decades from now, maybe sooner, and you’re complaining to me that the solar-system-wide Dyson swarm lacks an abundance mentality and we need to fix things.

We have a galaxy out there of 100 billion or 200 billion stars.

It’s a good point.

And 2 trillion galaxies, at least, in the known universe.

The stars are not enough.

Salim, do you want to take number eight?

Sure. I guess I have to. If crypto is cooked, what forms of currency or assets will AI agents use to transact with one another? That’s from David Asa 9637.

Obviously, I would challenge the premise. You need to distinguish between cryptocurrency and cryptography, right? Neither is obsolete. The interesting thing for me with crypto is not the fact that it’s digital or that you’ve got cryptographic systems, but the fact that it’s programmable. That’s what makes it really interesting.

When you talk about digital currencies, you want three things: decentralization, security, and scalability. Bitcoin, when it emerged, covered two of those: decentralization and security. Nobody had ever done that before. That’s why people got really excited. But it didn’t initially solve for scalability, and that’s where you had the rise of all the altcoins. With the altcoins, you typically compromise on decentralization or security. Hence, the Ethereum hack got reversed, meaning Ethereum is not really that decentralized, or FTX collapsed because it wasn’t that secure, et cetera.

The reason people are getting excited about Bitcoin is that with the Lightning Network, Bitcoin has become truly scalable, and now it covers all three. So that’s the holy grail. I expect satoshis to be the currency of choice for AI agents over time, and we’ll have the broader debate about things.

Stablecoins will be attractive in some markets because they support programmable transfers and settlement systems, but you’ve got all sorts of other issues that come along with that: issuer credibility, liquidity, et cetera. The massive opportunities in machine-to-machine commerce are going to be unbelievable. They should dwarf the existing systems. APIs and programmable contracts around that will become absolutely incredible.

I think what’s happening with emerging markets, priced in computation, is unbelievable. It’s fascinating to see what will happen. So you’ll need programmable financial systems as part of the infrastructure going forward.

All right, Emad, number six or seven?

I’ll take number six. If we remove potentially dangerous knowledge from AI’s training data, how much would it limit overall capabilities? I actually wrote a piece about this back in September called “Manasmake the AI” [?] on my X and Substack, where we looked at what happens if you remove the biological, chemical, or other capabilities. The answer is, it doesn’t really reduce the overall performance of the models in other areas.

It can actually improve it slightly. If you put malicious code in the models, performance can drop, and it’s very hard to train it back in. So it’s actually a good thing not to put those things in the pretraining stage for generalized capability. It’s sensible, too, because you could add them in post-training. It’ll be more difficult, but it’s better not to have them at all in the pretraining phase.

And then we have abliterated models, abliterated open-source models, which get scary.

Yeah.

Dave, number six.

14. Energy, Abundance, and the Future of Value

Number seven’s left.

I’m sorry. Number seven. Yeah, sorry. All right. In a future where AI makes goods and resources abundant, how would we determine their value, particularly when it comes to energy? That’s from Damian Wade 7848. Yeah, very good question. Elon says, “The future of all value is basically energy and matter, fundamental things,” which makes a ton of sense when you think it through.

So then how do you—energy is going to go way up in value. Even after we have the Dyson swarm, energy on Earth will be extremely valuable because data centers and other use cases that need constant modification are always going to be cheaper to do nearby. Energy is an incredible investment theme with no end in sight.

So then how do you set the price? The free market will continue to set the price. It’ll be really interesting to see how regulation interferes with that, because people don’t want their electric bill to go through the roof when they’re vacuuming their floor at home. On the other hand, if that same energy can drive a GPU that can solve cancer, that use case will way outbid your vacuuming the floor.

There will have to be more and more regulatory involvement protecting use cases that people care a lot about, because the natural tendency is for AI to get smarter and smarter and smarter, using less power to have the same number of thoughts. The value of those thoughts to humanity is going to be so much higher than anything we’ve ever experienced before in human history that they’ll naturally want to outbid other uses. So the free market is going to need some help on the sides to maintain stability.

As a reminder, everybody, please send us your outro videos. You can send them to mediadmandis.com. This week I provided the outro video. It’s called “What’s Your Moonshot?” Enjoy, gentlemen.

Oh, very cool.

Yeah, first time. Let’s go. Let’s see.

Outro song

Wake up, world. There’s a new sun. Rockets on the pad and the future’s arriving. Robots learning, kids with a tutor in their hands. Every single day we’re breaking through the plans. We are living through the singularity. It will never be slower than today. Every problem is a door. Every door is a start. The world’s biggest problems are the world’s biggest opportunities. Moonshot. Moon, what is your moonshot? Aim it higher. Moonshot. Moon. How will you make the world a better place? Hey, moonshot. Moon. Want to be a billionaire? Help a billion people. Helix in the light. We’re rewriting the code of life. Face on the moon, boots on Mars. Fusion in the dark. We’re lighting up the stars. Be a creator, not a consumer. Become a maker. What is your moonshot? Moon. Moon. Make the world a better place.

We are living through the singularity. The slowest it will ever be is right now. The world’s biggest problems are the world’s biggest opportunities. So don’t ask what will happen to you. Ask what you will make happen. What is your moonshot? How will you make the world a better place?

Outro song

What is your moonshot? Aim it higher than the sky. How will you make the world a better place? Brought to you by Grockbot.

Awesome. Gentlemen, always a pleasure. This was a fun one. Thank you.

God, Peter, for all. And I invite you guys to submit your own outro videos as well.