我们如何走到了史上最大IPO竞赛｜SpaceX、Anthropic与OpenAI

I'm Kevin Roose, a tech columnist at The New York Times.

Newton from Platformer.

And this is Hard Fork.

This week, SpaceX, Anthropic, and OpenAI are all heading to the public markets, but what do their IPOs mean? Then, author Kevin Hartnett is here to talk about why some mathematicians are sounding the alarm about the use of AI in their field. And finally, some hot ChatGPT.

Well, Casey, the big news this week is that the AI IPO race is heating up. It's hot IPO summer.

It really does seem like it is going to be a hot IPO summer, Kevin. I am told that we are on track to see what might be the 3 biggest IPOs of all time.

Yes. SpaceX is getting ready to go public, maybe as soon as next week. Then, just this week, Anthropic filed a confidential S-1 with the SEC, noting that it intends to go public. That is the first step in the process. There are reports that OpenAI is going to file its S-1 soon as well.

This is obviously long-awaited. People have been wondering when these giant private companies were going to go public, and now it seems like they're all racing to do it as quickly as they can and potentially beat each other to market.

Yeah, and the consequences are really important, and we're going to get into them. But before we do that, truly, there has never been a better or more important time to do our disclosures.

Yes. I work for The New York Times, which is suing OpenAI, Microsoft, and Perplexity.

And my fiancé works at Anthropic.

Okay. So those are our extra-special AI disclosures this week. Now, Casey, let's talk about these IPOs.

Let's do it.

Maybe we should start with SpaceX. What is going on with the SpaceX IPO? What are we expecting, and what does it mean?

Yeah, as you noted, Kevin, SpaceX is just the furthest along right now. They seem like they're getting very close to the finish line, and they just have some staggeringly ambitious plans. They plan to sell their shares at $135 a piece, which would raise $75 billion. That would make it the largest IPO of all time. It would also value the company at between $1.75 trillion and $2 trillion, which would instantly make it among the very biggest companies in the world.

Yeah, those are crazy numbers. Those are just numbers that we have not seen before in the history of capitalism. We should also remind people that when we say SpaceX, we are talking about the combined Frankenstein conglomerate. Tell us what is actually in SpaceX.

They make rockets. They make Starlink. They also, as of fairly recently, own xAI and X, the social network. And so that is all part of this Elon Musk conglomerate that is going public.

Hobby Lobby? Is Hobby Lobby part of it as well?

Not yet, but don't give their business development team any ideas.

Okay, fair enough.

We're going to do Hobby Lobby in space.

So this giant conglomerate is being positioned in the market as a way for people to invest in AI. Obviously, it does have an AI company inside of SpaceX, but I would say investors are more excited about the space part of it, which is the most developed. They make stuff that people actually use.

Yeah, I mean, look, there are 2 great businesses in here, right? One is a reusable rocket business that delivers satellites into space. It's very hard to build that kind of company, right? SpaceX just has an incredible moat. There aren't that many competitors to it. We saw Blue Origin, one of its biggest competitors, lose a rocket on the launchpad just over the past week or so. That's what makes that an incredible business.

And then Starlink is just on fire, right? They're using their ability to deliver satellites into space to also create a really powerful global internet access system that is just growing like wildfire. So there are 2 amazing businesses in there, and then they also have 2 terrible businesses called xAI and X. It'll be really interesting to see what the interplay of the good businesses and the bad businesses will be in the months to come.

Yeah, sometimes you just have to take the good with the bad, especially when they're all packaged in the same stock ticker.

What was interesting is that you didn't have to take the good with the bad. You could have just had the good. Up until very recently, SpaceX was just SpaceX and Starlink. It was just a pure good business, but then it seemed like Elon Musk decided he needed to hide his losses somewhere, and so they inherited the 2 worst companies he owns.

Well, what's interesting about that is that one of those companies, xAI, appears to be pivoting. So they are now renting out compute that they originally built for themselves to Anthropic, another one of these companies that's going to IPO this year. They are positioning themselves as something like a space company with a kind of AI neocloud business attached to it and a social network that is going to become the everything app.

All of it is a little mysterious, but basically, this is the long-awaited time when all of Elon Musk's Voltron-like companies are going to take a stab at going public together.

Okay, so that is the SpaceX IPO. Now let's talk about Anthropic. They have filed their confidential S-1. They are also expected to go public at something over $1 trillion.

What? I'm just shaking my head at the insanity of that, considering what their ARR was 1 year ago today.

Yes. This one is just wild to me. I was thinking the other day about the first time I ever visited this company. It was in 2023, 3 years ago. And when I tell you that this company was not only ambivalent about making money but seemed to actively resist the idea of making money, they were a small group of very earnest, AI-safety-obsessed people who were tinkering with and building models for unspecified purposes.

I remember your story about it, and it was about how glum and strange the office was, which you might expect for a bunch of very safety-focused people who hadn't even decided whether they really wanted to start making a product yet. But boy, howdy, did they make a product and decide that they actually liked making money and wanted to make a lot of it.

Now they're going to be one of the largest IPOs of all time, just 3 years after I was hanging out with them in a little Jackson Square walk-up office.

No, and in January 2025, this company had an annualized revenue run rate of about $1 billion. Recently, they've said it's 50. Who knows what it's going to be by the time they IPO? But that growth is unprecedented in Silicon Valley.

Yep. OpenAI—we don't know anything about their upcoming IPO except that they have said that they plan to do it. They may file their paperwork with the SEC as soon as this week to start that process. People also expect this to be a big, gigantic IPO.

All of this taken together, I think there are a few threads to pull on here. One of them is: What is this going to do to San Francisco? Here we have 2—call it 2.5 companies, because SpaceX has some headquarters and offices in Texas, Southern California, and other places—going public in the same year, based in San Francisco, minting hundreds, if not thousands, of new millionaires, decamillionaires, and centimillionaires. What will that do to the city's tech scene, the local real estate market, et cetera, et cetera? What are your thoughts on that?

I mean, my fear, Kevin, is that we are about to see a massive increase in inequality in a town that already had really significant inequality, right? I just worry that it's going to feel even worse.

Where I'm already starting to notice it is when I talk to my friends who have really good jobs paying maybe even mid-six figures. They're looking at what they're reading about the folks who got in early at OpenAI or Anthropic, and the comparison is not feeling good. They're starting to wonder, "What does this mean for me? Am I going to be able to lead the life that I wanted?"

It's had me thinking a lot about when I got to San Francisco in 2010. There was a sense of abundance here. There was a sense that anyone could do a startup and anyone could have the life that they wanted. It was true for many, many tens of thousands of people.

I feel like we're almost swinging back to this scarcity mentality here, which is, "Well, if you didn't make it in at one of these 2 companies, your future is in doubt." I don't know how true that is, but I can tell you that that is the anxiety that I'm hearing.

Totally. It's hard to feel much sympathy for these people who are very well-paid engineers and tech workers, looking at their slightly richer peers and thinking, "I got to get some of that." But I think this is a real status-anxiety moment in San Francisco and Silicon Valley, where even the people who thought they had made it in the world of tech are now looking at these people who have joined these insanely fast-growing companies, wanting to get in on that somehow but also thinking it might be too late.

People who I don't think were feeling precarious a year or 2 ago are now. I think that's just a really interesting social marker.

And just to say again, I think what you want in a society is for opportunity to be spread broadly.

And for everyone to feel like they have a chance to lead the life that they want. And so when you move into a world where there is what essentially amounts to a handful of lottery winners, and those are the only people who truly get to live the lives that they want, that just causes massive social instability and all sorts of other problems. So, yeah, I have a knot in the pit of my stomach about this.

Yep. I do too. And one thing that is making it slightly better is that I did see this coming, and I do get to crow about my correct prediction from our prediction episode back at the end of last year, which was that 2026 was the last year to buy a house in San Francisco. I’ll make one more quick AI bubble prediction, which is that 2026 is the last year to buy a house in San Francisco. Because I think this coming year, we will see the IPOs of at least 1 major AI company.

Anthropic is reportedly planning an IPO for next year. We may also see OpenAI go public. SpaceX may go public. Stripe may go public. There are these other highly valued private companies that may go public. When that happens, their employees get rich and go buy houses, which now appears to be true. The real estate market is going nuts. Houses are going for many multiples of their asking price. There was a story this week. Did you see this one about the San Francisco homes that are on sale asking for Anthropic or OpenAI stock instead of cash?

I think those people are smart. There’s a very real chance that that stock will appreciate in value even faster than your house. So I say, shoot your shot.

Yes. So there are 2 other things about these IPOs that I want to discuss with you. One of them is the effect it’s going to have on philanthropy. One of the strange characteristics of these particular companies is that many of their employees—and I would say this especially applies to Anthropic, but there’s also a piece of this at OpenAI—are committed to effective altruism and other similar philanthropic movements that basically teach you that, if you want to make a maximum impact on the world, you should make a bunch of money and then give it away.

Nonprofits, philanthropies, and donor advisory networks in and around San Francisco are now starting to ask, “What if we just have this influx of new philanthropic capital coming from these IPOs?” Our mutual friend Nan Ransohoff wrote a great post about this the other day calling it “the third wave of philanthropy,” basically where you have tens of billions or possibly hundreds of billions of dollars flooding into these charitable movements and causes. How does that affect what gets funded? Are there going to be new institutions that need to be built to absorb all of this philanthropic capital? I think this is something that people outside San Francisco don’t quite understand: how much money is going to be flowing into these philanthropies over the next couple of years.

Yeah, I also really loved Nan’s post and had a chance to talk with her about it in person recently, and it was just so fascinating to hear about the sheer volume of philanthropic capital that we’re expecting to emerge as a result of these IPOs and how little infrastructure there is to absorb it. Something that I think some people might not know about Anthropic is that, from the start, they have told people as they come on board, “If you pledge a percentage of your equity to philanthropy, we will actually match it.” And so this is just a program that is dramatically amplifying the amount of philanthropic capital that is about to become available.

Well, it was not just matching. So there are 2 things that Anthropic did that speak to their historical ties to effective altruism and that world of philanthropy. One is that all 8 co-founders pledged to give at least 80% of their wealth to charity. So right there off the top, we are talking about hundreds of billions of dollars potentially that are being earmarked for charity just by the 8 co-founders of Anthropic.

Then you have the stock-matching program, which, as you alluded to, not only offered to match employees who pledged a certain percentage of their stock to charity share for share, but matched them 3 to 1 in the case of some early employees. And when you actually lay out the numbers, as Nan did in her post, the amount of money is just staggering. We are going to see something bigger than the Gates Foundation every year, potentially, for the next few years. And it’s going to go to some stuff that will seem to the outside world fairly weird, right?

No, like what?

The joke going around the AI circles is, “Is this going to be a great year for shrimp welfare?” Because shrimp welfare, for arcane reasons that are probably not worth going into here, has become a sort of half-joking pet cause of the effective altruists. But it’s a great year to be a shrimp. It’s probably also a great year to be working on global health and pandemic prevention, AI safety, and all these other cause areas that are very closely affiliated with effective altruism.

Yeah, it makes me glad that we shredded the social safety net and made all those reductions in pandemic preparedness. So now the San Francisco billionaires can step in and rebuild it hand by hand.

Exactly. I want to bring up 1 other thing about these IPOs and ask for your opinion about it. The thing that makes me nervous about these IPOs is not that they could go sideways, people could lose money, or these companies are very speculative. All of that is true. What worries me is the safety angle here, because I am of the belief that these AI systems are getting more powerful, that those increased capabilities also bring increased risks, and I just know that many of these AI companies—OpenAI and Anthropic specifically—were started by people who were worried about safety and specifically worried about the ability of a for-profit corporation to develop AI safely.

At OpenAI, you’ve had this governance struggle that resulted in, among other things, Sam Altman being fired and rehired. At Anthropic, they have made themselves a public benefit corporation to try to lessen the influence of shareholder capital and fiduciary duty on their ability to make decisions related to safety. But I think all of that just gets much harder in a world where these are publicly traded companies and big investors, index-fund holders, and retirees are invested in them.

It was already going to be hard to slow down or maybe refuse to release something that was dangerous because of the enormous sums of money that these companies have raised, but it’s going to be a lot harder when the public markets are also pressuring these companies to raise and go as fast as they can.

Does it matter that both OpenAI and Anthropic are structured as public benefit corporations and are allowed to make social commitments that a traditional company like SpaceX, for example, will not be beholden to?

I’m of a couple minds on this. I think it’s probably better that they are public benefit corporations than not, because a public benefit corporation is this legal designation that allows you to take into account things like, “Are we being socially responsible?” It doesn’t mean investors can’t sue you as easily for breaching your fiduciary duty if you do something that is counter to their interests as shareholders.

But they’re still corporations, and they still exist at the pleasure of shareholders. There are certain concessions they can make to social issues and impact, but when it comes down to it, when the rubber meets the road and one of these companies develops a model that is truly dangerous, they’re now going to need to weigh not just what they think the right thing to do is, or what their private investors think they should do, or what their employees think they should do, but they’re also going to have the public markets breathing down their necks. They’re going to have activist investors and things like that. So I’m just nervous about the structure that is now going to grow up around these companies and push them in the direction of acceleration.

I think that’s fair, and I may be coping, but when I think about the possibility of a lab developing a really dangerous model and saying, “Well, due to shareholder pressure, we’re just going to put it out there,” you have to remember that shareholder pressure can work in the other way, too. Because if you put out a model that can create a new bioweapon, you’re probably going to get sued for securities fraud by a shareholder that said, “I trusted you to only release safe models.” So I do think that there are going to be some positive pressures here that hopefully keep them from doing anything too silly.

Yeah, how do you think this will impact the average person who’s not an investor or a shareholder in these companies, who may just want this stuff to be developed well and safely?

Mm-hmm. I mean, I think on balance, you’re right that because these are corporations, we do have to worry that capitalist pressures will lead them to cutting corners and doing things that are unsafe. So that is, I think, the right concern to have and to keep your eye on. On the other hand, I could also see an argument that when your company goes public, you are introducing more democratic oversight and governance into it, right? These folks will now have to report their earnings.

They will have to give us information about their financials. They will have to make certain disclosures as their products come out and as their company changes. Shareholders will be able to maybe vote on certain things. I think these are all good things because right now we have almost no levers whatsoever that people can pull other than trying to prevent a data center from being built in their backyard. So maybe these give folks some new ones.

Yeah, and I think there's been a lot of hand-wringing over these indexes and exchanges that have changed their rules. the Nasdaq 100 and the S&P have already, or are considering, loosening their rules around these so-called seasoning periods. Basically, it used to be that if you were a brand-new public company that had just IPOed, you could not be included on these major stock indexes because they wanted to see whether you were stable enough to become part of the basket of blue chips that people invest in when they buy an index fund.

Now, partially because of these looming IPOs, those rules have been relaxed, so these companies are not going to have to wait 3 months, 6 months, or a year anymore to be included on these indexes. Some people have said, “Well, that sounds bad, and we're exposing retail investors to these volatile and risky stocks.” I'm not that worried about it. I think investors want exposure to these companies. I know several people in San Francisco who have been devising these crazy, harebrained schemes to get pre-IPO stock in one of these companies.

I think that letting the public benefit from the upside of the AI boom is going to do more help than harm, but I could be wrong if all of this goes up in a conflagration and people lose their shirts.

Yeah, absolutely. Kevin and I are not financial advisors.

I am actually a certified financial advisor.

I am not a financial advisor. But I do think that if you're a retail investor and you believe in this stuff, you should have the ability to make that bet. Because we live in a country where you can bet on Bitcoin in an exchange-traded fund. You can go on a prediction market as a member of the military and bet on an operation that you're a part of. So in a world where those are restrictions on financial gambling, if you want to buy a share of OpenAI, I say Godspeed.

Yeah, I think part of the icky feeling that people are having about the AI industry now is that so much wealth is being concentrated in so few private companies and so few hands. In an optimistic scenario where these IPOs go off without a hitch and these companies keep growing at hyperscale, I think maybe having the benefits shared a little more broadly through things like index funds could be good for people's feelings, like, “Oh, there's something for me in this.”

I'm going to go further and say it's not just good; I'm going to say it is necessary. We cannot have a very small handful of companies that are growing this quickly, concentrating wealth and power that much into so few hands. It has to be shared more broadly than that. And while an IPO is a very small step in that direction, I do think it is a necessary one.

Okay, well, that is enough about the IPOs for this week. We will continue to cover these IPOs and everything that comes out of them, including possible space data centers. I don't know. I'm excited to learn more about those.

Yeah.

I will say, when it comes to Starlink, I was not a believer. And then I went on my first Starlink-equipped airplane last week. And Casey, this is going to be the biggest company in the world.

It's very—

When people get a taste—

Yeah.

—of 200-plus megabits in the air on a plane—

You're never going back.

Yeah, you can actually watch YouTube if you have Starlink on your plane.

Watch so many YouTube videos.

And the deal that they struck, I'm told, is that, at least with United, they're like, “We'll put Starlink on your plane, and we will charge United for that, but you can't charge your passengers.” So everyone's experience of Starlink is that it is a free miracle being delivered to them in their airplane seat.

Which is not a bad marketing strategy.

Not a bad marketing strategy.

Yeah.

Well, Casey, get out your TI-83 graphing calculator, because today we're going to talk about math.

Can I play Snake on it, or do we actually have to talk?

No, we have to talk because today we are going to talk about what is going on with AI and math. Now, this is a subject that we have talked about before on this show, but there's actually been a lot happening just over the past couple of weeks.

So, 2 weeks ago, on May 20, OpenAI announced that one of its models had reached this big mathematical milestone. Basically, it had disproved this long-standing geometry conjecture by identifying a new way of thinking about this famous math problem, one of these Erdős problems that no human mathematician had considered before. That was considered a very big deal in the world of mathematics.

At the same time, there's also this backlash brewing in mathematics to the use of AI. Just this week, a group of mathematicians have been passing around and signing something called the Leiden Declaration, which is basically an open letter about the use of AI in mathematics from people who are concerned that they're eroding the human foundations of this sort of academic mathematical discipline.

Yeah, and when I asked a mathematician why, they said, “y = mx + b.”

Okay, very good.

That, of course, is the classic slope-intercept formula for a straight line.

Which model did you use to look that up?

I'll tell you later.

So we just thought it was a really good idea to check in on the state of AI and math. And to help us make sense of what is going on right now, we have turned to one of the best guests I can imagine for this subject.

Kevin Hartnett is a journalist. He has covered math and computer science for many publications, including most recently as a senior reporter for Quanta Magazine. He's also the author of a book that comes out next week called The Proof in the Code, which is sort of about this formal math language called Lean and how it's transforming math and AI. Today he works as the editorial lead at Cursor, the AI coding platform. I just thought he would have a really good view of the situation.

Also, we just like to bring on people named Kevin because they tend to be really smart. Okay, let's bring in Kevin Hartnett.

Thanks, Kevin. Great to be here.

Well, Kevin, we've brought you here today to talk to you about AI and math. You just wrote a book called The Proof in the Code, which is, I'll just say it, the most interesting book I've ever read about math.

Thanks.

It's a short list. It narrowly edged out One Fish, Two Fish, Red Fish, Blue Fish.

So when we last checked on the field of mathematics and AI, it was last summer, and 3 of the big AI labs—Google DeepMind, OpenAI, and Harmonic—had all reported that their math models had achieved a gold-medal score at the International Math Olympiad. That was something that people had been saying for years would be impossible, or would take many, many decades, for computers to be able to do. But their AI models did it last summer. What has been happening in the field of AI and math since then?

Yeah, virtually everything. The IMO had been a benchmark for a long time. In my book, there's a whole chapter about the previous year's IMO, the 2024 IMO, where Google DeepMind got a silver-medal score, and that was considered a small watershed. Then, as you said, last year 3 labs got this gold-medal-level score, which had really been the benchmark that had been set out.

At that point, AI was still just doing essentially high school math—the hardest high school math in the world, but still just high school math. And I think for people who never went beyond high school math, it's hard to really appreciate how far that is from the frontier of research math. It's forever far. That's barely even wading into the field. It's 0% of the way to the frontier. So it was a proof of concept, maybe, but it certainly didn't mean much in terms of whether these models could actually do research.

That's very hard for me to hear as someone who was not that good at high school math, I have to say. I'm feeling a little defensive, but I believe you. It's making me defensive.

Why were the labs so focused on the IMO and on math in general? Was it because that was just a very hard challenge that they liked, or was it because they thought that being able to do math at a high level would enable their models to do other important things?

Well, it's definitely both. I think that the challenge—this IMO Grand Challenge, which was the name that a researcher at Microsoft Research gave to it—was really about, “Can we just create models that can do amazing math research?” It was really research for research's sake.

At a certain point, the labs and these startups you mentioned adopted that challenge themselves, and their motivations were a little different. There's very much this belief that if you can teach a model to reason about math problems and solve math problems, it would be much better at other things.

I always think about this statement that my high school math teacher would give when people would ask, “Why are we learning this? What's the point of this?” They would either say, “So you can balance your checkbook,” or, “To teach you how to think,” right? It's like one or the other, and the “to teach you how to think” is really the point here.

It's like if you can reason through a math problem and think logically, then you can apply that type of skill to all the parts of your life. And I think the labs believe that if you can teach a model to reason through math problems, it's going to be able to do all these other things that are probably much more commercially valuable as well.

So I'm having a flashback to when we first started to talk about AI and math, and the knock on these models was that they were actually quite terrible at it, right? If you would try to get them to do basic addition or multiplication, they would utterly fail. So, Kevin, sketch out for us a little bit what the labs did to navigate through that problem and get to a place where they could credibly try to advance the frontier of the science.

When ChatGPT came out in November 2022, mathematicians were passing around all these, “Ha-ha, look at this stupid model telling me that there are only finitely many primes when we all know there are infinitely many primes,” and “2 + 2 is 5.” Basically, that kind of thing.

I think, essentially, the models got better. There's definitely an element of reinforcement learning on math problems that makes these models better—RL on math—but I think it's also just the general improvement in these models that we all experience in the ways we use them that has led to these incredible reasoning tasks that they've become capable of.

So, let's talk about one of the areas where it seems like we've seen some creativity in math with AI lately, which is these Erdős problems. Can you tell us who this Erdős was and why he left us with so many problems?

Yes. Paul Erdős was a really colorful mathematician. He was essentially the Bob Dylan of math, in that he spent his life on the road. He died at age 83 at a math conference. He slept on mathematicians' couches his whole life, and as he went around, he compiled lists of problems that he thought were interesting. Either he would find them in the wild, or he invented many of them, and he created this Erdős list.

He endowed them with tiny little rewards, like $20 for solving this problem or $500 for solving this problem. That fund still exists after his death, and these things are paid out. I actually don't know if the LLMs have received the money or who gets the money when they do it, but anyway—

Don't give them the money. They don't need the money.

Marcus du Sautoy

If a human solves them, they get the money. Paul Erdős collected over 1,000 problems—1,200 problems—that he thought were interesting, and he just kind of left them out there.

The AI labs and these startups have looked for benchmarks, mountains they can climb, things they can do to prove that their models work well. The International Mathematical Olympiad was one of the most prominent ones. They got the gold medal there, and they needed to move on. They moved on to the Putnam exam, which is the premier college math competition, and started to do quite well there.

Then they started looking for new targets. These Erdős problems were sitting out there—1,200 or so problems that, for the most part, mathematicians had never looked at or had given very little attention to. So they essentially set their models to work on all of them: See what you can do. The models would cook up answers to them, and through the beginning of this year, we would see one announcement after another on Twitter: “Solved Erdős problem 737. Solved Erdős problem 63.”

And I think mathematicians viewed those hype-y announcements very differently from the energy behind the announcements themselves.

Mathematicians viewed them?

They said, “Something isn't adding up here.”

Thank you, Casey.

I'm going to have to think of some puns. I'm going to get one pun out before we finish this podcast. That's my goal.

So, mathematicians: There are just a lot of unsolved math problems in the world. Just because a math problem is unsolved, has been around for decades, and was dreamed up by a famous mathematician does not make it an important problem.

An important problem is one that the field, in its collective wisdom, determines either has an answer that will really change how we view math or, more significantly, requires methods that are going to remake the field. It's going to create important new math. These Erdős problems were just not viewed that way. They're kind of like sophisticated riddles—sophisticated arithmetic, numerical riddles.

These are like the Sudoku of math. It's like the Wordle of math.

It is like the Wordle of math. I think that's a fair statement. And so, anyway, mathematicians didn't spend a lot of time looking at them. I think the conventional wisdom was that these Erdős problems were toy problems, not serious problems.

But that's not true of all of them. About a month ago, OpenAI came out with this big new result. They've solved what many people think of as one of the most important Erdős problems, this thing called the unit-distance conjecture.

What was important about the unit-distance problem is that it was a problem a lot of people had looked at. So you couldn't just say, “No humans have really tried to solve this.” People had looked at it; they just hadn't solved it. The methods underneath it were very sophisticated and surprising. It was not just a clever cobbling together of obvious techniques.

And the result itself was just so good. Pretty unanimously, people agreed that it could be published in the Annals of Mathematics, the top journal of math. In a way, over the last year, there has been this shifting of the goalposts: AI did this, but it can't do that. Oh, it did that. Nope, it still can't do this.

There's still some shifting going on. There's still room to shift. They haven't solved a Millennium Prize Problem. But this proof of the unit-distance conjecture really said AI can do absolutely top-tier research.

I've been following the story of AI in math in part through people like Terence Tao, who is widely considered the greatest living mathematician. He has been experimenting, writing, and making videos about his experiments with these AI programs for use in frontier math research for a number of years now.

When he started working with them, he was like, “Oh, these aren't that helpful. Maybe they're a mediocre grad student who you'd have assisting you.” More recently, he seems to be saying, “This is revolutionary for the kind of frontier math research that I and other professional mathematicians do.”

He recently made a video with OpenAI talking about how he can now try a bunch of crazy ideas and experiments because the cognitive friction of using these models has gone down. You can have an idea, give it to the model, and say, “Go test this a bunch of different ways and figure out if there's anything there.”

Is that a widely held view among mathematicians, or is he just on the extreme AI-pilled end of the field?

Terry is an extremely interesting figure in this because he is so important, as you said. I have a whole chapter in my book about some of the early work that he did with AI, this thing he called equational theories. He's always been interested in different ways of doing math. He's very intensely collaborative, and he's interested in new ways of working.

Terry is representative of one of 3 attitudes toward AI and math right now. A couple of weeks ago, I was at the Institute for Advanced Study in Princeton, New Jersey, which is like the citadel of modern math—the biggest, densest collection of great mathematical minds in the world lives and works there.

In a single afternoon, I was walking around the campus, and I had 2 strikingly different experiences. I ran into 2 40-year-old mathematicians, both at the top of the field. One of them told me he had just tried to do some math with Gemini. He said, “That stupid thing told me that XYZ thing, which we know is wrong, is true.” Then he closed it and went back to doing math the other way.

An hour later, a guy who, on paper, looks a lot like the first guy said, “I think in 2 years, AI is going to put mathematicians out of business because it's just going to be strictly better than us at all of it, and we won't need mathematicians anymore.”

Those are the 2 poles. Terry is squarely in the middle. That OpenAI video you described, Kevin, is the “jetpack for your thoughts” view of AI. It's the Iron Man suit that will allow me to do more, better, and bigger things than I could before. He is clearly the leading figure of that point of view.

I don't really know how it would break down if you took a poll of mathematicians right now. I would guess that “it's going to put us out of business” would finish last. I would guess Terry's view would finish first, and I think “it's good for nothing” might actually have won a year ago, but it's definitely falling in the polls.

Let's talk about some potential paths to a world where maybe more people are agreeing with the person who thinks that AI threatens the job of a mathematician. Kevin, do you want to talk a little bit about this letter that these mathematicians put together?

I do. This was fascinating, and this is actually the reason we wanted to have you on today: to talk about this response, this declaration, the Leiden Declaration on Artificial Intelligence and Mathematics.

And this is basically a declaration signed by something like 800 mathematicians so far, which I would characterize as a very worried document. They are upset about the use of AI in their field and what they consider the irresponsible or reckless use of it. They make a bunch of statements about how these technologies are producing plausible but unreliable or even incorrect arguments, which are hard to distinguish from correct proofs. So help us understand this declaration and the perspectives of the mathematicians who are putting their names on it.

Yeah, I think you are right that it reflects a kind of deep concern and worry for the future of the field. Mathematicians, I would think, are deeply worried as a population or a community that largely was able to run itself and self-regulate for centuries. Now there's this massive exogenous force that's shaking it, and they want to be able to defend the field. They want to set up some kind of guardrails and also say, “This is our field. You don't get to tell us what's important and how it runs.”

This document you talked about, Kevin, is, I think, an effort to try and start making that kind of statement.

Well, I think they've realized that there's strength in numbers.

Oh, my God.

So here's what I want to know.

The hits keep coming.

Here's what I want to know: What exactly is so threatening to them about other people using ChatGPT to do math proofs?

Well, let me actually step back 1 second. I think there are 2 things the letter and document are trying to do. 1 is the same thing that all fields are trying to do, which is ask, “What are the new rules of the road in the age of AI?” It's basically saying that if you use AI when writing a proof, you need to tell us about it.

The arXiv, which is where proofs are posted online before they're published, recently issued a statement that if they see unedited use of AI in your uploaded PDF—if they see the metadata from the AI prompt that you copied and pasted in there, didn't even know was there, and didn't review—they're banning you from the platform for 1 year. That's some effort to set new rules of the road.

And then, too, I think they worry that the types of math and incentives to do math, and the types of math problems that LLMs are good at, are different from the kinds they care about. Their own priorities are just going to be steamrolled by the rapid pace of progress in the particular kinds of problems that AI is good at at the moment. All the attention and money that flow to them—I think they're worried the field will get squeezed out and they'll have no say in that direction. They're trying to have a say.

It's—I’m sure this is a naive perspective, but isn't this just what people in every discipline feel when there's a new technology that does the thing they used to do better than they do? Wouldn't the abacus guild have been writing declarations about the dangerous new calculators?

They wouldn't have bothered. They were very violent. They turned straight to violence.

So, I think there are definitely large elements of math that are worth preserving and that AI could, in a certain way, undermine without fully replacing. I think that is a valid worry. Math is a very rich discipline. It's the greatest flowering of the human mind.

I think there are things to worry about—that AI could strip a lot of the incentive and value out of it without fully replacing it.

I want to return to this because I'm still not sure I totally connected with what the folks who signed the Leiden Declaration are worried about. There's a version of this anxiety that I've seen in other professions, which is essentially that AI enables the instant creation of so much stuff—what is often called slop—that it overwhelms and crowds out the people in the industry who are talented and doing a good job. Is this primarily just a slop issue, where they feel like they're going to be overcome by a tide of AI-generated proofs, or is there something else there that I'm missing?

Oh, I think it's actually entirely that. If AI can generate proofs that are really good, and we can read them and all be super impressed by them, then there'll be no more reason for us to have jobs. We'll be hobbyists, like great chess players.

And is the anxiety there an economic one of, “Hey, this is how I feed my family,” or is it something in addition to that, like without humans steering the direction of math, something bad will happen?

I think the anxiety is about having been in possession of something for a long time that was very special, that was rare in the human population: a great ability at math. Now it's generally available. I think that's just a weird thing to reckon with.

I think it's also this belief that math, and the way the community has developed the norms around it, have produced a lot of basic discoveries that are both practically important—math fuels our understanding of the universe, it's important in engineering, and it's important in all sorts of technology. So there's a concern that if you squeeze it out in certain ways, we will lose those downstream benefits.

The more serious point is this belief that math is a quintessentially human endeavor. It's the pinnacle of human thought. It's like writing a sonata; it's like writing a novel. I don't know. I think we don't feel as threatened in those areas by AI because we understand that the human behind the creation is such an important thing.

Oh, I think writers and musicians feel very threatened by AI, just to offer that. I think there's a very similar response happening in the creative community.

Do you? I guess I was not so sure of that statement. Do I think that people would not be as interested in a novel written by AI?

They're not going to be interested if they know that it was written by AI.

Yes. Yeah, that is true.

Yeah.

Anyway, if a proof is not—if there's no human behind it who sat and wrestled with it—then maybe we lose some kind of essentially human endeavor. That is the worry.

You're right. I think programmers are an interesting exception to this sort of defensiveness because, for the most part, many programmers are very excited about the tools that allow them to do their jobs better and faster. Obviously, they're worried about the future of their own jobs, but I don't think you're seeing this kind of Leiden Declaration backlash to the use of AI for programming.

I think it's interesting to see the list of all-star mathematicians who have signed this declaration, including Terry Tao, who we just mentioned and who is fairly excited about the use of AI. He also signed this declaration saying, “Hey, wait a minute. We have to be careful and put some rules on the road here.”

Let me ask what I imagine will be a naive question, though. My extremely limited understanding of math is that mathematics consists of the natural laws of the universe, right? This field is not invented; it is discovered. And my sense is that, potentially, it is a field that could even someday be solved completely because we would simply understand the structure of math and all of those laws of the universe.

So I could imagine a group of mathematicians saying, “Hey, this is really exciting because this is going to accelerate us toward getting to the teleological end of our entire discipline.” But I'm not hearing that today. What am I missing?

Yeah, this idea of whether math is invented or discovered is a never-ending debate. A little more than 10 years ago, Terry Tao won a big prize and was asked that question. His answer, which I very much remember, was that when you're doing math, it feels like you're creating something, but ultimately he viewed it, I think he said, as an act of discovery.

I don't think mathematicians have any concern that we're about to run out of math to be discovered, and AI would have to be a lot better to discover it all. Maybe we already know effectively 0% of all the math there is to know. There's a lot more out there. So I don't think that's a concern, although that is a fun thing to imagine.

Yeah. I'm just wondering: Is there some future of AI and math that is more likely to be a bittersweet acceptance, or is there going to be this kind of principled resistance pushing back and saying, “No, this is still going to be the domain of humans”?

It's hard for me to believe, and I can only speculate. No one knows the answer to this. I ask mathematicians frequently, “Where do you think this is going? What is the future for you all?” No one knows.

It is just hard for me to believe that something that has been so important and central to human activity for so long is just going to completely disappear and be replaced by pushing a button. I think we will be surprised by the way it turns out. Put me down as voting in that middle camp, the Terry Tao camp: Human beings directing these machines in some important ways, choosing which problems to set them on, is going to continue to be important.

So, math is going to look a lot different. There's just no doubt about it. It's going to have to adapt in a lot of ways, just like everybody in almost any industry is. But I think there will be something quite impressive and different that comes out at the end.

Well, Kevin, thank you for helping us balance the equation.

Yeah, you've been integral to the show.

Thank you, guys. It was a pleasure to come on and check your work.

We really appreciate it. Thanks, Kevin.

All right, Casey, before we go, we did have some other tech news that we wanted to discuss this week in our segment HatGPT. HatGPT, of course, is our segment where we take new stories, put them in a hat, draw them out one by one, riff on them a bit, and then when one of us gets bored, we say, “Stop generating.” And then we solve a formal math theorem.

Mhm. That's a new twist on the game. All right, what's our first item today, Kevin?

First out of the hat: A San Francisco startup is secretly testing robots in Airbnbs and trashing them, lawsuit claims. This one comes to us from the San Francisco Standard. Casey, did you hear about this robot testing that's going on in the Airbnbs?

I did. This is a thing. There are all of these well-funded robotics companies, and they need to just practice doing household chores over and over and over again, trying to generate lots and lots of training data. And apparently, a lot of Airbnbs are now caught up in the crossfire.

Yes. On April 12, a Ring camera in San Francisco captured footage of people moving large black cases into a home in San Francisco. 2 days later, the house's owner stopped by the house to take out the trash, looked through the window, and saw black cables taped to the walls. A man was typing at a laptop sitting next to what appeared to be a robot.

And when the guest checked out 11 days later, the house was a mess. Everything had been removed from the kitchen cabinets, the dishwasher was scratched, and everything was a mess because this startup had been training its robots in this person's house without his knowledge.

I have to say, I've rarely felt less sympathy for anyone on our show than the people who opened up their Airbnb to a robot company.

Well, they didn't know they were opening it up to a robot company.

Doesn't matter. I subscribe to the ALAB theory. That's “All Landlords Are Bastards.” So, look, if you have enough money to buy a place, and then you just want to rent it out and charge people these usurious cleaning fees, make them take out the trash on their way out the door, I have no sympathy for you. Let a robot in there.

Airbnb was a 2010 phenomenon. Everyone's staying in hotels again. We have to do something with these spaces. We might as well let a robot mess them up. Am I wrong? Tell me I'm wrong.

You're wrong. You should not be able to just bring in a bunch of robots in body bags to an Airbnb and have them screw up the house. Or if you do, there should be a cleaning fee. The owner of this house is now suing The Bot Company for renting his Airbnb under false pretenses to do robot training, and he's seeking, wait for it, $12,383.50 in damages.

That's now the median cleaning fee on Airbnb, anyway. Okay, I'm telling you, this is a normal Airbnb situation.

Business.

Stop generating. Next up: Trump signs executive order seeking oversight of AI models. This is from Sheera Frenkel and Trip Mickle here at the Times.

President Trump signed an executive order on Tuesday that asked technology companies to voluntarily give the government oversight of new AI models before releasing them to the public. There's a big change here from an earlier version that got scrapped a couple of weeks ago, which is that the number of days that the government will get to review models is down from 90 days to 30 days. And apparently, that was enough for former White House AI czar David Sacks to give it his blessing.

Kevin, what do you make of the new Trump executive order?

So, Casey, what is going on here? Because when we talked to Sundar Pichai a couple of weeks ago, he was going to the White House to be there for this executive order signing. But then that got delayed because apparently David Sacks had a fit and didn't like something that was in the order. So now we have this 30-day provision. I am just very confused about the state of the AI order and whether David Sacks is still running the show from the shadows over there.

Well, he's not the AI czar anymore, but he definitely still has sway. And there are some within the administration who believe that if the government is able to hold up the release of a new model for 90 days, that could hurt American competitiveness. If you bring it down to 30 days, they're just not as concerned. So that's the change that was made.

I mean, the part that just makes me roll my eyes at all of this is the fact that all of this is still voluntary. I think we are now past the point where frontier models actually should have to submit to mandatory, required testing before they inflict these models on the public, but we're still not there yet.

Wait, so it's voluntary to do the 30 days?

It's voluntary to submit it, and then the government has 30 days to, I guess, give you notes.

I see. I'm still confused about the state of AI regulation in this country, and my going assumption is that until I see something that is passed into law and signed, we are just sort of operating in the vibes universe.

Yeah. If the state of AI regulation is, “Hey, what do you guys want to do? What sounds good to you? What would be good? What would work for you?”

Stop generating. All right. What's next?

Okay, we've got: The U.S. is said to be investigating George Santos—we haven't heard that name in years—over Kalshi betting. This one comes to us from our colleagues at The New York Times.

Federal authorities are investigating whether former U.S. Representative George Santos engaged in insider trading by betting on a prediction market about whether he would show up at President Trump's State of the Union address in late February. Kalshi has referred this matter to the Justice Department and the CFTC for further investigation.

Kevin, say a bit more, because am I right that there were maybe some indications that he was going to go to the State of the Union, but then it appears that he may have become aware of that and gone and placed a bet that he would not go—and then did not go?

Yeah. The pattern of facts as we understand them is that just before the State of the Union, George Santos goes on social media and says that he's going to attend, but he missed the speech. Around the time of the event, Kalshi detected that he had bet against his own attendance. So, we salute a legend.

This diva truly will go down in history. George Santos, icon.

Yeah.

Yes, very funny. It's always the ones you most expect, and this is going to be just a hilarious genre of silly crime stories over the next few years. It's just increasingly famous people getting caught with their pants down betting on prediction markets about events that they themselves control.

Yeah, I will say, as an aging millennial, the idea that I could profit from saying I was going to go to a party and then not go is very appealing to me. We've all had that dream. We've been doing that for free for years.

Truly.

What's wrong with us? Okay.

Stop generating.

All right, this next one comes to us from 404 Media: Hackers simply asked Meta AI to give them access to high-profile Instagram accounts, and it worked.

Hackers say they used a Meta AI support chatbot to break into a host of high-profile Instagram profiles by asking the support bot to change the email address associated with the target account. The claims coincide with a series of high-profile Instagram account takeovers, including the Barack Obama White House account, the Chief Master Sergeant of the Space Force's account, and Sephora's account.

Kevin, what do you make of the fact that you can apparently access someone else's Instagram account by asking Meta AI?

Well, I just want to say my old pal at Meta AI, Nasty Nancy, would never have done this. She would have upheld the integrity of these Instagram accounts. But now they've turned it over to this crazy chatbot who's just giving away people's passwords.

Look, I know it seems like this story is bad for Meta, but I actually think it's good because we have finally found something that Meta AI is good for. And I'm not sure that I could name a second thing. So, congratulations to the superintelligence algorithm.

We learned about this when someone posted the apparent account credentials and cellphone number of Mark Zuckerberg on X. I have not tried the number yet to verify that it works, but it appears that you can just go on and ask Meta AI for anything.

I'm going to guess that that's not actually his contact information and is, in fact, some sort of phishing scam that would harm you.

Only one way to find out. Everyone, try calling Mark Zuckerberg at—

Kevin, you know we have a strict no-doxxing policy here on the show.

Okay, well, I'll save that for our bonus content. All right, stop generating. Next up: United—oh, this is my favorite story of the week. “United flight forced to turn around because of a Bluetooth speaker name.” That's the Verge headline. A United Airlines flight from Newark to Mallorca, Spain, last Saturday night had to turn around about 2 hours after takeoff and make an emergency landing due to security concerns over a Bluetooth signal. The crew on the flight asked passengers to turn off their Bluetooth devices multiple times, and everyone complied except for one speaker that belonged to a 16-year-old boy and was named “Bomb.”

It's pretty funny because there's maybe only 1 word in the English language that you could name your Bluetooth speaker that would force an emergency landing, and you picked it, brother. Congrats.

It's so true.

I was following the story on Reddit, where there were, weirdly, a number of passengers on this flight who were active on Reddit. During this, they were saying, “The pilots and the crew are coming on and telling us that we all have to turn off our Bluetooth devices immediately. What's going on?” And then you follow it in real time. Finally, they figure out, “Yeah, there's this Bluetooth speaker, and when you connect to it, it shows up on the little list of Bluetooth devices as ‘Bomb,’” which is a very bad name for a Bluetooth speaker. Don't do that.

It's a very bad name for a Bluetooth speaker unless you plan on playing some bomb-ass tunes. You know? Look, it's all about context. Also, I have news for people who are running airport security: Most bombs that would blow up planes cannot actually be connected to via Bluetooth and are not named “Bomb” in the Bluetooth list. These are not discoverable devices that are advertising themselves as what they are. It reminds you of—do you remember when people used to give their home Wi-Fi networks names like “CIA surveillance van”? The real CIA surveillance van is not named that.

Yeah, look, I think we have a very important message to deliver to airline security, and that is: You guys have to have a sense of humor. You know what I mean? You guys need to relax, live a little. Take a chill pill.

It was a damn Bluetooth speaker. Do you know how irritated I'd be if I was 2 hours into my flight to Mallorca and now we had to turn around? I'm trying to get to that beach.

All right, last one out of the hat.

You didn't say “stop generating.”

Stop generating. Okay, last one out of the hat. Oh, Casey, this one's in your wheelhouse: “Survivor boss Jeff Probst says Kalshi and Polymarket are ‘incentivizing people to lie, cheat, and steal.’” This one comes to us from Variety. Apparently, there was some drama on Survivor recently when one of the episodes was spoiled due to widely circulated reports about the odds on these prediction markets, Kalshi and Polymarket. On both platforms, Aubry Bracco was forecast to have an above-80% chance of winning before this season even premiered.

Yeah, so this is a story that brings together 2 of my favorite things, which are Survivor and hating on prediction markets. We've talked on the show about the fact that one of the main things prediction markets do is incentivize you to betray your friends, family, co-workers, and possibly your country.

And so what Jeff Probst is noticing is that now all of his crew members could stand to make a ton of money by betting on one of these prediction markets. Now, it is important to say that we do not currently have information to suggest that this is what happened. We don't know of somebody in particular on the Survivor crew who may have leaked the information, but now we're just living in a world where everyone is suspicious, and I think it just contributes to the low-trust society that we're already living in.

So this is obviously very bad, but let me take this opportunity to say, Kevin, that while I do think Aubry played a great game and had a great season, I have to give a shout-out to the greatest player to never win the game, Cirie Fields, who truly came so close on this season and was amazing to watch every single week. I was absolutely crushed when she lost. So, just incredible work, Cirie, and to the Survivor team for putting out 50 really amazing seasons of television.

Only 1 or 2 of which you could bet on in prediction markets. They were just doing this for the love of the game.

Mhm.

We have to start, pretty soon, a segment that's just about people getting caught for doing stupid stuff on prediction markets. In addition to the George Santos thing and the Survivor thing, there was also a story this week about a Google engineer who was charged with using inside information to make $1 million on Polymarket by placing bets on what users were allegedly searching for. So truly no corner of society is safe from the corrupting influence of prediction markets.

Yeah, when you've got that Google money and you're still trying to make a little extra scratch by corrupting a prediction market, something's gone wrong in this society.

Yes. So that is HatGPT. Let's close the hat.

Close up the hat.