Why this prof bet against AI’s math skills – and then suddenly changed his mind

On March 9, 2025, Toronto-based mathematician Daniel Litt started a social media thread aimed at puncturing what he saw as overblown predictions of how powerful artificial intelligence was becoming.

“They’re going to automate you. AI will do your job. It’ll be better than you at everything. When you want to smell a flower it’ll smell for you, better than you ever could. It’ll be the flower too,” he wrote.

That post sparked a discussion about the merits and current capacities of AI, and that discussion sparked a wager between Litt, an assistant professor at the University of Toronto, and Tamay Besiroglu, a U.S.-based AI researcher and entrepreneur. Besiroglu bet $250 that by 2030, a large AI would be capable of writing a paper good enough to be published in The Annals of Mathematics, one of the top academic journals in the field. Litt agreed to pay Besiroglu three times that amount if it happened. 

They worked out terms on X, with Litt clarifying his position along the way. “FWIW I think it’s quite likely that AI will be involved in such a paper by that time, and quite unlikely that it will be able to produce one on its own,” he wrote.

Though the bet was intended to take five years to produce a winner, last week, less than 18 months after the terms were set, Litt conceded. No AI-generated paper has yet been accepted into Annals, but OpenAI announced on Aug. 1 that its new model, Astra, had produced significant new results for several long-standing mathematical challenges. This, along with other developments, changed Litt’s mind about how quickly things would progress.

Be Giant interviewed him to find out what led him to give up the wager so soon.

Tell us about your research and how you came to be thinking about the issues at the heart of this bet. 

I do algebraic geometry and number theory, which are pretty pure abstract branches of mathematics. I have been interested in AI to do math for a long time. Around the time of the bet, there were a lot of claims about what the models could do and couldn't do, which were dramatically overhyped at the time, though I think many of those claims have come true by now. I was talking about that with Tamay, having a discussion about where we thought the future was heading.

How did the two of you come to the idea of placing the bet?

He proposed it in our Twitter thread. In the conversation, I said models would be doing high-quality math research probably in the next couple of decades. He thought it would be faster. I had some general sense of skepticism about model capabilities.

What were the precise terms of the bet?

I gave him three-to-one odds that, by 2030, AI models would not be able to autonomously produce number theory papers that I judged as at the same level of quality as 2025 Annals of Mathematics papers. Annals of Mathematics is one of the top math journals. I conceded the bet early – it still has not happened, but I expect it will happen very soon.

You didn’t make the bet that long ago. What changed?

One thing is that I expected the models to be doing math in a slightly more human way. Right now, the models are very good at doing certain types of things and not others. I think I imagined that if they could produce an Annals paper, they could produce any Annals paper. That is definitely not true. Maybe it will be true next year, but right now the kind of high-quality work they're doing is a fairly small sliver of what mathematicians do.

I was also wrong about the timelines. I expected things to move slower than they did. At the time, it was very clear that the models were superhuman in some ways. They knew everything. A human who knew all those things would basically be guaranteed to be doing really good math – the models weren't. It seemed like they were missing some capabilities. It turned out to be much easier for them to get those missing capabilities than I expected. It also turned out [that] some of those capabilities weren't necessary. There are a lot of skills and capabilities a human uses to do high-quality math. I think the models are still missing a lot of those human skills, but they're also so superhuman in other areas, they can substitute for the missing ones.

Does this re-evaluation make you feel differently about the viability of mathematics as a human pursuit?

Of course, I would have preferred from a financial point of view to win the bet, but from a scientific point of view, it's probably better that I lost. I'm excited to see some cool math. I think the institution of academic mathematics needs to change pretty rapidly, and there are going to be a lot of growing pains as that happens.

There's a huge opportunity here. These new tools should allow us to do incredible work. OpenAI just released 10 solutions to pretty interesting open questions in mathematics. Who's going to read those solutions? It’s going to be academic mathematicians, the only people who have the capabilities to interact with these new tools. There's a plausible case to be made that this is the beginning of an amazing growth period for mathematics, and I hope institutions feel the same way.

There's some danger if we don't adapt appropriately or funders don't recognize this huge opportunity. But from a scientific perspective, this could be amazing if we take advantage of it.

Many people's experiences of mathematics end after high school. Can you talk about what makes a good mathematician and what makes an Annals-worthy paper?

The answers to those two questions are different. What do mathematicians do? Typically, you're trying to understand some mathematical objects: numbers, shapes or something like that. You ask a question about it. It's actually incredibly easy to ask a question about those objects that no one knows the answer to.

So what makes a good paper? Well, maybe you ask such a question and answer it, or maybe you answer a question someone else asked. Some questions get a reputation of being interesting and hard, meaning lots of people care about them. The nature of some of these recent [OpenAI] results is that they took some old questions that people had asked, and that the community had not fully answered.

To put it in perspective, let me just say, OpenAI published this list of 10 solutions. This is 10 out of, I don't know, a million or more. There are an unlimited number of questions. It's exciting to have answers to these problems – I would have been very proud to have solved any of them – but it's not like this is the end of mathematics.

If you were going to make another bet now on where we'll be in five years, what would it be?

I think that the models will continue to get better very, very quickly. So if you're asking me to predict a capability they will not have, that is probably not a bet I would make. What I would bet is that there's still going to be a lot of very interesting mathematics happening in five years and a lot of it will be done by people.

This interview has been edited and condensed.