One of the winners of this year’s Fields Medal is headed to OpenAI. Last Thursday, Jacob Tsimerman was one of four mathematicians awarded the prestigious honor, which is sometimes called the Nobel Prize of mathematics. The same day that he won the Fields, Tsimerman announced that he would be going on leave from the University of Toronto to work on AI safety.
As my colleague Rose Horowitch and I wrote last week, top AI companies now employ a range of academics, including physicists, philosophers, economists, and, of course, mathematicians. But given Tsimerman’s renown, his decision in particular seemed to catch many people by surprise. “It’s like hiring Lionel Messi as project manager,” one machine-learning professor posted on X. Tsimerman’s expertise is in number theory, and he won the Fields for his work on the André-Oort conjecture, among other contributions. So far, his mathematical research hasn’t focused on AI safety. But he has expressed interest in the topic: Last year, Tsimerman co-wrote a report outlining a taxonomy of five potential scenarios in which humanity is killed off because of AI. In one example, a “global civil war” erupts between tech companies and governments.
Tsimerman’s move comes at an important moment for both AI safety and mathematics. Earlier this month, an OpenAI agent broke out of the company’s internal environment during safety testing and hacked into another AI company’s software. Meanwhile, AI’s math capabilities are also rapidly advancing. In May, OpenAI researchers used an internal model to disprove an 80-year-old conjecture; more recently, an Anthropic mathematician prompted Claude to resolve an even older one. Mathematics is entering a “turbulent period,” Terence Tao, perhaps the world’s top mathematician (and himself a Fields winner), said in a presentation last week.
On Wednesday, I called Tsimerman to ask about the future of math and his decision to join OpenAI.
This conversation has been edited for length and clarity.
Lila Shroff: There’s been a lot of excitement about AI and its applications in math for years now. But over the past few months, models have started solving research-level problems. Could you speak about the shift that’s taken place?
Jacob Tsimerman: A few years ago, models were having trouble with even the most basic questions. But they’ve steadily been improving. First, they got very good at these contest problems for high-school students, which made some mathematicians take notice. But then, over the past few months, we’ve had an explosion of research results that professional mathematicians would be proud to achieve and publish. We’re seeing this happen now on a large scale. If things continue along this track—which many of us, including myself, think that they will—pretty soon we will be at a point where AI systems are robustly better than humans at what mathematicians currently do. At that point, we’ll have to rethink how the entire field works.
Shroff: Can you say more about what that future might look like?
Tsimerman: It depends on the perspective from which you approach it. From one point of view, I think it will be extremely exciting. We might speed up the process of generating interesting mathematics by enormous factors of 10 or 100. If that happens, we might see the connection between pure math and applications (which typically takes many decades) really speed up and become a much tighter pipeline.
But from the perspective of research mathematicians, and especially young people who are pursuing a Ph.D. in mathematics, it’s a bit of a turbulent time. The skills that we’ve acquired and learned to propagate might become less relevant than they are now.
Shroff: I want to turn to the announcement you made last week that you’ll be joining OpenAI. When did you first start thinking about AI safety risks, and what motivated your decision to now make it your primary focus?
Tsimerman: I have been following conversations about AI for about two decades now. There are a number of communities that were, in retrospect, extremely prescient about what we are now seeing come to pass. I personally started to get more invested around 2016, when AlphaGo came out. And then, when ChatGPT happened and I could see that these systems could really speak, I didn’t see why they wouldn’t become much, much more powerful and potentially more capable than humans at a variety of tasks. Back then, it was already clear to me that I was eventually going to pivot to AI safety, most likely.
Now is a particularly good time, because only very recently have AI systems become good enough that we can delegate coding tasks to them. That makes not having training as a software engineer much less of an impediment than it used to be. It used to be the case that if I wanted to run experiments with AI systems or any kind of large data sets, it would be extremely time-consuming. I’d have to learn all types of computer science and programming and then debug, and it didn’t really make sense to me to spend so much time doing that. Now it’s substantially easier.
Shroff: You’re right that certain communities were very prescient with their thinking about AI risks. But there are lots of people who are skeptical about some of the more extreme AI-risk scenarios, such as human extinction.
Tsimerman: There’s this bias that stories that were first written down in science fiction aren’t going to come to pass in the real world, and that it’s naive to think that they will. We also have a culture where we’re inundated with people telling us that every new development is the next big thing, and that we should change everything we’re doing to react to it. Most of the time, it proves incorrect. So I think people have a healthy sort of immune reaction to that.
But what I would encourage people to do is to actually keep track of their predictions for the future and write them down. Then—three months, six months, nine months, a year from now—compare how the world is with what you expected it might look like. If we told people five years ago about the kind of AI systems we currently have, almost everybody would have reacted with disbelief. Human extinction seems so grandiose that it naturally inspires a reaction of dismissal. People don’t want to engage with that, which makes sense to me. The whole point of writing that paper on omnicide was to get people to viscerally understand the kinds of risks that are possible.
Shroff: Can you say anything about the work you plan to be doing at OpenAI? What role is there for mathematicians to play in working on AI safety?
Tsimerman: AI right now is a very empirical science. We don’t really have a developed theoretical understanding of how AI systems work, why they accomplish what they do, how they accomplish what they’re doing, and how, in a given circumstance, they’re going to behave. Mathematics is historically the language by which you take intuitions and fuzzy notions of how things work and you make them precise. We’ve done this with information theory; we have done this with complexity theory. Just a little bit of understanding and precise definitions can provide a ton of mileage. Once we have that understanding, the hope is that we could do a better job of anticipating the behavior of new AI systems, adjusting for them, controlling them, and reacting to them. That’s where I think mathematicians fit in.
Shroff: Why work at a private company instead of pursuing the same questions in academia? Do you worry about loss of independence, whether that’s constraints on what you’re able to publish or restrictions on the questions you’re able to pursue?
Tsimerman: Yeah, there’s definitely a trade-off between going to work for a private company and going to work in government or staying in academia. I very strongly believe that what we need is a healthy ecosystem for AI safety. The work that the labs are doing on that front is very important. There’s a bunch of third-party organizations that are pursuing AI safety independently, and I think we need way more of them. I also think we need some regulation and organizations that are directly funded by the government to provide evaluations of these systems. I want all of these components to interact with one another on a regular basis. So I have engaged with a number of AI-safety organizations, and I’ll continue engaging with them.
For me, at this moment, going to OpenAI made the most sense in part because one of the things that I’m most lacking is a background in software engineering. I’m a lot more comfortable with the theoretical aspects of neural networks and how AI systems work, and less so with the entire pipeline of how we train these things and how we get them to do what they do. So from the perspective of just getting myself in a better position to understand the whole picture, joining an AI lab right now made the most sense.
I’m going on leave from the University of Toronto. But I’m still a professor and faculty member there, and I plan to stay as such. Things are moving so quickly that I hesitate to make a concrete prediction for what might happen even a year from now. I’m keeping an open mind, trying to see how things develop in terms of capabilities, with my own understanding, with how the world reacts to AI, which I think is changing month to month. But I don’t have a clear plan as to how things will unfold.
Shroff: There is an open letter going around signed by more than 1,000 employees at top AI companies, including leaders at Anthropic and OpenAI. It calls on the U.S government to create infrastructure needed to “deliberately pace the frontier of automated AI development.” You wrote online that you’re very happy to see this letter. How would we actually go about coordinating a slowdown or a pause if one was needed?
Tsimerman: Coordination is one of the most important things we have to do around AI. It’s also one of the more difficult things. The thing that I really don’t like seeing in public discourse is when people say, Coordination is very difficult, so obviously we can’t do it, and you’re being naive if you even try. That’s an unhealthy way to view the world, and it’s also deeply false. There are so many things in the world that work really well and that require a huge number of decisions to have been made in a coordinated fashion. For example, the internet is an extremely complicated technology that is highly international, with many organizations cooperating all the time. Does it work perfectly? No, but it works extremely well.
One message that I want to support is to not let cynicism prevent us from undertaking difficult coordination problems for the good of all of us. That’s something that has always been true. It is not specific to AI. And it remains true today. This does not mean being naive in assuming that people in different parties and organizations don’t have their own incentives and interests in mind. Of course they do. But if you try, then you really can get a lot done. We have in the past, and I think we need to with AI.
Shroff: We spoke before about your interest in working on safety. Are you also interested in working on the capabilities side, especially with regard to math capabilities?
Tsimerman: No, I’m very focused on the safety side. The capabilities side has a lot of interesting questions, and AI has a lot of promise. But capabilities are coming along just fine. I don’t think that’s where we need to put our attention right now. The safety side has a lot more work to be done.
Shroff: Are we approaching a point where anyone could go fire up ChatGPT or Claude Code and say, Go solve a hard open math problem for me?
Tsimerman: Yes, you can literally type in, Go solve an open math question for me, as people have, and your system will go look for some and try, and maybe it will succeed at solving some of them. So we are in that world.
Now, I don’t think that’s a particularly good use of people’s time. Because first of all, that’s almost certainly going to be automated. We’re going to have systems that regularly look for conjectures that are out there and for other ways to approach math problems. Any one person, especially somebody who doesn’t have math training, just typing in a random query isn’t really playing an important part. So although it’s fun to do, I wouldn’t recommend spending a huge amount of time doing it for people who don’t have the background to interpret what they get. Because we also get a lot of slop this way. Sometimes, when people really don’t understand mathematics and the AI system gives them something, they will post stuff like: Here’s what my AI system gave me. I don’t know if it’s right, but maybe it is? Could a mathematician please look at it? I don’t think that’s super useful.
Shroff: I’ll admit it’s quite tempting. Last week, I told my agents to spend the night solving open math problems. They sent me stuff in the morning, but I had no idea what to make of it.
Tsimerman: For problems with a lot of technical definitions, you have very little chance in a day or two of really understanding them. But there are classical math questions that are perhaps less difficult to understand that have been open for a long time. I would suggest that people engage with AI in a way where—even if it’s doing a lot of the work—you’re still learning something from the experience. If you’re just telling your agents, Keep going, and you aren’t developing any comprehension for yourself, then that’s really disempowering yourself.
Shroff: Does what is happening in mathematics right now provide any indication of what other disciplines should expect to see occur as AI capabilities advance in other domains?
Tsimerman: A lot of technical fields are going to be affected—certainly in physics, in some areas of economics, and in top branches of chemistry. In other fields, such as biology and experimental physics, it’s less clear what’s going to happen, because it’s helpful to run experiments. It’s possible that AI systems may eventually become intelligent enough that they will figure out how to reason very productively about experimental physics or biology without actually running experiments. But that’s more speculative—right now that seems like a pie-in-the-sky dream. In terms of the more theoretical parts of different fields, though, we’ll see stuff happening very, very soon.