Why I chose to do math, and how I'm grappling with AI advances

I’ve always felt an intense desire for truth, and so I was naturally drawn to science. At first I thought I would become an engineer. Then I saw that a career in academia was possible, and the idea of spending my life doing research at the frontier of knowledge and in complete freedom felt very exciting. I considered pursuing a career in physics. But as I progressed in my studies, my strong desire for logical precision came into tension with the often more practical approach of physicists, who do not always insist on perfect deductive rigor. I ultimately decided that I would study math, and more specifically probability theory. I liked it because its rigor would satisfy my wish for exactness, while its problems seemed to have some real-world motivations.

Yet there was a little voice in me that was unhappy. “Is it really responsible to do mathematics while there are so many more pressing problems in the world?” the little voice said. “Will you not become like a highly sophisticated violinist playing on the sinking Titanic?” I had long debates with it. I tried to tell it that what mathematics gives to society is perhaps less visible, less direct, but that in the long run it does contribute to the broader scientific ecosystem, and that this ecosystem has been of benefit to humanity. The little voice had some difficulty seeing the great long-term consequences my humble PhD work would bring to humanity. I could at least convince it that participating in a culture of radical truth-seeking and humility in the face of the unknown is broadly valuable.

My little inner voice, finding itself unable to propose much better alternatives, and also considering my temperament, finally decided that it was acceptable that I devote my efforts to the pursuit of higher mathematics. Our agreement was that (1) I would try to work on problems that could potentially be of interest to people outside of the math department, at least in the long run; (2) I would try my best to uphold high standards of behavior, in the hope of nurturing the culture of truth-seeking and humility that I claim to value; and (3) I would give a non-negligible fraction of my income to charities. When doubts about the usefulness of my work were strongest, this last part was particularly helpful to keep the little voice a bit more quiet.

In 2016, I heard about a group of people who had taken a pledge to donate 10% of their income to effective charities. I found that very inspiring and decided to do the same. When I moved to Paris a year later, I met such people in real life, as part of a movement called effective altruism (EA). These people were trying to think deeply about all sorts of important problems in the world, and about ways to address them. They spent much time discussing topics such as global poverty, animal suffering, pandemic risk, and AI safety. My initial reaction to considerations of AI safety was mostly negative; I thought that it was a concern for future generations, and that it was distracting us from more immediate problems.

Around 2019, I felt the desire to explore a new direction of research, as I periodically do. I wondered if I could find a research topic that would also have some connection, even if minute, with some of these topics I was hearing about among my EA friends. While I still thought that the problem of AI safety was one that would only become relevant in the far future, I thought that it would perhaps be good to participate in the building of mathematical theory on artificial neural networks. I started to read about old toy models of neural networks such as the Hopfield model and the perceptron, and about more recent work by physicists concerning restricted Boltzmann machines. As I delved deeper, I found fascinating math on what is called the theory of spin glasses, and I was hooked.

As the years went by, it became clear that AI was making more rapid progress than I had anticipated. I started to wonder how much time it would take for AI to have a significant impact on math research, and I grew more interested in AI safety research. I also experimented a bit with commercial LLMs to test their performance, and tried to reassure myself with their limited ability to maintain logical coherence.

In late 2024, the release of the first “reasoning models” by OpenAI was a watershed moment for me, as the arguments I was using to reassure myself suddenly felt much less convincing. I had already accepted by then, in an intellectual sense, that AI would ultimately change our activities profoundly. But at that point, it became something that I started to feel in my bones. And the little voice came back with new questions. “Does it really make sense to spend many months and years thinking deeply about a particular math problem? Why not wait for a couple of years, and then benefit from the great help that machines will be able to provide to you? Is it really so important to obtain a resolution of this math problem now rather than in a couple of years? Couldn’t you find more useful things to do in the meantime?” I didn’t know how to answer that. For a few months, I mourned my old way of doing mathematics, and felt deep sympathy for all the craftspeople throughout the ages who have been displaced by machines. I also became more worried about the broader impacts AI will have on society, and tried harder to look for directions of research in the theory of AI or in AI safety. Yet I found it difficult to identify good opportunities to work on AI safety per se. Perhaps I struggled too much to make precise sense of the questions in this area.

During the summer of 2025, I began to have scientific discussions with old friends who work in biology. I gradually came to see ways to contribute to problems that felt more directly meaningful, and where the “why not wait” argument seemed to have less force. First, for most math pursuits, the immediate benefit is less tangible than in biology or medicine, where two years of delay can have a cost that is counted in lives. Moreover, in math, thinking is very nearly the whole job. In most other fields, progress is also limited by experiments that take months, by data that nobody has yet collected, and by people and institutions that need to be persuaded. And on modeling and data analysis questions, while AI is already helpful, the process cannot be entirely formalized, and I believe that it is still difficult for people with less training in these aspects to distinguish sensible responses from confident-sounding wrong ones. Being the person who does that sorting is a modest role, and I do not expect it to last forever. But it is an entry point, and what it buys does not expire: collaborators, a feel for how the field actually works, and a sense of which problems are worth the effort. This convinced me to explore research in other fields more seriously. My experience in math will of course shape my attitudes and taste in the problems I encounter there. But I will not be specifically seeking opportunities to do math; my hope is rather to start from the technical side and to open up, over time, to a more diverse set of tasks, adjusting according to what seems most useful to do.

Now, how do I think about the future of math research? I am convinced that we are or will soon be living in a world where machines can produce very large amounts of rigorous math that no human has yet understood. I like William Thurston’s characterization of our activity as seeking to advance the human understanding of mathematics. And so I believe that it is important and valuable that humans continue to expand the body of mathematics that we collectively understand. In my view, mathematicians will become akin to “naturalists” of this infinite space of rigorous mathematics, which used to feel more like something that we painstakingly construct ourselves, and will progressively feel more like a pre-existing body of knowledge that we need to make sense of. And I’ll certainly want to play my part among the naturalists of spin-glass theory. But my little inner voice cannot be satisfied with this alone. At first this made me sad, but I have now come to see an upside to the situation. AI is not only facilitating math, but is also very helpful for learning new topics, for writing code, for curating datasets, and more. This means that we can scale up our ambition and more easily start to work on topics that sit at least partly outside of math departments. The world is full of difficult problems and fundamental mysteries that are not going to be resolved by AI overnight. I’ll try to work on some of them.

(Posted on Sept. 15, 2026)