Evacuate? In our moment of triumph? I think you overestimate their chances!

This week we expect one or more private AI labs to post another round of size $N$ ($N=10, 100, 400,\ldots?$) bundles of LLM-generated solutions to mathematical problems. At this point this feels more than anything like the swan song of the dying order in mathematics and academia. Maximizing your number of publications no longer lands with the same effect as it used to in the old days pre 2026. Somehow, AI labs themselves are landing on the same pitfall as those authors reported to be posting – on their own – hundreds of LLM-generated papers in a single year. If not already, then soon, it will become clear to everyone that such stunts will not impress anyone.

I think mathematicians at the AI companies understand this, however, releasing the data set of these solutions is not up to them. The decision makers at the AI companies are not experimenting with LLMs and seeing what they can do, nor are they invested in the long term health of a field they love and care for. On the other hand, these decision makers understand the value of headlines and know how to sell things, and see in these hyped-up press releases an opportunity to exploit the public’s feelings of fear and awe around mathematics. This perception leads AI companies to believe the public will take notice if their systems are able to do “things mathematicians have not been able to do before.” At the same time, mathematicians have used (and misused) “conjectures and open problems” to generate interest and excitement in mathematics within the broader public. As a result, those “conjectures and open problems” have become a convenient tool for AI companies trying to “farm-off” the aura of awe and power in mathematics.

By now, however, a sizable number of mathematicians have become familiar with LLMs, and can see these announcements in their proper light. I myself only started playing with LLMs early in the summer, and have used them more systematically in the last 6 weeks or so. My overarching impression so far is that the real fun and transformative practices around mathematics and LLMs are still to emerge, and I expect them to come after we all have enough time to experiment and play with these tools. Once that happens (in months? years?), the flood of AI-slop papers will die out, or so I hope.

Mathematicians, in short, are rapidly coming to understand what LLMs are and what they do, and (some) are beginning to see the dawn of a new era. What this new era means is something we are still figuring out: we are a big community; there are a multitude of perspectives and circumstances and backgrounds, and the arrival of LLMs is not evenly distributed around the field. This means some of us are more excited and optimistic, and others are more apprehensive and concerned.

I am in the excited and optimistic camp, which is not the same as thinking the transition will be completely smooth and devoid of problems. We now have at our disposal the immense power of mechanized cultural technologies, which we can use to produce genuinely new mathematics results, even results that a year ago would have taken a year or more of focused work. I hope that as this reality sinks in, pushing a button to generate a one-shot solution to a problem will become largely devalued and disincentivized.

We now have a formidable technology: the aggregated data of mathematical knowledge of all mathematicians from all time, digitized and primed for computer manipulation, to be explored and expanded in an automated way. Using this to generate just another paper or prove a known theorem as if solving conjectures were a competitive sport is a narrow and impoverished view of what mathematics is.

The old order was not perfect but it had its moments, but its time had been running out even before the arrival of LLMs. To the news that the arXiv will be rate-limiting submissions to 2 submission a month per person, many of us quipped “great decision, and even 10 years ago it would have been great”. The old order had a serious flaw: misplacing too much value and rewards in the competitive-like aspects of mathematics. For that, I am glad it will be gone soon.

It is poetic justice that this order is dying at the hand of the AI industry, which in recent years has become hyperfocused in zero-sum, competitive thinking. For the vast majority of mathematicians, mathematics is not a competition but a personal and profound labor of discovery, of creation, of truth-finding, and of expression. The nascent order will be ours to shape.