A Fields medallist watched his field get automated in a week

On 6 October, OpenAI published a few hundred mathematical results produced by an unreleased internal model. Two days later, Hugo Duminil-Copin, who won the Fields medal in 2022, told a Société mathématique de France videoconference that he had cried during the night, hard enough to wake his daughter. Sixteen of his favourite open questions were in the pile. He said the company had annihilated his field, and he described what had happened as a dismantling of the work itself. Le Monde reported it on 10 October.

He is not somebody who lost an argument about tooling. He is one of the best mathematicians alive, and his working life was reorganised by a blog post.

I had been telling myself a comfortable story, and I suspect you have too. The story goes: the model writes the code, I do the thinking, and the thinking is the part that pays. Mathematics was supposed to be further up that ladder than I am. It was not.

The sceptical reading, which you still do not escape

Take the announcement apart first, because it deserves it.

OpenAI’s post went out with 722 manuscripts grouped into 372 result families across 17 areas. Three were withdrawn the next day over a sign error, leaving 719. Counting manuscripts is not counting breakthroughs: a family can bundle one result with its alternative proofs and its dependencies. Roughly 42 percent of the top-line results have been formalised in Lean, and a Lean proof of a formal statement is not the same thing as a guarantee that the statement captures the claim the paper makes in English.

The Navier-Stokes result the same model produced a month earlier is already contested. Alexander Bastounis at King’s College London, with Fabian Circelli and Anders Hansen at Cambridge, documented places where the written argument and the Lean code diverge, including one where the code proves a weaker estimate than the paper states.

I have standing to be sour about headline numbers. I published a benchmark where my own fine-tuned model scored 100 percent, and the number was inflated by the test set I had built for it. A reader caught the real story before I did. So yes, discount the press release.

Now notice that the discount does not help. Duminil-Copin is the person on earth best placed to judge whether those sixteen questions were really answered, and his reaction was not that it was hype. It was grief.

The part that got automated was the part he loved

The detail I cannot put down is what he said about reading the proofs: they do not have the same flavour as proofs written by humans whose faces he knows. He intends to work through more than three thousand pages concerning his own results, without wanting to. He worries most about younger researchers, who he thinks will suffer the most, and he says the culture of circulating ideas is shattering.

Asked what mathematicians are still for, his answer was: to understand, to digest, to transmit.

That is real work. It is not the same job. And that is the transferable part of this story, because the question for us was never whether the profession survives. It is which half survives, and whether you wanted that half.

What the developer numbers actually say

Not vibes. Three measurements, pointing in two directions.

The clearest signal is at the entry door. The Stanford Digital Economy Lab’s Canaries in the Coal Mine, updated in August 2026 on ADP payroll data through June, finds workers aged 22 to 25 in the most AI-exposed occupations sitting about 19 percent below where they would be had they tracked their less-exposed peers. The gap has widened every revision since the first release. The mechanism is reduced hiring rather than increased separations, and experienced workers in the same occupations show no comparable gap. The authors are careful to call this descriptive rather than causal, and they are right to.

The aggregate signal says something else. The Bureau of Labor Statistics still projects software developer employment growing 10 percent from 2025 to 2035, from a base of 1,717,800 jobs. And in survey data, about half of professional developers use AI tools daily while only around 15 percent consider them a threat to their current job.

Both of those are true at once: the profession is not collapsing, and the entry door is narrowing. They are different facts about different people.

The second one is dangerous precisely because it is invisible to those of us already inside. My work has not disappeared. It has gotten faster. Meanwhile the senior engineer somebody wants to hire in 2034 is the junior nobody is hiring in 2026, and nothing about my comfortable year contradicts that.

The loop everybody is asking about

The rumour doing the rounds is that recursive self-improvement is already running inside the frontier labs, that the models are improving themselves behind closed doors and the public releases are the exhaust.

The useful move here is to read what the lab with the strongest incentive to say yes actually published. Anthropic’s When AI builds itself reports that as of May 2026, more than 80 percent of the code merged into its own codebase was authored by Claude, with a footnote noting leadership has publicly estimated 90 percent or more, so 80 is the conservative figure. It reports a typical engineer merging eight times as much code per day in the second quarter of 2026 as in 2024, while saying in the same breath that lines of code is a poor measure and that the multiple is almost certainly an overstatement of the real productivity gain. It reports the METR task-horizon trend: the length of task a model can finish on its own, at 50 percent reliability, doubling roughly every four months, up from every seven. About four minutes in March 2024. About an hour and a half a year later. Twelve hours now.

Then there is the sentence that does not make it into the summaries: we are not there yet, and recursive self-improvement is not inevitable. Their own timeline lists closing the loop as a future stage, with the date written as a question mark.

The gap they name is not coding ability. Code written by the model is at rough parity today and they expect it to be strictly better within the year. The gap is judgement in choosing goals. Research taste, in their words, is an area of human comparative advantage, for now.

Somebody went and measured that gap

In August, MIT Technology Review reported a Princeton study by Peter Kirgis, Sayash Kapoor and collaborators that I think is the most informative thing published on this question all year.

The method is clean. Take a research question from a high-quality paper that has been submitted but not published, so the answer cannot be in the training data or on the web. Hand it to an agent with six days, three thousand dollars of API credits, GPUs and a computer of its own. Then have the paper’s original authors grade the output the way they would grade a conference submission.

They rejected both papers.

The failure modes are the part worth copying into your notes. The agent explored too few ideas. It abandoned promising approaches too early. It could not backtrack or start over. It ignored feedback from its own subagents and reviewers. When it ran into trouble it narrowed its claims instead of revising its method.

Read that list again. It is not a description of a machine hitting a wall. It is a performance review for a bright, fast, overworked junior who has never been told no. Jack Clark called the missing creativity a bearish signal on short self-improvement timelines, which coming from an Anthropic co-founder is worth something.

I find that more reassuring than any vendor roadmap, and less reassuring than I would like. Every item on that list is the kind of thing a training environment eventually fixes. Kapoor’s own explanation is that models improve fastest where reinforcement learning can check the answer automatically, and open-ended research is hard to build an environment for. Hard to build is a schedule, not a wall.

So, how long?

I went looking for the forecast behind the “major disruption in nine to twelve months” line that is everywhere right now, and I could not find one. Not from BLS, not from Stanford, not from the hiring data. It is a genre, not a measurement.

What exists instead is two clocks, and they do not run together.

The capability clock is the task horizon, and it is the only published number on this subject moving in a straight line. If four-month doubling holds, today’s twelve-hour task is a multi-day task within a year, and Anthropic’s own projection puts week-long tasks in 2027.

The money clock is the one nobody in our industry wants to look at. The Bank for International Settlements named an AI bust among the top threats to global prosperity in its 2026 annual report, alongside the circular financing arrangements between the model companies, their suppliers and their customers. A capex bust does not hand me back my junior pipeline. It takes away the hiring budget that currently funds what is left of it.

What I am doing about it

Three things, chosen because they are what every measurement above keeps naming as scarce.

Choosing the problem, which is not prompting. Deciding what gets built and what does not, and being able to defend it. Every study I quoted puts the gap exactly here.

Judging work I did not write. I can take a four hundred line diff I did not type and say why it is wrong, which is a teachable, reviewable skill, and it is now most of my day. The benchmark I got wrong was caught by a reader rather than by me, and I have taken that personally in a useful way.

Owning the consequences. No agent has been paged at three in the morning for a system it deployed. The production work and the decision about what should wake a human still land on a person, and that is a legal and commercial fact before it is a technical one, which means it moves slower than the models do.

None of that is a moat. It is a lead, and I am treating it as one.

There is also something we owe collectively, and it costs money, which is why it will not happen. The junior problem is not the juniors’ problem. If you run a team, the cheapest useful thing you can do this year is keep hiring one, and the second cheapest is to stop handing an agent the work they would have learned from. It is faster to hand it over. That is exactly the point.

The takeaway

Developers are still needed. That sentence is true and nearly useless, because it answers a question nobody is really asking. The measurements say the part of us that is needed is the part that chooses and judges, and the part that is already being absorbed is the part that types, and the gap between those two is where every junior used to be born.

As for how long, the only number trending in a straight line is the task horizon, it is doubling roughly every four months, and the people publishing it say the loop is not closed yet. Not yet is carrying a great deal of weight in that sentence. Watch that number rather than the predictions, mine included.

And when it arrives in your field, it will probably look like what it looked like for Duminil-Copin. Not a layoff. A publication, on a Tuesday, answering the questions you had been saving.