The Riemann Hypothesis and the Industrialization of Mathematics

It’s time for an emergency update. Levent Alpöge, working with Anthropic’s Claude, has announced a proof of the Riemann hypothesis (along with its generalized variant)–arguably the most important problem in mathematics. The operative phrase is working with. An equally important detail is the computing power behind that collaboration. In his usual fashion, he glosses over this, but the computation is not a small detail to be relegated to a footnote: It’s central to understanding how the proof was possible at all, and establishing the limits of what an individual can feasibly hope to accomplish with AI.

The default response, as seen online, is skepticism, despite Alpöge’s claim that the proof has been formalized in Lean. Independent verification is still pending, but, for all practical purposes, it’s done. Alpöge is an expert. He wouldn’t casually release a half-baked proof of the Riemann hypothesis. Presumably, he has been working behind the scenes for some time, checking everything, and waited until he was confident it would hold up before going public. He also works at Anthropic, a company with enormous computing resources, although he hasn’t disclosed how much of that infrastructure was used. If the Lean formalization checks out, with no unproved assumptions and the correct theorem statement, there is little room left for the usual skepticism. My expectation is that it will.

The rise of AI-assisted mathematics has produced a range of emotions, from elation to anomie or even existential dread, along with endless think pieces and hand-wringing about the future of the field. With this discovery, assuming it holds up, the debate is effectively over in one sense: mathematics has been forever changed. At this point, it’s reasonable to assume that, given enough computing power, resources, and expert guidance, AI will eventually solve most, if not all, of the remaining important open problems. Mathematicians now have to decide whether it’s worth devoting years of their lives to a problem when they risk being scooped by AI. Or whether they can ever hope to make a meaningful contribution to mathematics when overshadowed by enormous AI-assisted discoveries by people and firms with vastly more computing power.

The other problem is that almost any individual contribution looks quaint by comparison. Compared to proving the Riemann hypothesis, what is left? Some say mathematics is more than solving problems or proving or disproving conjectures, but those are also what get all the attention. So how are we supposed to reconcile these two competing narratives — that mathematics is about some higher pursuit of understanding or truth, when solving hard problems is really what matters as far as careers or recognition are concerned?

I discussed this dilemma with ChatGPT weeks ago, below:


Discussion with ChatGPT about AI and mathematical research

(I’m glad I wasn’t working on the Riemann hypothesis–I spent maybe an hour total on it. Otherwise, I’d be pissed having wasted all that effort for nothing.)

You don’t get credit for an unsuccessful approach, nor can you recover the lost time. It’s all or nothing, unlike, say, patents, which can still be sold after a business fails. An approach that almost proves the Riemann hypothesis but ultimately fails gets little or no credit unless it produces some independently publishable result. (Note: this is not the same as the proof that over two-thirds of the nontrivial zeros are on the critical line, which is a finished theorem in and of itself, not a failed attempt at the Riemann hypothesis.)

To address earlier about feasibility, it’s not as if any individual could reasonably have done this on the same scale. Alpöge credits Claude, but he makes it sound as if he used an ordinary AI subscription, like anyone else. He glosses over the enormous computing resources that must have been involved. The final output is 17.4 million lines of Lean across 175,187 files, plus an unpublished AI-generated draft of over 250,000 pages. How much computation was involved? How much of the work was done in parallel? A 250,000-page document can’t fit into a model’s context window, so it had to be divided into thousands of chunks, with some way of managing dependencies and retrieving earlier results. What specialized infrastructure and tools did he have? These are hardly minor details.

To use myself as an example, and to give an idea of what is involved for a much smaller (albeit still big) problem, my recent result took nine days of computation for the second part and another week for an earlier part. The Lean formalization will take a month, minimum, with each iteration taking 10–30 minutes between rewriting the proof in Lean and compiling it on my end. The bottleneck isn’t how long it takes Lean to check a finished proof; instead, it’s how long it takes to produce a proof that Lean accepts in the first place, given all the failed attempts, revisions, and recompiling.

His project is vastly larger than mine, and going from a draft to a claimed complete Lean formalization in six weeks suggests extensive parallel computing and resources, well beyond a single computer and subscription. For my paper, I was using a single subscription each to Claude and ChatGPT on the same laptop ($200/month between them), for a total of $1,300 spent on subscriptions and extra “boosts” to use the most powerful models available. This pushed the limits of what a consumer could reasonably hope to achieve with these models at home. I’m guessing the amount of computation involved in his proof was at least 1,000 times what I used, if not considerably more. So yes, Claude helped, as he says, but that tells us almost nothing about the resources involved or how people actually use Claude.