Today is day three of the post-RH era. I wish to expand on Friday’s thoughts.
Overall, I am a fan of AI-assisted mathematics, but it’s imperative that authors provide full disclosure, including the models used, the division of labor between human and AI, chat logs and prompts, and usage tiers (e.g., Pro vs. free Claude, or a single agent vs. multiple sessions and browsers). AI-assisted mathematics makes math less of a purely theoretical subject and more of an empirical one, like a hard science. Therefore, AI usage should be fully disclosed in a “Methods” section, as would be expected in any scientific paper. I think this is the most honest compromise.
In a widely shared post, Terence Tao was way off in his forecast of AI proving the Riemann hypothesis–off by 24 years. I would have guessed 2030 if asked the same question a month ago. To look at it charitably, maybe he meant 2050 with a single prompt for an end-user of a single-instance commercial AI product, not a multi-agent array of an unreleased model?
Only a month ago:
Terence Tao said there is something beyond AI needed to solve the Riemann hypothesis.
"So maybe some combination of humans and AIs in 2050 or something will make further progress. Who knows?" https://t.co/Tf4sUXly4x
— sina (@sinavaziri) November 20, 2026
By mid-2026, an inflection point had been crossed, with AI helping to disprove the Jacobian conjecture in dimensions three and higher, a problem that had resisted human mathematicians for 87 years. Between 2023 and 2026, AI went from merely assisting mathematicians with calculations or checking proofs–to producing original solutions to longstanding open problems. In this sense, AI had arguably surpassed human mathematicians.
So in just two years, AI went from a silver medal at the Math Olympiad to solving what was arguably the most important open problem in all of mathematics, the Riemann hypothesis.
A point of contention is how much credit Levent Alpöge deserves for the result. How much was his intellectual contribution, versus AI’s? In his usual fashion, he is cagey, failing to disclose prompts or other pertinent information (similar to his other announcements), so we’re left to infer what he could have known with the scant information he provides. If I had to guess, AI did ~100% of the intellectual work. He may have fed it a concept or a paper idea, and then somehow Claude (presumably with many parallel agents, unreleased models, unlimited time, unlimited budget, unlimited compute etc.) did the rest and produced a huge document, which he quickly compiled to Lean (also using parallel resources).
I have also heard him described as brilliant and worthy of a Fields Medal. This is disputable. The vast majority of his papers are large collaborations, and his only solo postgraduate paper is in the Journal of Number Theory–a publication so dreadful its very name serves to remind anyone unfortunate enough to read it that they are reading a number theory paper.
Just reading his statement suggests he fed a paper to Claude and then Claude did the rest:
Tens of such occurred, but I will give one line which will suffice to make fellow mathematicians fall out of their chairs. In Stepanov’s argument resolving Riemann over function fields one uses a divisor involving multiplicity. This argument is a descendant of Stepanov’s, but multiplicity is hard to find over the integers, corresponding to something like differentiating whole numbers. The simple, new, Fourier-analytic ideas in the toy setting led us to examine chaining theta lifts if we were to have the same tricks recur but work in a
natively automorphic manner. The line: for the von Mangoldt sum of length X, Claude’s suggested length of chain of theta lifts, in view of each mimicking multiplicity, was √ X. (And yet, what else could it be?)
From what I can infer, there was no specific epiphany on his part that led to the proof. Again, without chat logs, we are left to speculate. For my own problem I fed it a paper (my hope was that the AI would just resolve it on the spot, but it failed). I also made a distinct novel contribution to the problem, which turned it from checking infinitely many cases to a bounded set of cases, effectively solving it. Just feeding it the paper would not have been good enough; it merely served as a starting point.
Upon further reading his statement, especially when he says “with Claude,” one may be under the mistaken impression that he used a single commercial Claude instance, like a regular end user, to produce his 250,000-page write-up and Lean proof. This is extremely implausible. The Lean formalization alone would be extremely time-consuming in its own right. To use myself as an example again, formalizing my much shorter proof may take between one and two months on my single machine. Rewriting a proof in Lean requires repeated generations, corrections, and debugging, all of which rapidly drain usage. Indeed, I also quickly used up most of the usage allotment on my “Ultra” Claude plan, despite its much higher limits than the standard paid tiers, and getting the message, Couldn’t switch models. You’ve hit a limit that covers that model. View usage.
For him to accomplish all of this in just six weeks for a vastly larger proof is breathtakingly fast, suggesting extensive automation or parallel computing resources, which he fails to disclose. How many agents, sessions, machines, or computing resources were involved? Without knowing this, it’s impossible to compare his results with what an ordinary AI user could accomplish. I think the honest framing of the result is something like, “He took the initiative, with Anthropic, to prove the generalized Riemann hypothesis.”
Yes, it may sound pedantic to litigate over credit for mathematical results, but when a result is monumental, who or what gets the credit becomes part of the story, and it matters.