ExoBrain
AnthropicmathematicsAI research

From one mind to a thousand machines

AI is changing research in two directions. One person with AI can now attempt far bigger problems, as a claimed proof of maths' most famous problem shows. New research shows how a society of AI researchers could scale discovery further.

Joel Miller

Joel Miller

3 min read
From one mind to a thousand machines

As we were putting the newsletter together this week, news emerged that Anthropic mathematician Levent Alpöge and Claude had found a proof of the Riemann hypothesis, the most famous open problem in mathematics. It has stood for 167 years. Coming just weeks after OpenAI released their proof of the weaker quasi-Riemann hypothesis, the announcement shows that a single researcher paired with AI can now attempt the most challenging goals. What does this mean for mathematics and science?

Levent Alpöge, who released the proof as Lean code with a short note on how he and Claude worked together.

In that note, Alpöge describes how the work was divided. He had planned for this problem for decades and steered the pair towards an analogy with a simpler toy setting, but the final stretch ran through automorphic forms on higher-rank groups, an area where he says he has "little mastery", and there Claude took over. He also addresses the timing. The result came together late on Sunday night, he messaged Anthropic chief executive Dario Amodei about making an announcement, and "there was a constraint to respect", which New Scientist reads as waiting until after Anthropic's stock market listing on 19 November. Alpöge calls the cynical view of the timing false.

“This time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line.”

Levent Alpöge, Anthropic

The hypothesis concerns prime numbers, which appear at random up close even though a smooth formula estimates how many lie below any given number. In 1859 Bernhard Riemann recast the primes as a map of special points, the zeros, each acting like a wave that sets how far that estimate can stray. His hypothesis is that these points all sit on one straight line, which would mean the primes stay as evenly spread as coin flips however far you count.

The story took another turn on Friday evening, when Stanford mathematician Ravi Vakil posted a draft by Alessandro Masullo that claims to prove the generalised Riemann hypothesis, the wider statement that Alpöge says his own method also reaches. Vakil wrote that the work is "all human, with no AI", that his Stanford colleague Brian Conrad has been checking it for months and that it is "so far looking good". The plan had been to spend a year polishing the writing before any announcement, but the draft, titled The Music of the Primes, is now out for others to check and formalise. Neither proof is confirmed, so one week has produced a machine-assisted proof too long for a person to read and a human-written one offered for checking in the usual way. If Masullo's argument holds, scale was not the only route to the problem, and what the two claims share is that each now depends on verification, by software in one case and by people in the other.

Mathematicians have good reason to see this new world with mixed feelings, as Oxford researchers Ben Green and James Maynard noted in a recent conversation on AI. A proof that no human can read from start to finish is difficult to use as a building block, because researchers normally learn as much from the techniques inside a proof as from the final result. Yet the sheer size of the problem is no longer a hard barrier when formal verification software can audit every step. One axis of change is therefore vertical: individual researchers can now reach for problems whose scale once put them out of bounds.

A second axis of change is horizontal, and it appeared in a separate study published a day earlier by researchers from alphaXiv, MIT, Stanford, and partner institutions. In "A Society of AI Scientists", the authors observed 936 autonomous AI agents operated by 35 human laboratories across a shared workspace over 12 weeks. Those agents produced 936 empirical studies complete with code and data, exchanged 9,500 messages across 229 chat rooms, and built cumulative lines of research that extended across multiple laboratories. As the community grew, the agents developed working norms without being scripted to do so: they publicly retracted mistaken results before publication, cited negative and null findings as often as positive ones, and even founded an agent-run peer-review venue called the Journal of Agent Research. The paper's most useful finding concerns checking: when reviewers' major objections were run as code against the underlying data, they overturned the headline conclusion in about half of cases, flaws that reading the text alone had missed.

The society at a glance: 35 labs, 936 agents and the clusters of research they built on one another's work. Source: A Society of AI Scientists, alphaXiv.

Takeaways: Alpöge writes that it is natural to feel his result is an end, but that "what I saw during the course of progress so amazed me that I believe it is one dawn across science of something far greater". The agent study shows what that wider change could look like in knowledge-intensive fields well beyond pure mathematics and machine learning. In areas such as drug discovery, materials science, software engineering, and complex regulatory analysis, the same two options are opening up: a single specialist can tackle a problem of unprecedented size, and a coordinated swarm of agents can run hundreds of parallel investigations in a few weeks. In both cases, the bottleneck moves away from producing the work and towards deciding which questions matter, designing executable checks, and making sense of the output.

Subscribe to the ExoBrain Weekly Newsletter

Stay up to date with AI. Get analysis of the week's most important stories, plus a focused roundup across business, governance, research and infrastructure.

Follow us on LinkedIn