Unreleased Anthropic AI model advances Riemann hypothesis
An unreleased Anthropic AI model has made significant progress on the Riemann hypothesis, demonstrating that automated agent networks can tackle some of the world's hardest scientific problems.
Anthropic has revealed that an unreleased artificial intelligence model successfully advanced work on the Riemann hypothesis, a legendary 150-year-old mathematical problem that carries an unclaimed $1 million bounty. Remarkably, the breakthrough was initiated by an Anthropic employee who lacked advanced mathematical training. The staff member simply prompted the system to "take a real stab" at the proof, leaving the model to autonomously coordinate the complex project over a day and a half.
To achieve this, the AI orchestrated a massive network of 60 subagents that tested 650 distinct ideas and consumed 31 million output tokens. Within this digital hierarchy, two subagents generated the core mathematical concepts, 13 contributed supporting ideas, 30 tried but failed to produce new concepts, 13 acted as validators to verify the logic, and two drafted the final paper. Two of Anthropic's internal mathematicians verified the results, which were then formalized using Lean, an open-source proof assistant. The model successfully raised the lower bound of solutions for which the hypothesis is known to be true.
This achievement follows other recent AI-driven mathematical milestones, including the resolution of several Erdos problems and Anthropic's disproof of the Jacobian conjecture. Additionally, OpenAI recently showcased ten major proofs generated by its internal Astra model. While these advancements excite many, they have also sparked debate. In June, a group of prominent mathematicians signed a declaration warning that AI could undermine the field's standards of personal attribution and responsibility. Conversely, Fields Medalist Timothy Gowers suggested that mathematics might evolve to a point where theorems are no longer tied to human authors.
For AI practitioners and software engineers, this development signals a shift from single-prompt LLM interactions to highly coordinated, multi-agent systems capable of long-horizon reasoning. By demonstrating that a non-expert can guide an AI swarm to make genuine scientific progress, the experiment suggests that future enterprise workflows will rely heavily on autonomous agent orchestration to solve highly specialized, complex problems across various industries.
This is our own summary of reporting by TechCrunch AI
