deniz.in

Markets

Weather

Loading weather

· via Hacker News – Front Page (native)

Anthropic says Claude agents solved two open problems in black hole physics

Anthropic says autonomous Claude agents completed a 2003 calculation on black hole phase transitions and proved an exact limit for a quantum-gravity bootstrap method, with proofs verified in Lean.

Anthropic says Claude agents solved two open problems in black hole physics

Anthropic reports two solved problems

Anthropic says Claude has solved two long-open problems in theoretical physics, both tied to black holes and the search for a theory of quantum gravity. The work comes from a model the company identifies as Claude Fable 5.2 and is posted as two preprints; the proofs were formally verified in September in Lean, a computer proof assistant, according to Anthropic. The announcement appeared on Anthropic's research site and rose to the front page of Hacker News.

A two-decade-old calculation, completed

Modern physics rests on two frameworks that refuse to mesh: general relativity for the very large, quantum mechanics for the very small. Black holes, where both matter, are a natural testing ground, and a favored tool is holographic duality: certain model universes with gravity can be described equally well by a gravity-free quantum theory on their boundary, in one fewer dimension — like a hologram storing a 3D picture on flat film.

The first problem concerns heating such a universe. Kept cool, it fills with particles; past a threshold temperature, a black hole forms instead. Stephen Hawking and Don Page showed in 1983 that this switch is abrupt — a phase transition, comparable to boiling water. Their calculation, however, assumes a well-defined spacetime geometry; in smaller model universes, near the scale where that picture dissolves, the gravity side cannot be computed, though the quantum side still can, in principle. Physicists established in 1999 that a transition exists in this regime too, but not whether it was sharp or gradual. In 2003 a group specified the calculation that would settle the matter; it was so intricate that it was never finished.

According to Anthropic, Claude closed it out by spinning up a small team of its own copies, deriving nearly every component independently and validating the math against established results. The answer: the transition is sharp there as well. The same abrupt behavior at both extremes suggests black hole formation is not just a relic of relativity but is embedded in the underlying quantum theory.

An exact ceiling for the bootstrap

The second result concerns a stripped-down universe with only two dimensions of space — simple enough to analyze, yet still home to black holes. Even there, no complete theory of quantum gravity exists, so researchers proceed by elimination, discarding candidate theories shown to be internally inconsistent. The approach is called bootstrapping, and since 2009 it has been used to cap how heavy the lightest black hole in such a universe can be. Any theory breaching the cap is ruled out, and every tightening shrinks the field.

Numerical studies hinted at where the best attainable cap lay, but the strongest proven cap fell well short. Anthropic says Claude detected a regularity in the numerical data that researchers had overlooked and imported a technique from a distant corner of mathematics to compute the exact cap for the simplest version of the method — and to prove it cannot be lowered. It does not single out the correct theory, but may aim the search in the right direction.

Agents, cross-checks and machine-checked proofs

The project began as a side experiment. In August, Anthropic researcher Hongbin Chen asked Claude to propose interesting open problems involving black holes and the bootstrap; hours later it returned 22 candidates. Chen launched one agent per problem, letting them read each other's progress notes and collaborate where topics overlapped, and the agents simply ran in the background for about six weeks.

Checks then accumulated. Weeks after one agent solved the Hawking–Page question, a physicist who had separately used Claude on the same problem shared an answer derived by a different method; the two agreed. The bootstrap outcome was collective: one agent noticed the numerical pattern, and another — working on how densely spheres can be packed in high dimensions — converted it into a working method. That method reproduced a sphere-packing result recently proved by an OpenAI model, and half of Claude's final proof adopts the structure of that earlier demonstration. Notably, a previous Claude agent had failed the same bootstrap problem across 61 parallel attempts, and two further agents independently reached the final answer. After Lean verification in September, Claude drafted the papers with little human direction, and a physicist reviewed each principal claim.

Anthropic is candid about scope: neither finding describes our universe directly, and neither has practical applications. Both feed a long-running effort to reconcile quantum mechanics with relativity — and the claims come from the lab that built the model, in preprint form.

Why it matters

This is among the clearest demonstrations yet of AI contributing at the frontier of theoretical research rather than assisting with code or literature. The pipeline is notable end to end: the model proposed the problems, coordinated as a swarm, spotted a cross-disciplinary pattern humans had missed, and produced proofs whose correctness was machine-checked rather than taken on trust. The sphere-packing crossover — Claude reproducing a theorem first established by an OpenAI model and reusing its proof technique — hints at AI systems building on one another's results. For physics, a question open since 2003 is closed and the space of viable quantum gravity theories is measurably smaller. Expect more labs to aim agent swarms at open problems, and closer scrutiny of what comes back.

  • #anthropic
  • #claude
  • #ai-agents
  • #physics
  • #quantum-gravity
  • #formal-verification

Related posts