In six weeks, Claude derived an exact c/9.1266 bound in black hole physics
Anthropic had several copies of Claude spend six weeks working through unsolved problems in black hole physics, and the model picked the 22 questions itself. In one problem it derived an exact bound of c/9.1266 for the lightest quantum state, where c is the "size" of the theory. The best previously proven estimate had a coefficient of about 8.5. The lesson for agent developers: a long task is safer to hand to a group of copies with independent verification.
AI-processed from Habr AI; edited by Hamidun News
Anthropic reported that several copies of its Claude model worked for about six weeks on unsolved problems in theoretical physics and came back with two preprints on black holes. The model itself compiled the list of 22 problems. In one of the papers, it derived an exact bound of c/9.
1266. This is a limit on the "weight" of the lightest nontrivial quantum state in a simplified model of gravity, where c (the central charge) measures the size of the theory. Until now, physicists could rigorously prove only a coarser estimate with a coefficient of about 8.
5, even though computer calculations hinted at a tighter limit. Imagine spending years proving that a service's latency is "no worse than X," and then someone derives the exact formula and confirms it in two ways. For developers, the interesting part is not the physics but the setup: how to run long autonomous agent work so that the results can be trusted.
How the experiment was set up
According to Anthropic, it all started with a simple request. Researcher Hongbin Chen asked Claude to compile a list of unsolved questions in his field. The model proposed 22 problems. Over the next six weeks or so, several copies of it worked in parallel, exchanged intermediate results and checked each other's calculations.
Verification was built into the process. In the first problem, almost every stage of the calculations was done in two independent ways: if two paths give the same answer, an error is unlikely. In the second problem, the results were additionally checked in Lean. This is a formal verification system that accepts a proof only if every step is logically correct.
Both studies rely on holographic duality. The idea is that some model universes with gravity can be described by an equivalent quantum theory without gravity on their boundary. Properties of spacetime can then be studied through the behavior of particles.
What exactly Claude found
The first problem concerns the Hawking–Page transition, described in 1983. In a special model universe, anti-de Sitter space, a gas of particles gives way to a state with a black hole as it heats up. The change is abrupt, like water freezing: there is no smooth transition. It was unknown whether such a jump persists in the quantum regime, where the familiar geometry of spacetime no longer works.
Back in 2003, physicists proposed a calculation that could answer this question, but it proved too complex. Claude carried out the calculations in supersymmetric Yang–Mills theory and found that the transition is also abrupt when the particles interact weakly. This supports the hypothesis that the two phenomena share a common nature. A continuous connection between the regimes has not yet been proven.
The second paper deals with the modular bootstrap. This is a mathematical method for testing theories of quantum gravity for internal consistency: it lets researchers discard models that violate certain symmetry conditions without knowing in advance what the correct theory looks like. The best previously proven estimate in the simplified three-dimensional model was set by a coefficient of about 8.5.
Claude spotted a pattern in the numerical results and applied the machinery of spectral curves. As a result, it analytically derived an exact bound for the lightest nontrivial state: at large central charge c, it equals c/9.1266. The authors also showed that the chosen variant of the method cannot improve this bound. Other variants of the bootstrap may yield tighter constraints.
What this means for people building agents
First: long-running agent work has become realistic. Six weeks of parallel work by several copies is no longer like a short chatbot session. Second: Claude did not just compute but also chose what to work on, compiling the list of 22 problems itself.
Third, and most practical: reliability came from verification, not from a single strong model. Calculations done two ways and formal verification in Lean produce a result that does not have to be taken on faith. Product development works the same way. If an agent writes code, migrations or financial calculations, add an independent second path of verification: tests, recalculation by a different method, a formal checker where one is possible.
Limitations
Both results exist only as preprints and have not passed independent scientific review. Until physicists have examined them, it is too early to talk about closed questions. In addition, all data about the experiment comes from Anthropic itself. It is hard to judge from the report how much human involvement there was in setting the problems and reviewing intermediate results.
Still, the picture is telling: AI obtained results that physicists could not derive for years. If the results survive reviewers' scrutiny, lab agents will become a working tool of science. The "parallel copies plus independent verification" scheme can be tried on your own tasks right now.
- Nov 20, 2026Nvidia Is Close to Buying Inception for $500 Million to Get Diffusion LLMs
- Nov 20, 2026Microsoft puts a dollar value on AI agent work across 6000 task types, but it's only an estimate
- Nov 20, 2026The /gitlab-duo-setup Skill Connects Claude Code to GitLab Duo With a Single Command
- Nov 20, 2026Nvidia's stakes in startups grew from $22 billion to $65.5 billion, but without Huang the deals are in question
Key Questions Answered
How the experiment was set up?
According to Anthropic, it all started with a simple request. Researcher Hongbin Chen asked Claude to compile a list of unsolved questions in his field. The model proposed 22 problems. Over the next six weeks or so, several copies of it worked in parallel, exchanged intermediate results and checked each other's calculations. Verification was built into the process. In the first problem, almost every stage of the calculations was done in two independent ways: if two paths give the same answer, an error is unlikely.
What exactly Claude found?
The first problem concerns the Hawking–Page transition, described in 1983. In a special model universe, anti-de Sitter space, a gas of particles gives way to a state with a black hole as it heats up. The change is abrupt, like water freezing: there is no smooth transition. It was unknown whether such a jump persists in the quantum regime, where the familiar geometry of spacetime no longer works. Back in 2003, physicists proposed a calculation that could answer this question, but it proved too complex.
What this means for people building agents?
First: long-running agent work has become realistic. Six weeks of parallel work by several copies is no longer like a short chatbot session. Second: Claude did not just compute but also chose what to work on, compiling the list of 22 problems itself. Third, and most practical: reliability came from verification, not from a single strong model. Calculations done two ways and formal verification in Lean produce a result that does not have to be taken on faith. Product development works the same way.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.