AI Card Bao WWW.AICARDBAO.COM

Industry News

22 Claude agents were idle for six weeks and returned two unsolved problems in black hole physics.

It started as a side project.

In August this year, Hongbin Chen, a researcher at Anthropic, asked Claude: What are the unanswered questions in black hole physics? Claude listed 22. He assigned an agent to each question and let them run on their own in the background, and each could see the other's progress posts.

Six weeks later, two of them turned in their answers.

The first one is a problem that physicists formulated in 2003 but no one has ever finished. A model universe is heated to a certain extent to form a black hole. The question is how this transition occurs. Claude showed that it is a mutation, like the boiling of water, and this holds true even when space has no definite shape.

The second one is more twisty. A previous Claude agent tried this problem 61 times in parallel and failed. Half of the final successful proofs followed the idea of ​​a proof given by the OpenAI model.

Verification is two-tiered: the results first pass through Lean, a computer proof checker, and then a physicist checks each major conclusion.

Both papers are currently preprints and have not been peer-reviewed. Anthropic themselves have said that neither result is of practical use.

Solved 2 out of 22 problems, a hit rate of 9%. This number is much more honest than the statement "AI can overcome physical problems". It is closer to the real state of scientific research: most directions are unavailable, and occasionally one or two are successful. The difference is that the cost of trial and error went from years of graduate school to weeks of computing power.

The experience with the second question is particularly illustrative of the current situation. All 61 independent attempts failed, indicating that relying on brute force to achieve parallel heaping did not work; in the end, we borrowed ideas from other models to get through. Models have begun to cite each other's proofs.

One more detail: the list of questions itself was written by Claude. Picking out from the 22 which ones are worth doing and which ones have actually been solved by someone before, no one has participated in this step. As for whether any of these 22 are fake questions made up by itself, there is no mention in the public information.

Related