Anthropic says Claude solved two open problems in black hole physics, with no named physicist credited on the papers
원문을 읽지 못해 Hacker News 토론으로 요약했습니다
- Anthropic's blog post, as titled on Hacker News, claims Claude solved two open problems in black hole physics, though the article body could not be read.
- A commenter quotes the announcement saying neither result describes our universe directly and neither has practical applications, but both feed a long-running effort to reconcile quantum mechanics and relativity.
- Commenters report the Hexagon page says the main result was first obtained by an internal Claude research model, and the public Claude Opus 5.5 later reached the same result independently.
- Commenters say one problem was solved with Lean, and the other was not disclosed. Neither the paper nor the blog post credits a human author beyond Claude and Anthropic.
Hacker News 의견들
Those theoretical physics papers look like the real thing, not the math physics stuff like the OpenAI Yang-Mills result. Feels like AI is finally branching out of math.
The announcement says neither result describes our universe and neither has practical applications. Honestly that's normal for theory work, since models always assume symmetries and homogeneity.
The bootstrap paper is brutal to read, but the blog post is easier. The first result is a messy, error-prone calculation. The second is the cool one: Claude apparently proved an upper limit on black hole mass tied to a central charge, but only in a 2d universe.
Careful with 'consistent theory of quantum gravity.' Bootstrap only shows a not-inconsistent theory. Other methods could still find an inconsistency.
There's no human author credited on the Hexagon entry or the PDF, only Claude and Anthropic. Isn't that new even for papers that are mostly AI output?
Hexagon's who page says the main result was first found by an internal Claude research model, and the public Opus 5.5 got the same result later. A physicist gave brief guidance and checked the main results, supposedly. No peer review though, so I'd take that with a grain of salt.
'Claude solves' is the usual anthropomorphizing. Someone set up the question and let it iterate until it got an answer. The word prompt doesn't show up anywhere in the blog post, and that's the part I'd want to see.
I think AI papers should always include the prompting history as an annex. Show your working, like we made kids do.
Someone in the thread says one result was done in Lean and the other wasn't disclosed, so I'd guess that's the gap in the claims.
Given the timing with the IPO, I'd say it's both IPO fodder and substantial.