Claude solves two open problems in the physics of black holes

Anthropic claims its Claude AI has autonomously solved two open problems in black hole physics, prompting both excitement and skepticism about what was actually achieved. Commenters note the results are in highly idealized, non-our-universe settings and resemble advanced mathematical physics more than empirically testable breakthroughs, while raising concerns about authorship, lack of prompt transparency, and the use of such work as PR ahead of an IPO. Several physicists stress that the true significance will only be clear if the wider research community engages with, cites, and builds on these papers.

Scope and significance of the results

  • Many note the work is in highly idealized settings (e.g., 2D universes, bootstrap bounds, not our universe directly) and see this as standard for theoretical physics, akin to Schwarzschild black holes or “in vitro / in mice” results.
  • One technical commenter highlights:
    • First paper = difficult, error‑prone calculation about black hole phase transitions.
    • Second paper = upper bound on black hole mass compatible with a given central charge, i.e., a bootstrap consistency bound in 2D quantum gravity.
  • Others stress that only community uptake (citations, integration into later work) will show whether the results are truly important; there are countless “open problems” in theory.

Physics vs math; rigor vs usefulness

  • Extended discussion about how physicists use mathematically non‑rigorous tools: divergent series, zeta-function regularization (1+2+3+… = −1/12 context), Dirac delta, path integrals, renormalization.
  • Some argue this makes formalization in systems like Lean difficult; others counter that historically “dodgy” techniques often later get rigorous foundations.
  • Contrast is drawn between mathematics (axioms, theorems, formal verifiability) and theoretical physics (models judged by empirical adequacy and internal consistency).

AI’s role, authorship, and anthropomorphizing

  • Debate over saying “Claude solves” versus crediting humans who prompted, steered, and verified the work, plus those who trained the model and supplied data.
  • Some view omitting human co-authors as academically dubious; others think credit should go to the organization and tool explicitly.
  • Strong calls to publish prompts, interaction logs, and code; pushback notes that real workflows involve gigabytes of evolving context, many dead ends, and non‑deterministic runs.
  • Disagreement over whether anthropomorphizing (“escaped the sandbox”, “Claude did X”) is misleading or just normal shorthand for tool behavior.

Empirical science vs theoretical / AI capabilities

  • One side argues AI already does well at empirical pipelines (data wrangling, experiments by proxy via simulations) when data exists.
  • Others emphasize that designing good experiments and collecting high‑quality data is deeply intellectual, non‑rote work; much experimental labor (e.g., dissections, complex lab protocols) is highly skilled and hard to automate.

PR, timing, and community reaction

  • Many see the announcement as IPO‑adjacent marketing and “AI PR,” question whether problems were cherry‑picked low‑hanging fruit, or worry about undisclosed use of others’ research.
  • Some are impressed or intrigued by the capabilities; others express strong skepticism or hostility toward corporate AI’s impact on science and attribution norms.