ZICQ
中 Log in / Sign up
ZICQ Info Research & Papers #Claude #Anthropic #AI for Science #Black-Hole Physics #Formal Verification

Claude Reports Solutions to Two Open Problems in Black-Hole Physics, With Validation Still Pending

Avatar of Mr.Xu

By Mr.Xu Community Post

Published:

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:Anthropic Research reports that Claude solved two long-standing open problems in black-hole physics. If independently validated, the result would be a significant AI-for-science case showing how a language model can help formalize problems, generate candidate derivations, and invoke external verification tools. The collected item, however, contains only the headline, a link to the first-party page, and Hacker News engagement figures. It does not identify the two problems, mathematical methods,


Claim and Evidence Boundary

The title published by Anthropic Research states that Claude has solved two open problems in black-hole physics. The collected material contains only that headline, a first-party link, and Hacker News engagement figures. It does not include the problem statements, paper version, authorship, proofs, numerical data, code, or peer-review status. The defensible conclusion is therefore that Anthropic has made the research claim, not that the two propositions have already been independently validated. Engagement metrics measure attention rather than scientific correctness.

If the results survive independent scrutiny, their significance would extend beyond supplying two answers. A language model would be participating in a research process that requires long-horizon reasoning, symbolic derivation, and numerical validation with mutually constraining checks. Generating a plausible solution and producing a verifiable solution remain fundamentally different achievements.

Technical Mechanism: What Counts as a Machine-Solved Open Problem?

Open problems in black-hole physics may take the form of proving a previously unknown general result, constructing a counterexample to an existing conjecture, or obtaining a new high-precision numerical solution satisfying the relevant field equations. The available extract does not identify which categories Anthropic addressed. The following describes the verification stack required for this class of AI-assisted research rather than claiming an undisclosed implementation.

A language model is trained with a next-token objective, which does not directly optimize proof obligations. It may produce internally plausible derivations while hiding invalid assumptions, omitting boundary conditions, or mistaking numerical agreement for an analytic result. A credible workflow needs at least four layers:

  1. Problem formalization: Translate the natural-language description into an explicit theorem, boundary conditions, physical assumptions, and falsifiable numerical targets.
  2. Candidate generation: Use the model to combine literature patterns, candidate mechanisms, derivation paths, and counterexample strategies. Multiple samples can increase diversity, but they also increase verification cost and may share correlated errors.
  3. Tool grounding: Check identities with a computer algebra system, explore parameter space with high-precision numerical solvers, and cross-check results with an independent implementation. Near black-hole singularities, catastrophic cancellation and ill-conditioned convergence make simple floating-point precision increases insufficient.
  4. Formal and expert review: For a theorem, state every assumption, quantifier, and degenerate case, then seek proof-assistant checking or independent mathematical review. For a numerical result, publish convergence behavior, error budgets, and reproductions from independent code.

For context, a general-relativistic solution must satisfy the Einstein field equation G_{μν} + Λg_{μν} = 8πG T_{μν}/c⁴, together with the relevant horizon geometry and null-geodesic conditions. These equations illustrate the required standard; they are not an assertion that they define the two problems announced by Anthropic.

Engineering Trade-offs

Long contexts can preserve papers and derivation histories, but they also increase retrieval noise, attention dilution, and inference cost. A production system must balance full-context reasoning against retrieval-augmented evidence selection and structured state memory rather than simply enlarging the context window.

Multi-model, multi-sample, and multi-tool verification can reduce one-off hallucinations, yet it multiplies generation and validation expense. For open problems, a small set of high-value candidates may be more efficient than indiscriminate sampling. Agreement among many samples is not independent evidence if all candidates share the same systematic bias. Symbolic computation offers exact identity checks, while numerical methods are better suited to parameter-space exploration; neither substitutes for the other. Closed APIs introduce version drift and weak reproducibility, so researchers should preserve model versions, prompts, retrieval snapshots, tool arguments, random seeds, and intermediate artifacts to turn a demonstration into an auditable research record.

Measured Evidence: What the Current Extract Supports

The extract supports three limited facts: Anthropic Research published the relevant title, the link points to its first-party page, and the Hacker News submission attracted substantial discussion. It does not establish whether the two problems were genuinely open, whether the solutions are analytic or numerical, which tools Claude used, whether the work underwent peer review, or whether another team can reproduce it.

A stronger evidence level would require precise problem statements and the relevant literature boundary; complete proofs, numerical data, or executable code; convergence and error analysis; tests covering limiting cases and counterexamples; and independent review by domain specialists or external teams. An official research blog is a high-authority first-party release, but it is not automatically equivalent to peer-reviewed consensus.

Developer Playbook

AI-for-science teams can operationalize this kind of result through an auditable verification pipeline:

  • Maintain a structured problem card containing assumptions, targets, known results, and pass criteria, preventing the model from optimizing toward a malformed proposition.
  • Record prompts, model versions, retrieved evidence, tool calls, random seeds, and intermediate derivations so every step can be replayed.
  • Recompute results with two independent numerical implementations, or with one symbolic and one numerical implementation, while varying mesh resolution, precision, and extreme parameters.
  • Include negative controls and adversarial counterexample searches to detect confirmation bias. Final promotion should require domain-expert sign-off rather than model self-evaluation alone.

The defensible present-day role for Claude is as a candidate-solution generator and validation orchestrator, not an unsupervised judge of physical theorems. Its longer-term engineering value depends on whether a one-off breakthrough can be converted into a repeatable, inexpensive, and auditable scientific-discovery pipeline.


Source: Hacker News AI Feed (2026-11-19)

— END —

Tags: #Claude #Anthropic #AI for Science #Black-Hole Physics #Formal Verification

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…