Implication: For technical research, the near-term opportunity may be to accelerate execution on bounded problems while keeping scientific judgment with domain experts.
Watch condition: Watch for independent replications that report comparable task difficulty, expert time, error rates, and validation outcomes.
Thesis
This case shows a domain expert using Claude as a supervised executor for a bounded, checkable theoretical-physics calculation, not as an autonomous scientist.
Observed facts
Mechanism
Schwartz deliberately selected a well-defined, graduate-level problem with established methods and a result he could check. That structure let him decompose the work into prompts, inspect intermediate outputs, and correct errors. The reported speedup therefore depends on expert task selection and continuous evaluation; it is not evidence that a model independently identifies worthwhile questions, validates its own reasoning, or completes open-ended research.
Why now
The account offers a concrete example amid claims about end-to-end AI scientists, and Schwartz explicitly argues that intermediate, supervised steps matter. Its most useful contribution is a description of a workflow and its limits, rather than proof of generalized scientific autonomy.
Implications
For technical research, the near-term opportunity may be to accelerate execution on bounded problems while keeping scientific judgment with domain experts.
Evidence boundary
These figures and judgments come from Schwartz’s first-person account published by Anthropic; the supplied source provides no independent replication or comparative evaluation.
Watch conditions
Watch for independent replications that report comparable task difficulty, expert time, error rates, and validation outcomes.