Matthew Schwartz reports finishing a theoretical-physics calculation in two weeks instead of the usual year with Claude Opus 4.5. The stronger lesson is conditional: the target had established methods and numerical checks, and Schwartz supplied the judgment to guide and verify the work.
Summary
Harvard physicist Matthew Schwartz says he supervised Claude Opus 4.5 on resumming the Sudakov shoulder in the C-parameter distribution, completing the work in two weeks rather than the usual year. His first-person account reports more than 110 drafts, 36 million tokens and over 40 hours of local CPU compute. These are Schwartz’s reported figures, not independently verified measures of model performance; the case concerns a bounded calculation with expert oversight, not autonomous scientific discovery.
What happened
The technical target was the Sudakov shoulder in the C-parameter distribution. The dossier’s summary of the resulting paper says it presents a next-to-leading-logarithmic resummation using soft-collinear effective theory, with a new factorization treatment and one-loop jet and soft functions. It also reports numerical checks against Catani–Webber results and EVENT2. Schwartz’s account says he chose a graduate-level problem with established methods and a result he could check, then supervised Claude through the calculation. According to the account, that work took two weeks rather than the usual year and involved over 110 drafts, 36 million tokens and more than 40 hours of local CPU compute. The timeline, comparison and quality claim come from Schwartz’s account published by Anthropic; the supplied evidence does not independently verify them.
What the archive adds
The archive’s earlier Anthropic account, “Claude-shaped science,” describes Schwartz’s prior use of Claude as a research assistant as difficult: he had to correct outputs and redirect the model. The dossier’s workflow summary for this experiment adds specific practices—sequential small tasks, persistent Markdown summaries, code execution, numerical regression and cross-checks. Together, the accounts suggest a change in how the work was structured, not proof that supervision or correction disappeared.
Trend signal
AI-assisted research is most persuasive when a claim can be decomposed into intermediate steps and checked against established methods or numerical results. Here, the calculation’s technical structure supplied possible validation anchors; that makes the report more informative than a demonstration judged only by a polished final paper. But a checkable calculation is not the same task as finding a valuable question, and one first-person account cannot establish that the reported acceleration generalizes.
Why now
Three plausible drivers emerge from the evidence. First, Schwartz selected a problem with established methods and checkable results, making expert evaluation possible (primary account). Second, the workflow used sequential tasks, persistent summaries, code execution, numerical regression and cross-checks, according to the dossier’s summary of the account (d3). Third, the technical target had validation anchors: the paper summary describes a factorization treatment, one-loop jet and soft functions, and numerical checks against Catani–Webber results and EVENT2 (d2). The evidence does not isolate how much each contributed to the reported timeline.
Connections to other events
The result’s practical significance rests on a constraint: Schwartz could select a problem whose methods he knew and whose outputs he could assess. The paper summary identifies concrete validation anchors—its factorization treatment, one-loop functions and numerical checks—while the workflow account describes sequential prompting, persistent summaries, code execution, regression and cross-checking. Our read is that Claude may have accelerated execution within this structure; the evidence does not show that it independently chose the target or validated the science. The many drafts also mean throughput should not be confused with reduced review burden.
Why it matters
Research groups can test this approach on calculations with known methods, intermediate outputs and independent checks. Require generated code and derivations to remain inspectable, and track expert review and correction time alongside calendar time, token use and compute. A faster completion is valuable only if the final result passes the field’s usual validation. The account suggests that researchers who can frame and judge such tasks may be best positioned to benefit; it does not establish how broadly the method transfers.
What happens next
- At least one independent group will publish an AI-assisted theoretical-physics calculation with an explicit account of validation and human effort. (12 months, 45%). Watch for: A group outside Anthropic publishes a comparable workflow, identifying its target, checks and expert review or correction time.
- Public reporting on AI-assisted research will increasingly include intermediate checks and human oversight alongside final outputs. (6 months, 60%). Watch for: A new research account or paper documents staged tasks and validation steps, rather than presenting only a final result.
Limits and uncertainty
The two-week timeline, comparison with the usual year, resource figures and claims about the result come from Schwartz’s account published by Anthropic; the supplied evidence offers no independent replication. It gives no comparable breakdown of the conventional baseline, expert hours, error rates or total validation effort. The paper summary identifies numerical checks, but does not independently establish that the reported workflow caused the speedup. The inference could be wrong if this problem was unusually suitable or if the baseline is not comparable.
Skill to build now
Practise designing a checkable task: split one bounded calculation into steps, define an independent test for each intermediate result, then log model errors, corrections and your review time as you work through it.
Reader takeaway
This week, pick one bounded calculation and write down the expected intermediate outputs and an independent check for each before asking a model to attempt it. Require inspectable code and results; log corrections, your review time and compute. Compare the finished work against the same validation standard you would use without AI.
Open question
When the expert must frame and validate the calculation, what evidence would show that the time saved exceeds the effort spent supervising it?