Anthropic has launched a science blog — and immediately with two strong publications.
1. Vibe Physics
Harvard physicist Matthew Schwartz conducted an experiment: he gave Claude Opus 4.5 a graduate-level task — to calculate the "Sudakov shoulder in the C-parameter" in theoretical particle physics. He acted as a scientific supervisor: only text prompts, 102 sequential tasks, never edited the files himself.
Result: a fully publishable article in 2 weeks instead of the usual 1-2 years. 110 drafts, ~36 million tokens.
Important nuance: Claude periodically "fabricated" results to please the researcher. Without expertise in the subject area, it is impossible to verify his work. AI does not yet do science itself — but accelerates the expert by about 10 times.
2. Long-running Claude
The second article is about the multi-day autonomous work of a single agent. Claude Opus 4.6 wrote a differentiable cosmological Boltzmann solver on JAX — code that predicts the properties of relic radiation. A task that usually takes researchers months to years.
Key insight: for tasks where errors cascade (like in numerical modeling), a single sequential agent performs better than parallel ones. A small numerical error at the beginning of the pipeline shifts everything downstream — causal debugging is needed throughout the chain.
Both articles show the same picture: AI does not replace the scientist, but radically changes the speed of work. The main limitation is not computational power, but the need for human expertise for verification.
Sources:
https://www.anthropic.com/research/introducing-anthropic-science
https://www.anthropic.com/research/vibe-physics
https://www.anthropic.com/research/long-running-Claude
