Will physics be the next field to be turned upside down by AI? — SkimNews
Get the Health newsletter
Daily health & science — research, biotech, public health, the studies worth knowing. Free.
- John Sous at Yale found six physics benchmark tests contained errors and ambiguous definitions; correcting them pushed AI model scores from as low as 30% to as high as 94%.
- Sous argues AI will excel at any "well-defined" physics problem but may struggle with creative, model-building work, calling physics's open-ended frontier "a very artistic, a very creative process."
- Logan Wright's Yale lab is using Anthropic's Claude to control laser pulses on a rubidium atom, with the AI predicting outputs, running tests, and gathering data without human intervention.
- Wright says the AI-controlled experiments can compress a week of human work into a single overnight run, and researchers can submit ideas from a phone app while away from the lab.
- The article attributes AI's math dominance to formalization — turning theorems into verifiable code — a technique still in early days for physics, where problems are messier and harder to pin down.
- Wright raised the philosophical stakes, saying AI could soon design an experiment humans don't understand and ask them to build it, questioning whether incomprehensible results still "count as science."
Why it matters: The benchmark correction at Yale suggests published AI-versus-physics comparisons have systematically understated model capability, while Wright's rubidium-atom experiment shows AI is already moving from code-writing assistant to overnight experiment operator in real labs — making formalism, not raw model power, the near-term bottleneck for AI physics.
Ask SkimNews
