Live·Open questions in longevity research
All news
Scientific ComputingScience Research

AI system solved 49 of 50 theoretical physics calculation problems; scientists in other fields see barriers in the lack of automated verification, laboratories, and sufficient data, Scientific American reports

13 November 2026· 261113002

AI system solved 49 of 50 theoretical physics calculation problems; scientists in other fields see barriers in the lack of automated verification, laboratories, and sufficient data, Scientific American reports

On November 12, Scientific American published an analysis by journalist Dan Garisto exploring whether the changes AI has brought to mathematics will extend to other sciences. The researchers interviewed disagree in their assessments, but Garisto says they all agree that science faces major changes.

In May, OpenAI models working under human guidance disproved a conjecture by the mathematician Erdős. In October, the company released hundreds of AI-written preprints (papers that have yet to undergo peer review) on unsolved problems. Better models explain only part of this progress, Garisto writes. The other reason lies in the nature of mathematics: exact solutions exist, and the Lean software can check proofs written in its language even when no one fully understands the mathematics behind them. “In the sciences, we have no such verifier,” says Anoop Krishnan, an engineer at the Indian Institute of Technology Delhi who studies AI.

Theoretical physics could be next, the magazine suggests. In a preprint, 34 authors assembled 50 unpublished calculation problems. Each extends a known result using established methods and has an answer, a formula or a number, that can be checked. The best of the five systems tested, OpenAI’s GPT-6-Astra, solved 49. The authors note that a verifiable calculation comes late in the research process: first, a physicist develops a theory and asks the right questions. In their experience, AI still lacks “taste,” the shared sense among physicists of what makes a problem interesting. Humans therefore set the direction of research.

Harvard theoretical physicist Matthew Schwartz, whom the magazine identifies as a consultant to the AI company Anthropic, and his coauthors released 36 preprints on problems suited to AI: those with verifiable answers and reliable data. For work outside physics, he brought in experts in the relevant fields. When the Claude model claimed to have found a way to predict earthquakes 15% better than current methods, an independent expert readily disproved the claim. Schwartz has yet to identify a profound insight in physics that has come from AI: “Is this a fundamental limit, or are the tools simply not ready yet? I’m almost certain the tools aren’t ready.”

Princeton chemistry professor Leslie Schoop, also an Anthropic consultant, describes the leap in AI capabilities over recent months as undeniable. “It can now do pretty much everything I can, except it doesn’t have a laboratory,” she jokes. Her group discovered a new structure in the laboratory through chance, intuition, and repeated attempts. None of the models she asked to predict that structure succeeded.

Neuroscientist Mark Histed of the US National Institute of Mental Health believes that the difficult questions in his field “won’t be solved by adding intelligence.” Better data are needed, and for a complex system, the amount required will eventually exceed what could be collected “before the heat death of the universe.” Garisto explains that the problem is chaos. In a system such as Earth’s weather, and possibly the brain, tiny differences in initial conditions grow over time, so calculations remain accurate only for a limited period into the future.

Two years ago, Garisto wrote that AI would remain merely a tool for scientists. Now this skeptic, who still does not use generative AI, reluctantly reports: “We are facing a fundamental shift in how information is produced, an event in human history that will probably prove as consequential as the invention of the printing press.”

Originally published on Telegram by Ukhvat NewsView on Telegram
Sources
#gpt-6-astra#theoretical-physics#automated-verification#laboratory-access#scientific-ai#chaotic-systems