AI Came for Math. Now It’s Chasing the Rest of Science

Last week, I attended a speech astatine MIT by Neil Thompson, who studies advancement successful computing and artificial intelligence. Over the people of conscionable fractional an hour, Thompson made a convincing lawsuit that respective scientific fields are astir to beryllium upended by AI.

As mathematicians reel from the mode AI is automating their field, Thompson’s latest survey suggests that probability, statistic theory, economical theory, high-energy theory, gravitation and cosmology, and quantum physics whitethorn beryllium next, acknowledgment to the rapidly evolving capabilities of frontier models. Between mid-2024 and mid-2026, the survey shows, the champion AI models went from being capable to correctly lick astir 2 percent of theoretical problems successful these fields to astir 35 percent of specified problems. And determination are nary signs that this advancement volition dilatory down.

“Within conscionable a fewer years, we expect a batch of mentation questions to beryllium wholly transformed successful presumption of AI's quality to bash them,” says Thompson.

Broadly, Thompson’s laboratory digs into however advances successful machine subject interaction nine and impact innovation. This study—which Thompson conducted alongside Gabriel Manso, a PhD student, and Claudio Mayrink Verdun, a probe scientist—took problems stated successful probe papers posted online betwixt December 2025 and May 2026, and past tested whether antithetic OpenAI models—o1, o3, GPT-5, GPT-5.1, GPT-5.4, and GPT-5.5—could lick them.

Crucially, solutions to these problems were recovered successful papers published aft the grooming cutoff for GPT-5.5 and preceding models. This means that the models they tested had to enactment retired solutions for themselves, alternatively than relying connected answers fed successful arsenic grooming data. (Such papers served arsenic crushed information for investigating the models’ solving skills.) The researchers besides blocked net access, which prevented the models they tested from looking up solutions online.

The areas covered by the survey were limited—the trio focused connected fields successful which AI models could fig things retired without needing to tally simulations oregon experiments. But the enactment suggests that the kinds of daze waves presently being felt successful mathematics are apt to ripple retired to different math- and theory-heavy areas of technological inquiry.

Last month, OpenAI rocked the mathematics world by announcing that an unreleased exemplary had solved a long-standing, expansive situation successful mathematics known arsenic Navier-Stokes, which concerns equations that explicate the behaviour of fluid motion. The occurrence proved other arguable aft a mathematician astatine NYU— who had been moving with different world associated with Anthropic—accused OpenAI of gazumping them aft learning of their progress.

OpenAI and others person since announced hundreds much solutions to longstanding mathematics problems, triggering thing of a situation successful the field. Some mathematicians person called for limits connected the usage of AI, informing that the process of solving problems is often arsenic utile arsenic having the last answer.