MIT study: AI solves far more theoretical science problems; physics seen as next

Left · 1Wired reported that MIT researcher Neil Thompson, who studies progress in computing and AI, and colleagues found that the best AI models went from correctly solving about 2 percent of theoretical problems in several fields in mid-2024 to about 35 percent in mid-2026. 1 Left · 1The fields named as possibly next to be affected are probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics. Thompson said, "Within just a few years, we expect a lot of theory questions to be totally transformed in terms of AI's ability to do them." 1 Left · 1The study, by Thompson, PhD student Gabriel Manso and research scientist Claudio Mayrink Verdun, used problems from research papers posted between December 2025 and May 2026 and tested OpenAI models from o1 to GPT-5.5. Solutions appeared in papers published after the models' training cutoff, and internet access was blocked, so the models had to work out answers themselves. 1

Left · 1The study was limited to fields where models could reason without running simulations or experiments. Wired says it suggests the upheaval in mathematics may spread to other theory-heavy areas. 1 Left · 1The mathematics context, according to Wired, includes OpenAI's claim last month that an unreleased model solved the Navier-Stokes challenge, which was disputed by an NYU mathematician who accused OpenAI of gazumping them. Some mathematicians have since called for limits on AI use. 1 Left · 1Manso said people will still be needed to generate ideas and verify results, since a model can wrongly claim a right answer; Wired notes OpenAI withdrew several solutions from its latest dataset after publication. 1

Centre · 1New Scientist reported that Yale's John Sous believes AI is underrated at physics because of flawed tests: models scored as low as 30 per cent on six physics benchmarks, but after errors and ambiguous definitions were corrected, scores reached as high as 94 per cent. 2 Centre · 1Sous said AI may be very good at anything well defined, but it is unclear whether it can do the full job of a theoretical physicist, which he called a creative process, in the fuzzy, cutting-edge parts of physics. New Scientist adds that mathematics has advanced faster partly because formalisation lets AI check its answers, which is far less developed in physics. 2 Centre · 1Yale's Logan Wright said AI is already used in experiments where inputs and outputs can be computer controlled. His lab uses Anthropic's Claude to control laser pulses on a rubidium atom, with experiments running overnight that match about a week of human work. He expects this to work only in constrained domains and noted a debate over whether results nobody understands count as science. 2