AI Now Solves 35% of New Theoretical Science Problems, MIT Study Finds
Models jumped from solving 2% to 35% of theoretical science problems in two years. Scientists are about to become AI fact-checkers.
Research · Source: AI Daily Post
What happened
MIT researcher Neil Thompson just dropped a study that should wake up the academic world. Between mid-2024 and mid-2026, frontier AI models went from solving roughly two percent of theoretical science problems to about 35 percent. The study looked at six specific areas. These include probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics. The rapid jump in capability has triggered calls for limits on the reach of AI into scientific fields.
This was not a cheap memorization trick. Thompson ran the study with PhD student Gabriel Manso and research scientist Claudio Mayrink Verdun. They pulled problems straight from papers posted online between December 2025 and May 2026. They tested these problems against a run of OpenAI models, ranging from o1 through GPT-5.5. The design of the study matters heavily. The solutions appeared in papers published after the training cutoff for the newest models tested. There was no way for the models to simply recall an answer from the internet. Whatever these systems produced, they had to work out entirely on their own.
The testing happened offline using proprietary models. Now, human experts are stuck checking if the AI answers are actually right. The incentives here are completely warped. Recently, OpenAI dumped over 700 manuscripts online from a secret internal model. They kept the model secret and only published the results they liked. A Princeton advisory group recently called out this exact behavior. They criticized AI companies for relying on secret internal models to mass-produce science. OpenAI essentially said they do not have to listen.
Key facts
- 35% — The portion of theoretical science problems top AI models can now solve.
- 700 — Manuscripts OpenAI recently dumped online from a secret internal model.
- December 2025 to May 2026 — The publication window of the papers used to test the models, ensuring answers were not in the training data.
Why it matters
The value of academic work is getting redefined faster than legacy institutions can adjust. Journals, tenure committees, and funding bodies still reward final proofs and clean results. They are not built to credit the reasoning process that AI tends to skip over. Mathematicians are already arguing that the human reasoning process was the entire point of the work. Theoretical science is becoming incredibly cheap to automate because you do not need a physical lab. You just need massive servers. Money and prestige will naturally flow to whoever owns the biggest models.
A massive power imbalance is forming in the scientific community. AI labs get the flashy headlines for mass-producing science. Meanwhile, academic referees bear the heavy cost. They are forced to do the unpaid labor of checking the complex math generated by these models. Worse, they cannot even test the proprietary models that generated the answers. Universities without massive compute budgets will soon lose their status as centers of discovery. They will just become dependent clients of big tech companies. The role of the scientist is shifting from discoverer to mere fact-checker.
For builders
Build tools for transparent reasoning
Working scientists need systems that preserve and expose intermediate reasoning, not just final answers. The demand for transparent processes will soon outpace the demand for raw output. Founders who build tools showing exactly how AI reached a complex conclusion will win over those just shipping another benchmark-topping solver. Journals and universities will pay for this transparency.
Automate the verification process
Academics are drowning in unpaid labor checking AI-generated math and science. There is a massive, untapped market for tools that help referees verify AI outputs quickly and accurately. If you can build a verification layer for theoretical science, academic institutions and publishers will become your biggest customers.
My take
Compute is power, and the labs with the biggest servers are turning brilliant scientists into unpaid editors. If we do not build tools that expose the reasoning behind these AI breakthroughs, science just becomes a black box owned by OpenAI. The future of discovery belongs to whoever controls the compute, but the immediate business opportunity is in building the verification layer.
Original reporting: AI Daily Post. This is my rewrite and opinion.