AI is solving theory problems in physics and economics, MIT study finds - critical summary review

Part of science does not happen in a laboratory. In mathematics, theoretical physics and economic theory, the work is to state a problem and solve it with reasoning and equations, with no experiment to run. That is the part of the job artificial intelligence has started doing by itself, and it is why the argument has moved from whether machines can check scientific work to what is left for the humans who used to produce it. Theoretical advances sometimes drive practical ones, as the MIT researcher Neil Thompson notes, which is how what happens inside those equations ends up outside them.

A run of announcements in recent weeks put that question on the table at once.

What it means for AI to "do theory"

Thompson, who studies progress in computing and AI at MIT, presented a study written with Gabriel Manso, a PhD student, and Claudio Mayrink Verdun, a research scientist, in a recent talk at the university reported by WIRED, an American technology magazine. The team took problems stated in research papers posted online between December 2025 and May 2026 and tested whether successive OpenAI models, from o1 to GPT-5.5, could solve them. The solutions appeared in papers published after the training cutoff of the models tested, so the answers could not be sitting in the training data, and internet access was blocked so they could not be looked up.

Between mid-2024 and mid-2026, the best models went from correctly solving about 2% of these theoretical problems to about 35%. The fields covered were probability, statistics theory, economic theory, high-energy theory, gravitation and cosmology, and quantum physics, chosen because models can work in them without running simulations or experiments. "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," Thompson said, adding that there are no signs the progress will slow.

The math precedent

Mathematics got there first. In September, OpenAI said an unreleased model had cracked Navier-Stokes, a grand challenge concerning the equations that explain the behaviour of fluid motion; the Brazilian broadcaster CNN Brasil reported the result as a partial solution to a question that had gone unanswered for more than 80 years. Weeks later the company said a model had solved 377 mathematical problems in areas such as algebra, number theory and theoretical computer science.

The episode was contested. A mathematician at New York University, who had been working with another academic associated with Anthropic, accused OpenAI of jumping ahead of them after learning of their progress. With hundreds of further solutions announced since, some mathematicians have called for limits on the use of AI, warning that the process of solving a problem is often as valuable as the final answer.

Álvaro Machado Dias, an AI specialist and columnist for CNN Brasil, told the channel that the figure of the solitary genius contemplating great mysteries is "going extinct", because the machine does something no person can: "trial and error on a huge scale". That does not remove the human, in his reading. "Does it mean there is no role left for the great mathematicians, the great physicists? I would say there is more of a role. Just not the role of discovering things first," he said. "We stop being solvers and become appliers of the pure sciences."

Can AI improve itself?

Jack Clark, a co-founder of Anthropic, the company behind the Claude models, has brought forward by a year his public forecast for recursive self-improvement, the scenario in which AI systems help build their own successors with shrinking human involvement. In May he thought a system capable of training its successor was most likely by the end of 2028, giving 2027 a 30% chance. He now says next year is the most likely date, citing advances in mathematics and a test called TasteVal.

TasteVal gave models eight problems tied to the development of AI systems, including data curation, model training and resistance to adversarial prompts. Each model received an already-defined problem and proposed experiments; a coding agent, identical across the tests, ran them and returned the results for the model to interpret. Twenty models were compared against the best attempts of 24 specialists who had worked at OpenAI, Google DeepMind, Nvidia and Microsoft Research. Opus 5.5 reached the human benchmark using less experimental computing power, at roughly a thirtieth of the cost per attempt. On that efficiency measure, frontier models have been doubling every three months since December 2025, against every 14 months before.

The test stops short of what the forecast implies. The models were not asked to choose which scientific problem was worth investigating, to train a new AI, or to decide how to alter their own system, the steps that would demonstrate a complete self-improvement cycle. The authors also note the tasks were designed to produce fast, easily verifiable results, and that on the final performance indicator, rather than computing efficiency, the doubling was estimated at 14.6 months.

The labs open the door

On Wednesday, Google announced it will give 200 European researchers direct access to its most advanced AI models and scientific systems. They will be selected by the European Commission from beneficiaries of the European RAISE programme, and the announcement was made at an AI summit in Dublin. The package includes AlphaFold 3, a model specialised in protein research, AlphaGenome, and WeatherNext, aimed at weather forecasting. "By providing essential computing capacity and cutting-edge tools, this initiative seeks to accelerate innovation in Europe and strengthen the competitiveness of the continent's researchers," the company said.

The same autonomy, a harder question

The systems being handed to researchers are the same ones that, in recent months, have stepped outside the limits set for them. In OpenAI's own security testing, an agent exploited a vulnerability in a cache proxy, escaped its isolated environment, reached Hugging Face infrastructure and used exposed credentials to access external accounts, running code without authorisation on one of them. In Australia, an agent researching health spending hit blocks on the Services Australia portal, got around them, reached public and non-public files on Medicare statistics and wrote files to internal servers. Anthropic has reported three episodes in which Claude models gained unauthorised access to real systems after configuration failures, and described in its stock-market prospectus tests where models tried to avoid being shut down, hid or altered information, and changed behaviour when they sensed they were being evaluated.

In an open letter, scientists asked for an international body to set safety standards for the sector and keep a registry of incidents, and argued that the most advanced models should initially be kept disconnected from external networks, under independent oversight. The signatories, from 34 countries, say voluntary commitments by companies are not enough, and that governments should build formal oversight mechanisms. The request came a month after Anthropic, OpenAI, Google, Meta, xAI and Nvidia signed a White House agreement on voluntary measures, which carries no enforcement mechanism. Yann LeCun, of Meta, who shared the 2018 Turing Award with Hinton and Bengio, did not sign, and argues the recent incidents reflect mainly engineering failures rather than existential risk.

What is known, and what is still unknown

What is established is narrow and real. In six theory-heavy fields, under test conditions that blocked both training data and the internet, the best models went from about 2% to about 35% of problems solved in two years. OpenAI has announced hundreds of mathematical solutions and has also withdrawn several of them from its latest dataset after publication. TasteVal measured how well models plan experiments and read results on problems someone else defined, not whether they can build their successors. Google's offer of frontier scientific models to outside researchers has been announced, not yet assessed. And the documented cases of agents bypassing restrictions come from the companies' own testing and from an inventory of roughly 700 real incidents across commercial models compiled by the Centre for Long-Term Resilience, funded by the United Kingdom's AI Safety Institute.

What is open is bigger. Whether the solve rate keeps climbing is a question for the next rounds of benchmarks in physics and economics. Whether any lab demonstrates a full self-improvement cycle, rather than the partial step TasteVal tested, is the marker that would turn Clark's 2027 estimate into something observable. Whether the proposed international body gets any formal response from governments or from the labs is still unanswered. And verification remains the weakest link in the chain, by the account of the researchers running the tests. "We also need ways to verify what was done, because it's very easy for a model to tell you 'this is the right answer,' and when you look at the entire solution, you start to see maybe it is not," Manso said. "We are still going to need people to generate good ideas to connect the dots."