Startup Physical Superintelligence (PSI) is trying to build a business around scientific discoveries made with artificial intelligence. One of its employees used early access to Anthropic's Fable 5.2 model to solve a quantum computing problem that had remained open for 18 years. For the company, this is not just a demonstration of AI's capabilities, but an example of how specialized systems can accelerate fundamental science and eventually lead to new technologies, TechCrunch reports.

AI solved a problem scientists had been working on for nearly two decades

David Turturan, a recent MIT graduate, joined PSI instead of pursuing a master's degree. While working at the startup, he gained early access to Anthropic's Fable 5.2 model and used it to find a solution to the Aaronson–Ambainis conjecture, one of the major open problems in quantum computing theory.

The conjecture concerns how much faster quantum computers can solve certain problems compared with classical machines. The solution appears to close off one avenue through which quantum computers might gain a significant advantage in solving certain difficult problems. However, it does not mean that the capabilities of quantum computing as a whole have been limited or that all promising approaches have now been ruled out.

The result is primarily a negative one: it allows researchers to stop looking for an advantage in one particular direction and focus on other possibilities.

Scott Aaronson, one of the scientists who formulated the conjecture, acknowledged that he does not yet fully understand the proof. Nevertheless, he believes it is correct: the solution was formalized using Lean, a language and system for verifying mathematical proofs. PSI also shared the result with a group of leading researchers, who prepared a detailed explanation.

Aaronson described the development as both impressive and exhausting. He was also surprised that the problem had not appeared in a recent roundup of mathematical problems solved using OpenAI models: in his view, it was considered one of the most significant unsolved problems in quantum computational complexity.

For Turturan, the experience prompted reflection on the future of scientific publishing. He suggested that this might turn out to be his last paper in the traditional academic format: if AI can solve such problems independently within a few months, the very need to publish papers could change.

PSI is betting on an entire research system, not a single model

This achievement is also important for PSI because the company is trying to develop its own approach to using AI in science. It does not expect to simply ask a general-purpose model questions and receive ready-made discoveries. Instead, the startup is developing specialized systems that combine models, tools, memory, simulations, and additional software mechanisms known as harnesses.

PSI CEO Matt Pines believes an effective research system must be able to explore a problem both broadly and deeply: considering many possible solutions, testing alternatives, and using appropriate computational tools.

In other words, what matters is not just a model's ability to reason, but also the infrastructure that enables it to carry out sustained research. The company envisions this combination turning AI from a scientist's assistant into an independent tool for scientific discovery.

For now, PSI makes money from a more applied task: optimizing data centers. Its systems model the physical parameters of facilities, including heat, energy consumption, water, and noise. These characteristics affect both a data center's operational efficiency and its impact on the surrounding community.

The company's long-term goal, however, is far more ambitious: to use theoretical discoveries in physics to create real devices and technologies. Pines acknowledges that, until recently, he had to convince investors that fundamental physics could soon become a field dominated by machine-led research. He says that point is still a long way off, but the shift in that direction has already begun.

Why mathematical proofs are only the beginning

AI's successes in mathematics and computer science do not yet mean that machines will be able to make discoveries just as effectively in every scientific field.

In mathematics, a result can be expressed in a formal system and verified using software. A proof written in Lean, for example, can undergo rigorous automated verification. In physics and biology, the situation is more complicated: a computational model must correspond to the real world, and its predictions must be confirmed through experiments.

“Ultimately, physics differs from mathematics in that there is a real world. Nature also gets a vote on which models of reality are correct,” Pines explains.

PSI is therefore trying to determine how best to turn computational results into experimental tests. AI may propose a new idea or identify a pattern, but researchers must then establish whether it truly describes a natural phenomenon. If confirmed, the discovery could become the basis for a technology or device that was previously impossible to create.

Pines describes this process as a series of cycles: a model proposes a hypothesis, computations help test it, experiments compare the result with reality, and the resulting data is used to guide further exploration.

The challenges of this approach are already drawing attention from other companies. For example, startup Discovered Materials announced a competition with a $1 million prize pool for researchers using AI to tackle problems such as superconductivity and heat dissipation. This shows that interest in automated scientific discovery extends beyond mathematical problems.

Science is still developing rules for working with AI

Rapid progress is raising not only technical but also organizational questions for the scientific community. How should discoveries made by models be verified? Who should be responsible for the correctness of a proof? How thoroughly must a researcher understand a solution obtained using AI? And how should such results be published so that other scientists can assess them independently?

PSI tried to take these questions into account while working on the proof of the Aaronson–Ambainis conjecture. Because the company had to wait until Anthropic's model was close to public release, the team had time to show the result to specialists in the field and obtain their assessment.

Jim Halverson, a professor at Northeastern University and a scientific adviser to PSI, believes the academic community is only beginning to develop best practices for this new era of research. He says openness and early engagement with experts are important conditions for the responsible use of AI.

Turturan himself also faced the need not just to get an answer from the model, but to explain it to people. He said he first read the AI-generated solution carefully only after it had been prepared for publication, and tried to express it in language that felt more natural to humans. This helped him grasp the scale of the new technology's capabilities.

PSI's story shows that AI-assisted scientific work is gradually moving beyond automating individual calculations. Models are already capable of helping solve problems that long remained open for specialists. But turning such results into reliable knowledge, reproducible experiments, and commercial technologies will require not only more powerful models, but also new verification methods. It is precisely this transition from a machine-generated hypothesis to a confirmed discovery that PSI is trying to build its business around.