Physical Superintelligence, a startup lab known as PSI, is seeking to build a business around AI-assisted scientific discovery after one of its technical staff used early access to Anthropic’s Fable 5.2 model to solve a long-standing problem in quantum computing. The result concerns the Aaronson-Ambainis conjecture, an 18-year-old question about the ability of quantum computers to solve certain problems faster than classical computers. The work has attracted attention because it suggests that advanced AI systems may be able to contribute to difficult mathematical research, not merely summarise papers or write routine code. The finding is significant, but it should be understood as a specific result under expert review rather than proof that AI has automated science.
A formal proof changes the research conversation
David Turturean, a recent MIT graduate who joined PSI instead of immediately pursuing a master’s degree, used the model to work on the conjecture. The problem was posed by researchers Scott Aaronson and Andris Ambainis and concerns the limits of quantum advantage for particular computational tasks. According to the reporting, the result is a negative one: it closes off one possible route by which quantum computers might achieve certain speedups. Negative results are important in theoretical research because they can prevent other teams from spending years exploring an avenue that cannot provide the expected outcome. They do not eliminate all possible quantum advantages or invalidate the broader field of quantum computing.
A key part of the verification process was formalisation in Lean, a proof assistant used to express mathematical arguments in a form that software can check. Formal verification does not automatically guarantee that every assumption or interpretation is correct, but it can reduce the risk of unnoticed gaps in a long proof. Aaronson said he did not fully understand the proof at first but believed it was likely correct because it had been autoformalised in Lean and explained by researchers with whom PSI shared the solution. The combination of machine-generated reasoning, formal checking and expert review provides a useful model for evaluating AI contributions to mathematics.
The result also shows a possible change in the economics of research. Mathematical work often requires specialists to spend long periods exploring conjectures, checking edge cases and translating informal intuition into rigorous proof. AI systems that can assist with these tasks could reduce the time needed to test ideas or make formal verification more accessible. But the workflow still depends on humans to define the problem, evaluate the assumptions and determine what a result means for the field. A model can produce an answer that looks convincing while hiding a flaw, and a formal proof can only verify the statement encoded in the system. Scientific reliability So requires both technical validation and expert interpretation.
Verification remains essential when AI contributes to mathematics
PSI’s commercial ambition is broader than publishing a single mathematical result. The company is positioning itself around AI systems that can contribute to research in physics, mathematics and other technical disciplines. That raises questions about what customers would pay for: access to a model, a research service, validated discoveries or a combination of software and expert support. Different scientific fields have different standards of evidence, data requirements and routes to commercial use. A discovery in theoretical mathematics may be valuable because it changes what researchers pursue, while an engineering or life-sciences result may need experimental confirmation before it can influence a product or clinical decision.
The most important next step is independent scrutiny of the proof and a clearer account of the research process. The conjecture’s co-author has expressed confidence while acknowledging that the argument required explanation from experts. That is a meaningful but qualified endorsement. Researchers will need to examine the proof, verify its assumptions and understand how the AI system contributed to the final reasoning. If the result survives that process, it could become an example of AI accelerating a specialised research task. It would not mean that every mathematical or scientific problem can be solved by a model, but it would strengthen the case for carefully designed AI research systems that combine reasoning, formal tools and human review.