Physical Superintelligence used early access to Anthropic's Fable 5.2 model to solve the Aaronson-Ambainis conjecture, an eighteen-year-old problem in quantum computing. The work was done by David Turturean, a recent MIT graduate who deferred a master's at MIT's computer science and AI lab to join the startup lab. Scott Aaronson, one of the authors of the conjecture, said he believes the proof is correct because it was autoformalized in Lean. For business, this matters as a first commercial test of AI systems built specifically for scientific discovery.
How PSI proved the quantum conjecture
At issue was the ability of quantum computers to solve certain problems faster than classical machines. The proposed solution closes one avenue by which quantum systems could achieve high speed on difficult tasks, while leaving other avenues open. Turturean described it as a negative result that tells researchers where not to look anymore. Aaronson called the conjecture one of the last centrally important open problems in quantum query complexity. He wrote that he felt satisfied, awestruck, disappointed and fatigued, and noted surprise that the problem was missing from OpenAI's Mathocalypse release.
PSI attributes the result not to a base model alone but to specialized harnesses and scaffolding around it. CEO Matt Pines said an effective research system needs harnesses, tools, memory systems and simulation infrastructure. The goal is to pursue a problem with breadth and depth while considering a wide range of alternatives. Turturean worked from an AI-elicited solution, then rewrote it for human readers before publication. Aaronson noted that PSI shared the solution with leading researchers who helped explain it, which supported confidence in the Lean formalization.
The case arrives amid rapid change in how mathematics and computer science use AI proofs. In those fields, results can be expressed in validated code, which makes machine checking practical. Physics and life sciences remain harder because models must match physical reality, where nature gets a vote. Last week Discovered Materials released a $1 million grand challenge for AI research into superconductivity and heat dissipation. Six to nine months ago Pines said he had to stretch to convince investors that theoretical physics would become mostly the domain of machines. He now says that shift is not complete but is on its way.
What AI-driven discovery means for business
For companies operating compute infrastructure, PSI already offers a concrete product based on the same physics capability. Its current revenue comes from optimizing data centers by simulating heat, power, water and sound variables. Site performance and community impact are treated as a direct function of physics, which allows trade-offs to be modeled before construction or expansion. For a small operator this can mean fewer design iterations with consultants, while for a large operator it can mean standardized simulation across a portfolio of sites. The long-term plan is to convert theoretical physics results into working devices.
The same approach carries clear limits for buyers evaluating vendors of AI for science. A Lean-checked proof does not equal an experimentally validated physical effect, and PSI says its focus is learning how to validate computational results with experiments efficiently. Buyers should ask what tools, memory and simulation layers sit around the model, how external experts review results, and what experimental step confirms a claim. The news alone does not show that the system can repeat such results across biology, materials or energy. Questions about error rates, cost per project and ownership of resulting intellectual property remain open.
The marker to watch is the summit at MIT next week, where leading physicists will discuss standards for AI-driven science. Northeastern professor and PSI adviser Jim Halverson said best practices are still emerging and pointed to transparency and contact with experts. If that meeting produces shared rules on disclosure, review and experimental confirmation, corporate R and D groups will gain a clearer basis for procurement. If not, each vendor will continue to define proof in its own terms.
