PSI Reports Proof of 18-Year-Old Quantum Conjecture

By Grace Kim

An 18-year-old quantum-computing conjecture now has a reported, machine-assisted proof. Physical Superintelligence used early access to Anthropic’s Fable 5.2 model to address the Aaronson-Ambainis problem. The result appears to close one route toward quantum speedups, though it leaves faster solutions to other hard problems open.

What did the AI system prove?

The conjecture concerns whether quantum computers can solve certain problems faster than classical computers. Scott Aaronson, one of the two researchers who posed it, said he does not fully understand the proof.

Aaronson believes it is correct because it was autoformalized in Lean, a coding language for mathematics. It was also explained by a team of leading researchers whom PSI shared the solution with. That formal record and the team’s review gave him confidence.

Aaronson wrote: “A proof of full [Aaronson-Ambainis conjecture] was one of the last remaining open problems in quantum query complexity that everyone seemed to agree was centrally important.” He said the news left him satisfied, awestruck, disappointed and fatigued.

David Turturean, a recent MIT graduate and PSI technical staff member, used Anthropic’s model for the work. He deferred plans to pursue a master’s degree at MIT’s computer science and AI laboratory to join PSI. Turturean said he was accepting that the paper might be his last and questioned why anyone would publish another within months.

The proof removes one candidate path for quantum speedups on difficult problems, but it does not restrict quantum advantage across the broader class of computational tasks. Turturean called it a negative result that tells researchers they no longer need to pursue that avenue.

Aaronson had noticed the problem missing from OpenAI’s Mathocalypse dump several weeks earlier. He wondered whether OpenAI had tried it and found it harder than the problems it solved.

How does PSI turn models into revenue?

PSI’s approach combines models with specialized harnesses and scaffolding for scientific problem-solving. CEO Matt Pines said: “An effective kind of research system should have lots of things related to harnesses, lots related to tools, memory systems, simulation infrastructure.” The goal is to pursue a problem at breadth and depth while considering many potential alternatives.

At present, PSI earns revenue from using its tools to optimize data centers. The tools simulate the heat, power, water and sound around each site. Those variables determine both site performance and community impact, so their physical behavior remains central to the product. The company did not disclose a revenue amount.

The larger goal is converting theoretical physics breakthroughs into actual technologies. Six to nine months ago, Pines said he had to persuade investors that theoretical physics would soon become mostly the domain of machines.

He said the capability is not there yet, but it is on its way. The theoretical-physics business remains a roadmap; current revenue still comes from data-center tools.

Can AI validate physical discoveries?

AI has produced important solutions in mathematics and computer science because researchers can express them in validated code. Less abstracted disciplines, including life sciences and physics, have proved more difficult.

Pines said physics differs from mathematics because nature gets a vote in models of reality. PSI is focused on turning computational results into efficient experimental validation.

If a result captures a true fact of nature, researchers can exploit it in a device. AI suggests an idea, the physical world tests it, and a validated effect can support something researchers could not build before. Pines described these as the loops PSI wants to build.

Last week, startup Discovered Materials released a $1 million grand challenge for researchers using AI to study superconductivity and heat dissipation.

How was the result disclosed?

PSI held the result until Anthropic’s model was nearly ready for release. It then shared the solution with relevant researchers to improve their understanding. This gave experts time to examine the proof before wider disclosure.

Jim Halverson, a Northeastern University professor and PSI adviser, said academics wanted to do the work correctly. He sees transparency and direct contact with experts as the right path. Best practices remain unsettled as “neolabs for science” seek a role in advanced research.

Halverson will attend a summit at MIT next week. Leading physicists there will consider how the field should respond to the accelerating pace of AI-assisted science.

Turturean said this was the first AI-elicited solution he took to print. He worked on presenting it naturally for human readers and expects the academic community to ask him to explain the result. He also described the experience as a revelation of AI’s weight.

The commercial contest now shifts from showing that AI can solve research problems to proving which systems can validate physical effects, support devices and produce revenue.