MIT graduate solves 18-year-old quantum computing problem with the help of an AI model

Solving an 18-year-old problem

David Turturyan, a recent MIT graduate and employee of the startup lab Physical Superintelligence (PSI), received early access to the Anthropic Fable 5.2 model and solved a problem that had remained open for 18 years — the Aaronson–Ambainis conjecture about when quantum computers can solve problems faster than classical ones. Turturyan postponed his plans to enroll in MIT's master's program at the Computer Science and Artificial Intelligence Laboratory in order to join PSI. “I fully accept that this may be my last paper”, he said, adding that in a few months publishing papers may lose its meaning.

Aaronson's reaction and the meaning of the result

Scott Aaronson, one of the two scientists who formulated the problem, wrote that he does not fully understand the proof but believes it is correct: the solution was autoformalized in Lean — a programming language for mathematics — and explained by a team of leading researchers to whom PSI passed the result. According to him, a full proof of the conjecture was “one of the last remaining open problems in quantum query complexity that everyone considered centrally important”; he described his feelings as satisfaction, excitement, disappointment and fatigue, and also expressed surprise that the problem was not on OpenAI's recent “Mathocalypse” list of solved problems.

What the result means for quantum computing

The solution closes only one of the possible paths by which quantum computers could solve certain complex problems at super speed, but not all of them. “This is a negative result that lets people understand: okay, we no longer look there,” Turturyan explained. The conjecture concerned the question of quantum machines' advantage over classical ones, and its resolution narrows the search space for quantum speedup.

PSI's business model and the limits of AI science

PSI is betting on solving scientific problems with specialized “harnesses” and scaffolding for models. “An effective research system must have a lot of things connected with harnesses, tools, memory systems, simulation infrastructure,” said the company's CEO Matt Pines. For now, the company's revenue comes from data center optimization — modeling heat, energy, water and noise, where a site's performance directly depends on physics. The long-term goal is to turn breakthroughs in theoretical physics into real technologies. Pines admitted that six months ago investors had to “stretch” to justify this thesis: “theoretical physics will soon become mostly a domain of machines” — “that's not yet the case, but it's on the way.” At the same time, he noted that physics differs from mathematics in that “nature votes in your models of reality”: in disciplines less abstract than mathematics and computer science, AI solutions are harder to verify. Last week, the startup Discovered Materials announced a $1 million grant for using AI on problems such as superconductivity and heat dissipation.

New standards for AI science

Because PSI had to hold back the result until Anthropic's model release was nearly ready, the company managed to share the solution with relevant researchers. “As academics, we wanted to play our part in doing things right,” said Jim Halverson, a professor at Northeastern University who advises PSI. Next week he will take part in a summit at MIT, where leading physicists will try to work out a response to the new pace of science: “Best practices are still forming… We decided that transparency is a good path, and contact with experts is a good path.” Turturyan himself said that for the first time he reworked an AI-derived solution for a “more natural” human presentation, since the scientific community will expect him to explain the result — and that this made him realize “the weight of what AI can do.”