A Startup Wants to Turn AI-Solved Physics into Profit

An 18-Year Puzzle Falls to AI

David Turturean, a recent MIT graduate working at Physical Superintelligence (PSI), used early access to Anthropic's Fable 5.2 model to resolve the Aaronson-Ambainis conjecture, a problem that had stood open in quantum computing theory since 2008. The conjecture concerns whether quantum computers can solve certain classes of problems faster than classical machines. Scott Aaronson, one of the researchers who originally posed the question, confirmed that the solution appears correct after it was formalised in Lean, a formal verification language for mathematical proofs, and reviewed by a group of specialists with whom PSI shared the work.

The result is a negative one: it rules out a particular pathway quantum computers might have followed to achieve speedups on certain difficult problems. That closure matters because it narrows the landscape for researchers investigating quantum advantage. Turturean described it as a way to eliminate false trails, allowing the field to redirect effort.

He also acknowledged a broader implication. "I'm totally accepting the idea that this might be the last paper I ever put out," he said, having deferred plans to pursue a master's degree at MIT's computer science and artificial intelligence laboratory. "In a few months, like, why would you even put out a paper?"

The Business Case for Theoretical Physics

PSI's revenue today comes from applying its AI tools to optimise data centres. The company simulates variables such as heat dissipation, power consumption, water use and acoustic output, treating site performance and community impact as functions of underlying physics. That work is immediate and commercial. The longer ambition, however, is to convert theoretical discoveries into technologies that can be built and sold.

Matt Pines, PSI's chief executive, said the company is constructing what he calls "research systems" - scaffolds that combine model harnesses, memory architectures, simulation infrastructure and tool libraries to pursue scientific problems across both breadth and depth. The goal is to enable machines to explore a wide range of potential solutions and validate them efficiently.

Six to nine months ago, Pines recalled, he had to persuade investors that theoretical physics would soon become machine-dominated territory. "And it's not there yet, but it's on its way," he said.

Validation Remains the Hard Part

AI has demonstrated strength in fields such as mathematics and computer science, where solutions can be expressed in code and verified algorithmically. Extending that capability to disciplines such as the life sciences or physics, where experimental validation is required, has proven more difficult. Last week, Discovered Materials announced a $1 million challenge for researchers using AI to tackle problems including superconductivity and heat dissipation, a signal that the gap between computational results and real-world confirmation remains wide.

Pines emphasised that physics differs from pure mathematics precisely because nature imposes constraints that models must satisfy. "At the end of the day, physics is different than math because it's the real world … nature gets a vote in your models of reality," he said. PSI is focused on designing loops that take computational findings, validate them through experiment, and - if they reflect genuine physical principles - turn them into devices or processes that enable something new.

Standards Still Emerging

Because PSI had to hold its quantum conjecture result until Anthropic's model was close to public release, the team used the delay to share the proof with relevant researchers and refine its presentation. Jim Halverson, a professor at Northeastern University who advises PSI, said transparency and expert consultation were chosen as guiding principles. He will attend a summit at MIT next week where leading physicists plan to discuss how the field should respond to the accelerated pace of AI-assisted discovery.

"Best practices are still emerging," Halverson said. "We figured that transparency is a good way to go, and contacting experts is a good way to go."

Turturean spent time reworking the AI-generated proof to make it more naturally readable for human researchers, anticipating that the academic community would expect him to explain the result in accessible terms. "And I think it also made me realise, oh my God, the weight of what AI can do," he said.

What Happens to Scientific Careers

Aaronson described his reaction to the proof as a mixture of satisfaction, awe, disappointment and fatigue. He noted that the Aaronson-Ambainis conjecture was widely regarded as one of the last centrally important open problems in quantum query complexity. He also observed that OpenAI's recent release of solved mathematical problems - which he referred to as the "Mathocalypse dump" - did not include a solution to this conjecture, raising the question of whether PSI's approach offered a material advantage.

The incident highlights a larger tension. If machines can now solve problems that once anchored doctoral theses and tenured careers, the structure of academic research may need to adapt. Turturean's remark about possibly never publishing another paper reflects uncertainty about what role human researchers will play once AI systems can generate, verify and even explain proofs at scale.

PSI's model assumes that the bottleneck will shift from discovery to validation and application. The company is betting that the real value lies not in producing theoretical results, but in building the infrastructure to test them against physical reality and convert the ones that survive into commercial technology. Whether that loop can be closed efficiently, and at scale, will determine whether AI-driven science becomes a viable business or remains an expensive research experiment.