PSI is trying to make a business out of AI-driven scientific discovery

Startup lab Physical Superintelligence used early access to Anthropic's unreleased Fable 5.2 model to solve the 18-year-old Aaronson-Ambainis conjecture in quantum computing, as detailed in TechCrunch's report. David Turturean, a recent MIT graduate who deferred his master's studies at MIT's computer science and AI lab to join the company's technical staff, developed the solution using specialized model harnesses and scaffolding. The mathematical proof was autoformalized in Lean, settling a core theoretical question regarding how quantum algorithms compare against classical counterparts.

This report examines how Physical Superintelligence engineered its research system to resolve complex mathematical conjectures, how the commercial laboratory funds its ongoing work through physical data center simulations, and why translating code-verified logic into physical hardware presents a distinct operational hurdle. It is aimed at research engineers, quantum computing scientists, infrastructure architects, and machine learning practitioners tracking automated theorem proving and computational physics.

01Quantum query complexity and autoformalized proofs

The Aaronson-Ambainis conjecture directly addressed the bounds of quantum query complexity, focusing on whether quantum machines could achieve speedups across specific problem types. Turturean's solution established a negative result, shutting down a theoretical path researchers had explored for nearly two decades. Scott Aaronson noted that while he did not fully understand every step of the derivation immediately, he accepted its validity because the complete proof was autoformalized in the Lean verification language.

Aaronson observed that the conjecture represented one of the final unresolved foundational questions in quantum query complexity. The result surprised theoretical computer scientists, particularly after it failed to appear in OpenAI's earlier Mathocalypse release. The successful formalization confirmed that targeted model scaffolding could resolve highly complex open conjectures that had resisted prior algorithmic analysis.

02Specialized harnesses and simulation infrastructure

Physical Superintelligence structured its approach around bespoke model harnesses, external tool integrations, persistent memory systems, and simulation infrastructure rather than raw model prompting. Chief executive officer Matt Pines explained that tackling scientific questions requires evaluating alternatives across broad and deep solution trees. The architecture combines large language models with rigorous algorithmic scaffolding that can execute and inspect symbolic reasoning steps.

Because the team had to wait for Anthropic's model release schedule, researchers had time to review the generated output and reformat the steps for peer review. Turturean translated the machine-elicited proof from its autoformalized format into structured mathematical prose suitable for human academic review. The methodology demonstrated how autonomous discovery engines can work alongside human verification to clarify technical proofs.

03Commercial data center simulation revenue

To finance its long-term theoretical research, Physical Superintelligence generates near-term commercial revenue by modeling operational variables for mission-critical data center sites. The company deploys physical simulation tools to optimize real-world facility dynamics, including heat dispersion, electrical power consumption, water cooling demands, and acoustic signatures. These physical factors directly govern the operating performance and local community impact of high-density compute facilities.

Pines noted that while convincing investors six to nine months ago required defending the premise that machines would soon drive theoretical physics, automated tooling is steadily closing that gap. The firm treats industrial facility optimization as a live testbed for complex physical modeling while its core team develops broader tools for theoretical science.

04Physical experiments and natural constraints

Unlike pure mathematics and computer science where proofs can be completely validated within code environments, physical discoveries face real-world experimental constraints. Pines emphasized that nature retains a vote in physical modeling, meaning computational predictions must be validated by laboratory hardware. The company is designing structured feedback loops that take algorithmic hypotheses and verify them using physical experiments.

The broader market is pursuing similar hybrid validation workflows across physical sciences. Last week, startup Discovered Materials introduced a $1 million grand challenge targeting AI researchers working on heat dissipation and superconductivity. Physical Superintelligence plans to harness confirmed physical mechanisms directly into hardware devices, turning validated scientific rules into functioning commercial systems.

05Academic peer review and emerging best practices

The rapid arrival of automated scientific proofs has prompted researchers to establish clear disclosure and evaluation standards across the scientific community. Northeastern University professor and PSI advisor Jim Halverson explained that the startup adopted an explicit policy of academic transparency, sharing the solution with domain experts before public release. Halverson is attending a dedicated summit at MIT next week where senior physicists plan to formulate working standards for AI-assisted discovery.

Academic researchers emphasize that as automated laboratories generate proofs at accelerating rates, human researchers must retain the ability to understand and audit the underlying principles. Turturean pointed out that the scientific community still expects authors to interpret and defend their results. Developing rigorous verification norms will determine how traditional scientific institutions incorporate machine-generated findings into formal academic literature.