PSI Leverages AI to Crack Long-Standing Quantum Conjecture

Physical Superintelligence (PSI), a startup founded by MIT graduates, has demonstrated that specialized AI systems can resolve decades-old problems in theoretical computer science. Using early access to Anthropic's Fable 5.2 model, researcher David Turturean produced a proof for the Aaronson–Ambainis conjecture, an 18-year-old question in quantum query complexity. The proof was autoformalized in the Lean theorem prover and vetted by leading experts, including Scott Aaronson himself, who confirmed its correctness despite initial surprise.

Why It Matters for AI Infrastructure

PSI's approach goes beyond generic large language models. The company builds dedicated research harnesses that combine tool use, memory, simulation environments, and iterative reasoning scaffolds. This architecture enables the system to explore a problem at breadth and depth, a capability that generic chat interfaces lack. For GPU cloud providers like AiGpu, the implication is clear: scientific AI workloads demand not just raw compute but also tailored software stacks that manage long-horizon reasoning, formal verification, and domain-specific simulation.

Currently, PSI monetizes its platform by optimizing data-center physics—modeling heat, power, water, and acoustic constraints to improve site performance and community impact. The longer-term vision is to translate automated theoretical breakthroughs into tangible hardware and materials innovations. However, CEO Matt Pines acknowledges that physics presents a higher bar than mathematics: nature ultimately validates or rejects a model, making experimental feedback loops essential. Recent initiatives such as Discovered Materials' $1 million challenge for AI-driven superconductivity research underscore the growing investment in this direction.

As AI moves from pattern matching to autonomous discovery, the infrastructure layer must evolve to support continuous, tool-augmented research pipelines. Organizations that provide flexible GPU clusters, low-latency storage, and integrated MLOps tooling will be best positioned to capture this emerging workload class.