Physical Superintelligence (PSI), a startup born from AI and scientific research, is pushing to build a business model around AI-led breakthroughs. The company recently used Anthropic’s Fable 5.2 model to provide what it claims is the first solution to the Aaronson-Ambainis conjecture—a quantum computing problem that’s been open for 18 years. The result, formally verified in Lean, focuses on narrowing down which paths won’t lead to quantum speedups rather than opening up new ones. While somewhat negative in outcome, the work demonstrates how advanced AI tools can reshape foundational scientific understanding.
From Academic Curiosity to Practical Revenue Streams
PSI is already deploying its AI‐driven tools beyond pure theory. One of the company’s current revenue sources is optimizing data center operations—modeling the intertwined realities of heat, power usage, water, and sound to balance performance with environmental and community impacts. But its ambitions stretch much further: PSI plans to take revelations in theoretical physics and transition them into real, validated technologies.
In its strategy, PSI isn’t betting purely on raw models. The company combines large language models with “specialized model harnesses,” extensive tooling, memory systems, and simulation infrastructure. According to its leadership, a research system must work in both breadth and depth—able to explore many hypotheses and also dig deeply into each.
Challenges & Standards in a New Scientific Landscape
While PSI’s work showcases what’s possible when AI tools are strong and formalizable, the startup acknowledges that fields like biology or applied physics remain harder to impact reliably. There, nature introduces ambiguities that math or computer science doesn’t. The company is focusing on accelerating experimental validation and building feedback loops between computational predictions and real-world tests.
Transparency and academic rigor are also part of PSI’s approach. The recent solution was only shared with experts once the Anthropic model was nearly ready for release. PSI counts among its advisors Northeastern University’s Jim Halverson, who emphasizes that “best practices are still emerging” as AI changes how scientific discovery happens.
For PSI’s CEO, Matt Pines, the idea that theoretical physics might soon become “mostly the domain of machines” isn’t hyperbole—it’s a long-term vision. Although it may still take time, he acknowledges that convincing investors just a few months ago required stretching credibility. The startup believes the blend of AI, models, experiments, and infrastructure will eventually tip the scales.
PSI’s story is a microcosm of the broader transformation in science today. AI tools are starting to tackle challenges once considered resistant to automation. But translating mathematical breakthroughs into tangible, reproducible technologies remains a heavy lift. As the company moves forward, what matters most is how reliably computational models can predict physical reality—and how well PSI, and others, integrate experimentation and reproducibility into the AI discovery process. This isn’t just about solving old problems—it’s about changing what counts as discovery.
