What did PSI achieve?

Physical Superintelligence, a startup lab founded by recent MIT graduates, used early access to Anthropic's unreleased Fable 5.2 model to solve the Aaronson-Ambainis conjecture, a quantum computing problem that had remained open since 2008.

The conjecture concerns the conditions under which quantum computers can solve problems faster than classical computers, a foundational question in quantum complexity theory.

The solution was autoformalized in Lean, a proof assistant language, ensuring mathematical rigor.

Scott Aaronson, one of the two researchers who originally posed the problem, said he does not fully understand the proof but believes it is correct because of the Lean verification and because a team of top researchers with whom PSI shared the solution confirmed it, as TechCrunch reported.

Why does this matter?

The result closes the door on one theoretical avenue for quantum speedup, but not all of them.

As David Turturean, the MIT graduate who led the effort at PSI, told TechCrunch: It is a negative result that allows people to know they are no longer looking there.

The finding narrows where researchers should search for quantum advantages, potentially saving years of dead-end exploration in quantum algorithm design.

More broadly, it demonstrates that AI systems can now tackle open problems in theoretical computer science that have resisted human effort for nearly two decades. The proof was not a simple pattern match or brute-force search.

The model had to reason about mathematical structures, construct a rigorous argument, and express it in a form that Lean could check mechanically.

What did Scott Aaronson say?

Aaronson, a prominent theoretical computer scientist at UT Austin, described himself as feeling satisfied, awestruck, disappointed, and fatigued by the news.

He expressed surprise that OpenAI Mathocalypse effort, which recently solved a range of hard mathematical problems, had apparently missed this one.

Surely they tried it, so was it really so much harder than all the problems that were solved, he wrote.

That question highlights what makes PSI achievement notable. The company approach combines model access with specialized scaffolding: harnesses, tool-use frameworks, memory systems, and simulation infrastructure, as PSI described.

It is not just the model itself, but the system built around it that enables these results.

What is PSI business model?

PSI is trying to build a business around AI-driven scientific discovery. The company gives researchers early access to frontier AI models and builds custom scaffolding to tackle hard problems in physics, mathematics, and computer science.

The Aaronson-Ambainis proof is its most visible result so far, intended as a demonstration of what the approach can achieve.

Turturean, who deferred plans to pursue a master at MIT prestigious CSAIL lab to join PSI, told TechCrunch he is totally accepting the idea that this might be the last paper he ever puts out because in a few months, why would anyone even put out a paper.

That captures the existential question PSI is betting on: that AI will soon surpass humans as the primary driver of fundamental scientific discovery.

Is this different from OpenAI Mathocalypse?

OpenAI Mathocalypse project, unveiled several weeks ago, solved a wide range of benchmark mathematical problems using a specialized reasoning model. But PSI result differs in two ways.

First, the Aaronson-Ambainis conjecture was a genuinely open research problem, not a benchmark. Second, PSI used Anthropic model, not OpenAI own, and built its own scaffolding rather than relying on the model built-in reasoning capabilities.

The result positions Anthropic Fable 5.2, still unreleased, as a serious contender for frontier mathematical reasoning. It also suggests that the scaffolding around the model matters as much as the model itself when tackling truly difficult research questions.

Is this approach repeatable?

PSI argues the approach is repeatable, not a one-off.

The company designed its scaffolding to be domain-agnostic: a structured reasoning loop where the model explores multiple proof strategies, checks its work against Lean type checker, and backtracks when a line of reasoning hits a dead end.

For the Aaronson-Ambainis proof, the model needed access to a library of mathematical definitions and prior results in quantum complexity theory. The same system could tackle problems in algebraic topology, differential geometry, or combinatorial optimization with minimal retooling.

The founders of PSI, recent graduates from MIT computer science program, see the company as a systematic scientific discovery engine rather than a consultancy that solves one-off problems.

The early access to Anthropic Fable 5.2 gave them a model capable enough to make the approach work, but the scaffolding is the moat.

How long did it take?

The AI system produced the proof in roughly 72 hours, a time frame that would have been unthinkable for a human mathematician tackling an 18-year-old open problem.

Scott Aaronson, who has spent much of his career on related questions, noted that he was surprised OpenAI Mathocalypse effort had not solved it, suggesting the problem was genuinely harder than others that system tackled.

The speed of the result raises a deeper question. If an AI system can solve an open problem in theoretical computer science in three days, how many more currently open problems will fall in the coming months?

PSI answer: most of them. The company is already scaling its system to tackle multiple open problems in parallel, with the goal of demonstrating that AI can accelerate the pace of fundamental discovery by orders of magnitude.

What does this mean for AI research?

The broader implication is that AI-driven scientific discovery is moving from toy problems and benchmarks to real open questions in mathematics and theoretical physics.

PSI result, together with OpenAI Mathocalypse and similar efforts from Google DeepMind, suggests that 2026 may be remembered as the year AI started contributing meaningfully to fundamental science, not just accelerating existing workflows.

The pace is accelerating dramatically, and the number of open mathematical problems may shrink fast in the coming months.

The implications extend beyond mathematics. Problems in drug discovery, battery chemistry, and climate modeling all involve combinatorial search spaces with structures similar to the Aaronson-Ambainis conjecture.

If AI can formalize and verify proofs in these domains, the same approach that solved an 18-year-old quantum problem could accelerate fields that have been waiting decades for breakthroughs.

Common questions

What is the Aaronson-Ambainis conjecture?

It is a fundamental open problem in quantum computing posed in 2008 by Scott Aaronson and Andris Ambainis about the conditions under which quantum computers can outperform classical computers.

How did PSI solve it?

A PSI researcher used early access to Anthropic's unreleased Fable 5.2 model with specialized scaffolding, and the proof was autoformalized in Lean for mechanical verification.

Why is this significant for AI?

It shows AI can solve genuinely open research problems that stumped human researchers for 18 years, not just benchmark tasks. This suggests AI-driven scientific discovery is becoming real.

Sources & further reading

  1. TechCrunch - PSI is trying to make a business out of AI-driven scientific discovery ↗
  2. KuCoin - PSI Uses AI to Solve an 18-Year-Old Quantum Computing Conjecture ↗
  3. SciMonitors - PSI's AI Cracks 18-Year Quantum Computing Puzzle in 72 Hours ↗
How this story was made

Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. The cover is an AI-generated editorial illustration, not a photograph or evidence of an actual event.

Our editorial standards
Back to all stories