“I Heard That a Robot Solved the Math Problem You Worked On”

Summary

1. AI is beginning to solve major open problems

Scott Aaronson discusses AI-generated mathematical results, including an OpenAI release of hundreds of proposed solutions to open problems. One prominent example was a claimed proof of the Unique Games Conjecture, a problem Aaronson’s wife, complexity theorist Dana Moshkovitz, had worked on for years. She and other experts were trying to understand and verify the proof.

Aaronson stresses that a proposed solution is not the same as a result the research community understands. Many AI-generated manuscripts were difficult to read and poorly explained, so mathematicians still had work to do: checking the arguments, identifying their key ideas, and writing clearer accounts.

2. What mathematicians do may change—but understanding still matters

Aaronson agrees that mathematical research is not only about obtaining theorem statements. Open problems are valuable partly because working toward their solutions creates new ideas, understanding, and questions. If AI can produce an answer without the human process that traditionally generated those insights, something important may be lost.

He nevertheless argues that AI’s ability to solve problems cannot simply be ignored or reversed. The technology will become widely available, so mathematicians will need to adapt and place human understanding at the center of mathematical culture. He also considers it reasonable to say that models “understand” mathematics in an operational sense: solving difficult problems appears to require an internal representation of the subject. He distinguishes this from consciousness or sentience.

AI explanations may improve, but current systems often seem poorly calibrated to their readers. They can spend too much time on basics and too little on the central, difficult steps. AI tools can also help experts read and clarify AI-generated work, while humans produce more accessible explanations.

3. Credit, responsibility, and how results should be published

The discussion compares different ways of releasing AI-assisted research:

• OpenAI’s approach, as described in the interview, was to publish many solutions, often with the company listed as author and without an individual who had fully understood each result.
• Anthropic’s approach, in a cited example, was to involve human experts who examined results, wrote clearer presentations, answered questions, and accepted responsibility for the work.

Aaronson describes an emerging norm for research papers: disclose how AI was used. If researchers use AI to find important ideas but then understand the result, explain it, and take responsibility for it, they can reasonably be authors. When nobody has yet checked or understood an AI-generated result, authorship and credit become harder to assign.

He criticizes the idea that someone should simply submit a problem to a model, publish the answer, and expect recognition while leaving others to do the work of understanding it. At the same time, verifying, simplifying, and explaining a result can itself merit credit.

4. Effects on researchers and mathematical education

Aaronson describes a mix of excitement and grief in the mathematical community. Some researchers miss the traditional experience of pursuing a hard problem over months or years and reaching a solution through their own efforts. The disruption may be especially difficult for graduate students, who have often established themselves by solving well-defined open problems.

He argues that the community should not respond by pretending the technology can be stopped. Instead, it should work out how to preserve mathematical culture, make the field attractive to students, and continue teaching problem-solving skills. Those skills remain important for judging whether AI is wrong and for taking on more advanced work.

Aaronson uses chess as a cautiously hopeful comparison: computers became better than humans, but human chess continued to thrive as a sport, pastime, and culture, with AI changing how people train and analyze games. He suggests that people may continue to value doing and understanding mathematics even when AI can solve problems more effectively.

5. Open problems are open to anyone—but disclosure still matters

Aaronson rejects the implication in a Human Mathematics Association statement that mathematicians had not been asked to do AI’s work. The point of an open problem, he says, is that anyone may attempt it: a professional, a student, an unknown amateur, or a corporation. He cites Ramanujan as an example of an outsider who made major contributions.

That does not mean AI companies should be exempt from scrutiny. Aaronson says they should be held to high standards, particularly on safety, security, and consistency. But the mathematical community cannot expect the public or companies to preserve the old way of doing research for its sake.

6. Hypothetical disclosure of a practical P = NP proof

Asked whether a company should release a proof of P = NP, Aaronson says the implications would depend heavily on the algorithm’s practical efficiency. A proof involving an algorithm with enormous exponents or constants could be theoretically important without immediately breaking real-world systems.

A genuinely practical algorithm could have sweeping consequences:

• Much of today’s internet cryptography might become vulnerable.
• Digital signatures and cryptocurrency security could be compromised.
• AI development might accelerate if important computational tasks became much easier.

Aaronson connects this to responsible disclosure in cybersecurity, where a discoverer may notify affected parties and allow time to respond before publishing. A breakthrough affecting widely used cryptography would raise harder questions: whom to notify, how to demonstrate the discovery without exposing the method, and whether governments or companies should make the first decisions. He also notes that cryptographic vulnerabilities could arise without P = NP—for example, through a faster algorithm for a particular problem.

7. Why AI has advanced further in mathematics than in much of physics

Aaronson suggests that AI succeeds especially well on problems with clearly stated claims and mechanically checkable answers. Such problems allow systems to test many candidate approaches and receive a reliable signal about whether an answer works.

Many physics questions are less sharply defined. Researchers must decide whether a model is appropriate, whether an approximation is acceptable, and how well a theory captures the real physical situation. Aaronson therefore expects mathematical physics—where some questions can be formulated precisely—to be affected earlier than fields that rely more heavily on ambiguous models or experiments.

8. A result overturning Aaronson and Kuperberg’s conjecture

Aaronson describes a problem he posed with Greg Kuperberg about unitary synthesis: whether every unitary transformation on (n) qubits could be implemented efficiently using access to a suitable classical Boolean function as an oracle. They had expected the answer to be no, partly because they could not see how to avoid an exponential number of queries.

Aaronson says an OpenAI paper presented a solution in the opposite direction: such a Boolean function could allow the unitary transformation to be applied efficiently. He says subsequent AI-assisted work improved the result. The result matters because, under the assumption P = PSPACE, it would imply that a broad class of quantum transformations could be performed efficiently.

9. Computational complexity and black-hole firewalls

The conversation connects that result to the black-hole information problem. Hawking radiation is expected, under quantum mechanics, to contain information about what fell into a black hole, but that information would be extremely scrambled. Recovering it would require applying a suitable transformation to the radiation.

Aaronson recounts the Harlow–Hayden argument that decoding the radiation would be computationally too difficult to perform before the black hole evaporated. That difficulty helps prevent a “firewall” scenario, in which decoding the radiation and then falling into the black hole would lead to a destructive effect at the event horizon.

The new connection Aaronson describes is conditional: if P = PSPACE, then the decoding task would be easy, and the computational-complexity basis for the Harlow–Hayden argument would fail. This does not establish that P equals PSPACE. Rather, Aaronson emphasizes how hard it may be to prove the relevant task is difficult, and how that difficulty connects to major open questions in complexity theory.

10. Maxwell’s demon, entropy, and computational hardness

Aaronson also outlines an argument about whether computational complexity may help explain the apparent second law of thermodynamics in certain systems:

• In a gas, particles initially gathered in one region will generally spread out. A return to the original, low-entropy arrangement is possible in principle but typically takes an extremely long time.
• Maxwell’s demon appears able to reduce entropy by observing particles and selectively opening or closing a gate.
• The standard resolution is that the demon must store or erase information. Entropy is not eliminated; it is transferred to the demon’s memory or environment.
• In a reversible classical system, suppose the demon knows the governing rules and records observations about the particles. It may need to solve a difficult computational problem to infer which initial state produced those observations.
• If P = NP, or if certain related computational barriers such as one-way functions fail, the demon might solve that problem efficiently and compress its stored memory.
• In that case, what looked like thermodynamic entropy could be exposed as pseudoentropy: data that appears random but has a shorter underlying description. The second law would not necessarily be violated, but its apparent effect could be circumvented in that system.

Aaronson adds important qualifications: chaos, the need for infinite measurement precision, or quantum uncertainty could prevent this strategy from working. The argument is mainly relevant to suitable classical systems that are not disrupted by those effects.

11. Aaronson’s criticism of Integrated Information Theory

At the end, Aaronson criticizes Integrated Information Theory (IIT), associated with Giulio Tononi. IIT uses a quantity called phi to assign a level of consciousness to a physical system.

Aaronson’s objections are that the definition of phi has changed between versions of the theory and that, in his view, it is not convincingly derived from the stated axioms. He also points to systems such as large grids of XOR logic gates that could receive very high phi values despite seeming to have no meaningful conscious activity. He questions what such a system would actually be conscious of and argues that IIT’s predictions can be absurd rather than merely too vague to test.

Speakers and sources featured

• Curt Jaimungal — host and interviewer.
• Scott Aaronson — computer scientist and interview guest.
• Dana Moshkovitz — Aaronson’s wife and a complexity theorist; discussed in the interview but not heard speaking.
• Aaronson’s nine-year-old son — quoted by Aaronson in an anecdote; not heard speaking.

People and works cited include:

• Subhash Khot — Unique Games Conjecture.
• Greg Kuperberg — unitary-synthesis problem.
• Timothy Gowers and Edward Witten — members of a cited advisory council.
• Tristan Buckmaster and Levent Alpoge — researchers mentioned in connection with Navier–Stokes.
• Vijay Vasil(e)va and Josh Alman — human researchers cited in Anthropic’s approach.
• Daniel Harlow and Patrick Hayden — black-hole decoding argument.
• Stephen Hawking — black-hole information.
• Leo Szilard, Rolf Landauer, and Charles Bennett — Maxwell’s demon and information.
• Giulio Tononi — Integrated Information Theory.
• Srinivasa Ramanujan — example of an outsider mathematician.

Companies and groups mentioned include OpenAI, Anthropic, and the Human Mathematics Association.

Sponsor advertisements for Base44, Mod, and CleanMyMac are interspersed in the subtitles but are not part of the substantive discussion.