AI Crisis in Math: Navier-Stokes, Riemann Hypothesis & Complexity Theory | Lex Fridman Podcast #503
📚 Main topics
- AI’s sudden mathematical advancesThe guests discuss AI results on Navier–Stokes, the quasi-Riemann hypothesis, the Unique Games Conjecture, and many other problems, and how these breakthroughs have unsettled expectations across mathematics. 0:00
- What mathematicians do beyond proving theoremsThe conversation considers whether mathematical work is only about solving well-defined problems, or also about asking fruitful questions, creating concepts, understanding proofs, and finding beauty and connections. 13:29
- Formal proofs and collaborative mathematicsAlex explains Lean, an interactive theorem prover, and Mathlib, its shared mathematics library; formalization can verify proofs while a more explanatory account communicates the ideas behind them. 55:00
- The Riemann hypothesis and prime numbersAlex describes how the zeta function encodes information about the primes and explains that AI’s quasi-Riemann results establish zero-free regions, but do not prove the full Riemann hypothesis. 1:06:11
- Navier–Stokes, turbulence, and applicationsSteven explains the fluid equations and the AI-produced blowup result, while emphasizing that it does not make practical fluid simulation or engineering design easier by itself. 1:29:39
- Complexity theory and the Unique Games ConjectureAnkur outlines satisfiability, the PCP theorem, and why the Unique Games Conjecture shaped expectations about the limits of approximation algorithms. 2:04:43
- AI, learning, and natural dataThe guests discuss why real-world data can make problems tractable even when worst-case complexity says they are hard, and how subtle patterns in model training can produce unexpected behaviors. 2:49:36
✨ Key takeaways
- A theorem is not the same as understandingA formally checked proof can establish that a result follows from its assumptions, but mathematicians also want to know why it works and what ideas it reveals. 1:00:17
- AI expands the reachable landscapeEven when a proof combines existing techniques rather than introducing a wholly new concept, AI can connect ideas at a scale and speed that changes what researchers believe is attainable. 33:37
- Verification and communication are different tasksLean can check detailed reasoning, while papers and talks can focus more on the motivating story, the key choices, and the insights gained along the way. 1:00:49
- Applications do not follow automatically from a mathematical resultThe Navier–Stokes proof is important for understanding the equations, but it does not directly provide better predictions or designs for real-world flows. 1:48:31
- Worst-case difficulty does not describe every practical instanceReal data and specific problem structures can make tasks much easier than general complexity bounds suggest, though identifying those structures remains a challenge. 2:42:39
- AI raises institutional and career questionsThe surge of results puts pressure on students, journals, research norms, and the ways mathematicians recognize contributions. 25:09
🧠 Lessons learned
- Stay curious and build fundamentalsThe guests advise students to learn core subjects, use AI to test and deepen their understanding rather than simply obtain answers, and preserve the ability to think independently. 3:18:12
- Make room for uncertaintyWith the old academic playbook becoming less reliable, students and researchers should develop their own sense of purpose and focus on work they find meaningful. 3:24:34
- Do not reduce a profession to its most visible outputThe “doorman fallacy” is a reminder that automating an obvious task does not necessarily replace the less visible value of a whole role; proving theorems is not all mathematicians do. 19:55
- Protect time for deep thought and playThe conversation emphasizes that exploration, patient effort, and playing with ideas can be valuable even when they do not immediately produce a landmark result. 2:16:15
- Use AI where outputs can be checkedThe guests favor applications such as proof verification and design exploration when errors are detectable, while cautioning against handing high-stakes decisions to systems without adequate safeguards. 3:02:35
- Keep human connection centralThe guests point to family, community, and collaboration as sources of meaning that remain important even as machines become more capable. 3:32:40
🏁 Conclusion/next steps
- Rework research and education practicesThe discussion calls for evolving how mathematics is reviewed, shared, taught, and credited, while preserving careful verification and opportunities for students to learn. 52:53
- Choose enduring questions and valuesRather than rushing to match AI output, researchers are encouraged to pursue problems they expect to care about over the long term and to define success beyond prizes or publication counts. 2:26:23
- Treat AI as a powerful collaborator, not a substitute for judgmentThe guests see substantial potential in using AI to learn, explore, and solve problems, while stressing the continuing importance of human understanding, curiosity, and responsibility. 3:11:19
Transcript excerpt
0:00 - All right, so this is a kind of emergency episode I think it's fair to say, about the state of mathematics and artificial intelligence. A kind of crisis moment in the history of mathematics. Over the last few weeks, machines claimed to have solved some of the hardest problems in the history of math. Navier-Stokes, the quasi- Riemann hypothesis, the unique games conjecture, and hundreds of other results. So here with me, we have Steven Brunton,
0:33 who's a mathematician, fluid dynamics expert at University of Washington, who can speak to Navier-Stokes, turbulence, all things fluid, and many other topics. Also, here's Ankur Moitra, a mathematician and theoretical computer scientist at MIT, visiting Institute for Advanced Study in Princeton this year, who can speak to many things, unique games, conjecture, and complexity theory in general, and the mathematics of how the AI learns and
1:05 all things theoretical computer science. And Alex, a mathematician, number theorist at Rutgers, who can speak to also very many topics, including Riemann hypothesis and Lean, the automated proof and the automated verifier and checking and so on, and so many other topics in mathematics. All three are brilliant. All are very kind to show up this early morning and talk to me, assemble this kinda multi-agent mini swarm of human brains to try to interpret what happened
🔒 The full, searchable transcript is available with Pro.
Questions & Answers
Common questions about this video
¿Por qué Alex Kontorovich considera que la llegada de la IA a las matemáticas no necesariamente destruye la disciplina?
Porque el valor de las matemáticas puras no consiste únicamente en obtener teoremas, sino también en comprender por qué son ciertos, desarrollar ideas y explorar conceptos. Aunque la IA pueda resolver problemas, Kontorovich sostiene que la comprensión matemática sigue teniendo valor cuando vive en la mente de las personas. 12:59
¿Qué es Lean y para qué sirve Mathlib?
Lean es un asistente de pruebas interactivo que verifica si cada paso de una demostración se deduce lógicamente de los anteriores. Mathlib es la biblioteca de matemáticas formalizadas que se utiliza en el ecosistema Lean, para que los proyectos puedan apoyarse en resultados ya establecidos en vez de reconstruirlos desde cero. 57:04
¿Qué demuestra la cuasihipótesis de Riemann y cómo se diferencia de la hipótesis de Riemann completa?
La hipótesis de Riemann completa afirma que los ceros relevantes de la función zeta tienen parte real igual a un medio. La prueba de la cuasihipótesis de Riemann descrita en la conversación establece una región libre de ceros entre las partes reales siete octavos y uno; llegar de siete octavos a un medio requeriría una idea nueva. 1:24:20
¿Qué problema de Navier-Stokes se probó y por qué el resultado no cambia de inmediato la ingeniería de fluidos?
Se probó una construcción con una fuerza externa suave que puede producir una singularidad de vorticidad en un flujo tridimensional. Aunque es un resultado matemático importante, ocurre en un régimen no físico y no hace que, en la práctica, simular o controlar fluidos como la turbulencia sea más fácil para diseñar alas o corazones artificiales. 1:45:12
¿Qué es la conjetura de juegos únicos y por qué su resolución importa en ciencias de la computación teórica?
Es una conjetura relacionada con la dureza de aproximación de ciertos problemas de optimización. Si se cumple, ayuda a determinar límites óptimos para algoritmos de aproximación en una gran familia de problemas, por lo que su demostración puede cambiar cómo se organiza y estudia ese campo. 2:09:54
¿Por qué Ankur Moitra advierte que los modelos de aprendizaje profundo pueden resolver problemas de formas distintas a las humanas?
Los sistemas pueden aprovechar patrones estadísticos, atajos o «puertas traseras» presentes en los datos, en lugar de basarse en las características que una persona usaría y podría explicar. Por eso, no conviene asumir que la IA razona como un matemático humano ni trazar con demasiada confianza una frontera sobre lo que puede hacer. 32:02
¿Qué consejo dan los invitados a estudiantes que sienten ansiedad ante los avances de la IA?
Recomiendan conservar la curiosidad, aprender los fundamentos y usar la IA para comprobar y ampliar la comprensión, no solo para obtener respuestas rápidas. También sugieren tener motivos intrínsecos para estudiar, tolerar la incertidumbre y pensar qué harían si ya hubieran alcanzado sus metas profesionales. 3:18:12
Related summaries
DHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501
Visiting Taiwan w/ Jensen Huang | Lex Fridman
Sundar Pichai: CEO of Google and Alphabet | Lex Fridman Podcast #471
Saagar Enjeti: Trump, MAGA, DOGE, Obama, FDR, JFK, History & Politics | Lex Fridman Podcast #454
Unlocking Persistent Motivation & Extreme Resilience | Ross Edgley
