Key Moments
AI Crisis in Math: Navier-Stokes, Riemann Hypothesis & Complexity Theory | Lex Fridman Podcast #503
Lex FridmanKey Moments
AI systems recently claimed to have solved hundreds of long-standing mathematical problems, including the quasi-Riemann hypothesis and Navier-Stokes, raising existential questions about human mathematicians' role and the future of mathematical discovery and education, as experts grapple with both awe and dread.
Key Insights
Over the past few weeks, AI systems claimed solutions to hundreds of hard math problems, including Navier-Stokes and the quasi-Riemann hypothesis, a feat comparable to a Fields Medal for a human.
Terence Tao, a Fields Medal winner, expressed horror at AI disrupting years of human research, particularly for graduate students, as knowing a problem's solution can diminish the exploratory journey.
The 'doorman fallacy' suggests reducing a complex job (like a mathematician proving theorems) to its most visible task, overlooking deeper values like building new theories, conceptual connections, and the human story of discovery, which AI currently struggles to replicate.
The quasi-Riemann hypothesis proof by AI involved complex, known ideas combined in a 'shockingly beautiful cascade,' providing a vertical strip (no zeros between 7/8 and 1) where no zeros exist, a significant leap from the 70-year-stagnant 'zero-free region' but still far from the full 1/2 line.
Fermat's Last Theorem was formalized by Anthropic AI in 13 million lines of Lean code from axioms, a task expected to take humans years, demonstrating AI's brute-force capability even if the proof lacks human-like elegance or conceptual depth.
The Navier-Stokes blow-up proof, specifically for smooth external forcing in 3D, constructs a non-physical, specialized vortex that stretches and thins to infinite vorticity while maintaining finite energy, offering mathematical insight but not immediately impacting trillion-dollar engineering problems like drag reduction.
AI solves foundational mathematical problems, challenging human purpose
In a recent and 'historic' event, AI systems reportedly solved hundreds of the most challenging problems in mathematics, including the Navier-Stokes equations, the quasi-Riemann hypothesis, and the unique games conjecture. This rapid progression has sent shockwaves through the mathematical community, eliciting a mixture of awe, terror, and hope among leading figures like Terence Tao. The sheer magnitude of these achievements, which for a human would guarantee a Fields Medal, has prompted mathematicians to re-evaluate their roles and the very nature of mathematical discovery. The core tension lies in the potential for AI to 'solve' math by finding solutions without the human-centric process of understanding, intuition, and the 'why' behind a proof. This moment has sparked intense debate about whether this heralds a 'golden age' of mathematical knowledge or a 'crisis' that could strip away the meaning and motivation for human mathematicians, especially for graduate students and postdocs whose research paths might be 'casually disrupted or destroyed' by readily available AI solutions. The shift forces a reflection on human purpose beyond 'being famous' or winning prizes, emphasizing curiosity, understanding the world, and the inherent satisfaction of tackling challenging problems.
The 'solved axiom' and the value of human understanding over mere solutions
The 'solved axiom' posits that any mathematical problem solvable by human-AI interaction could eventually be solved by AI alone, leading to a scenario where human input merely involves 'pressing a button.' While not fully believed to be a current reality, this axiom frames a critical discussion: If AI can autonomously solve problems, what is the role of humans? Mathematician Alex Kontorovich argues that the value of pure mathematics lies not in theorems existing as 'ones and zeros in data centers,' but in their 'living in the heads of human beings.' He highlights the distinction between 'solving' and 'destroying' math. Ankur Moitra adds that human mathematicians, computationally limited, often develop 'more generalizable and abstractable ways' of reaching solutions, which then explain connections and open new research directions – a process AI may not replicate in the same conceptual manner. The human pursuit of understanding involves developing 'systems of ways of understanding problems' that apply to a 'sea of problems,' not just single theorems, often reasoning through 'sloppy metaphors' that AI might bypass with brute-force statistical relationships across vast datasets. The ability to ask 'the right questions,' which requires significant mathematical insight, remains a uniquely human contribution, guiding the focus of mathematical inquiry towards areas of human care and aesthetic value.
AI's impact on a 70-year-old problem: The quasi-Riemann hypothesis
One of the most astonishing breakthroughs was the AI-driven proof of the quasi-Riemann hypothesis (quasi-RH). The Riemann hypothesis, a central problem in number theory, concerns the distribution of prime numbers and specifically, the locations of the non-trivial zeros of the Riemann zeta function in the complex plane. Riemann famously linked these zeros to the prime numbers, proposing that all non-trivial zeros lie on the 'critical line' where the real part is exactly one-half. Until the AI breakthrough, mathematicians had established a 'zero-free region' — a thin strip where no zeros exist — which had not seen significant improvement in 70 years. The AI proof, described by Alex Kontorovich as 'extraordinarily shocking,' provided a much broader 'vertical strip' where no zeros exist, specifically between 7/8 and 1. This is a monumental step, but the methods used are believed to be fundamentally incapable of reaching the full one-half line, indicating that a genuinely 'new idea' will be needed for the complete Riemann hypothesis. The 'speed of progress' and the nature of this discovery, integrating existing brilliant ideas in a novel sequence, profoundly impacted experts who had deemed such progress impossible within their lifetimes.
Fermat's Last Theorem formalized: A brute-force spectacle
Adding to the sense of a paradigm shift, Anthropic AI successfully formalized Fermat's Last Theorem (FLT) from axioms in 13 million lines of Lean code. This achievement was particularly striking as it was accomplished with minimal human domain expertise, primarily by humans prompting the AI to 'keep going.' This massive formalization, a 'sorry-free' Lean proof, defied expectations, as mathematician Alex Kontorovich had previously estimated such a feat to be years away, given the theorem's complexity and its distance from existing MathLib 'halo' — the well-structured, canonical library of formalized mathematics. The AI's method involved 'jerry-rigging' and blasting through numerous technical steps, creating a functional but inelegant proof. While it provides absolute certainty of correctness, it highlights a qualitative difference: human mathematicians strive for canonical, beautiful, and generalizable proofs that offer deep conceptual understanding, whereas AI can achieve 'sorry-free' solutions through sheer computational power and combination of existing knowledge, even if the resulting structure is cumbersome.
Navier-Stokes blow-up: Mathematical insight vs. engineering application
The AI's proof of the Navier-Stokes blow-up problem for smooth external forcing in three dimensions also captivated the scientific community. The Navier-Stokes equations describe the conservation of momentum in fluids, fundamental for understanding everything from airplane drag to artificial hearts. While conceptually simple (akin to F=MA for fluids), their complexity arises from the 'billion little boxes' of interacting elements, making fluid behavior, especially turbulence, incredibly difficult to predict. The 'blow-up' problem questions whether solutions to these equations can reach infinite velocity or pressure in finite time, a phenomenon analogous to a mathematical singularity. The AI's counterexample constructs a highly specialized vortex that stretches and thins in the third dimension, causing localized infinite vorticity while overall fluid energy remains finite. This mathematical result confirms a theoretical possibility but, as fluid dynamics expert Steven Brunton explains, operates in a 'non-physical' regime, making its immediate impact on 'trillion-dollar engineering problems' like predicting airplane drag or designing artificial hearts negligible. The core engineering challenge of inverse design (e.g., designing a wing for specific lift/drag) remains 'a much harder problem' that requires understanding the full range of possible phenomena from the equations, a goal still far from reach.
The 'doorman fallacy' and the future of mathematical research
Ankur Moitra introduced the 'doorman fallacy' to contextualize the AI's impact on mathematics. This fallacy reduces a complex job to its simplest, most visible task (e.g., a doorman merely opening doors), overlooking the nuanced, valuable aspects like recognizing guests, providing security, and signaling prestige. Similarly, mathematicians are not merely theorem-provers; they build new theories, make conceptual connections, and cultivate a sense of beauty and understanding. While AI excels at the 'visible task' of proving theorems, the deeper, more creative, and communicative aspects of mathematics — such as asking insightful questions, constructing elegant theories, and explaining the 'story' behind a proof — remain profoundly human. The community must 'recenter' its values, recognizing the 'plurality of different aesthetics and goals.' As problem-solving for well-defined conjectures becomes 'cheap,' the orthogonal aspects, like building new abstractions and fostering deep understanding, become 'all that much more valuable.' This shift necessitates rethinking academic structures, from graduate student mentorship to publication models, to encourage broad, curious thinking and collaborative verification, rather than solely focusing on individual 'first through the gate' solutions.
Rethinking education and motivation in an AI-powered world
The rapid advancement of AI poses significant challenges to mathematical education and human motivation. Terence Tao's concern about the 'opportunities lost' in the frantic race to develop technology resonates with the potential for students to lose the 'play and learn' aspect of math when solutions are readily available. Alex Kontorovich highlights the historical precedent of mathematical 'crises,' like the 19th-century foundational crises of calculus, which ultimately led to deeper understanding and rigorous rebuilding. He argues that 'math has been killed multiple times before' and always recovered. However, the current challenge is unique: ensuring humans remain motivated to learn and engage deeply, even when AI can provide answers. The discussion emphasizes the importance of 'critical reasoning skills,' 'advising skills' (how to question information), and distinguishing between factual truth and deeper understanding ('Math is not the fact, it's the why'). Educators face the task of cultivating curiosity and encouraging students to grapple with 'messy, uncertain things' and to model the natural world, rather than solely solving well-defined, worst-case problems. The 'doorman fallacy' implies that education should focus on the irreducible human aspects of learning and discovery, fostering an independence of thought from machines, even as AI becomes an invaluable tool for accelerated learning and exploration.
The nature of mathematical discovery: Invented or discovered?
The philosophical debate about whether mathematics is 'invented' or 'discovered' gains new relevance in the age of AI. Alex Kontorovich offers a pragmatic view: humans 'invent' the questions based on their current knowledge and worldly concerns, but once a question is precisely asked, the 'answer is what God decides the answer is'—meaning the truth is objective and 'discovered.' Steven Brunton adds that math is 'both created and discovered.' Humans create formalisms and approximations to understand the world, and it is 'a miracle' that these descriptions are 'unreasonably effective.' The simplicity and generalizability of fundamental laws like F=MA or E=mc² suggest an underlying mathematical structure that we uncover. However, even seemingly fundamental concepts like integers are abstractions we've created to interact with quantities, taking millennia to 'discover' concepts like zero or negative numbers. The human aesthetic for 'beautiful simplicity' serves as a guide for mathematical inquiry because humans are 'fundamentally simple' and can only handle so much complexity in their minds. AI's ability to find complex solutions might challenge this aesthetic, forcing a re-evaluation of what constitutes 'beauty' or 'understanding' in mathematics if AI-generated proofs are correct but conceptually inscrutable to humans.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Concepts
●People Referenced
Common Questions
Over the last few weeks, AI systems have claimed to solve some of the hardest problems in math, including Navier-Stokes, the quasi-Riemann hypothesis, and the unique games conjecture, among hundreds of other results.
Topics
Mentioned in this video
A landmark result in complexity theory proving that every language in NP has probabilistically checkable proofs, laying the foundation for hardness of approximation.
A set of partial differential equations that describe the motion of viscous fluid substances, central to fluid dynamics. Their solution for existence and smoothness is one of the Millennium Prize Problems.
A famous topological conjecture, proved by Grigori Perelman, which required years of community effort to verify after he posted his papers online.
A weaker form of the Riemann Hypothesis, whose proof by AI systems was a major surprise to mathematicians, indicating significant progress in number theory.
A problem in theoretical computer science, a strengthening of the PCP theorem, whose truth implies optimal hardness of approximation for many constraint satisfaction problems.
A theoretical model of computation used to explore the limits of computability, referenced in relation to Navier-Stokes-like equations simulating Turing machines.
A hypothetical axiom stating that any mathematical problem solvable by human-AI interaction could also be solved by AI alone. This is debated, with Alex Kontorovich not believing it.
A problem whose solution by AI and human collaborators had a major effect on fine-grained complexity theory, proving it to be solvable much faster than previously thought.
A partial differential equation that describes how the quantum state of a quantum system changes with time, cited as an example of general equations whose phenomena are hard to predict.
A famous theorem in number theory, proved by Andrew Wiles, whose recent formalization by AI in Lean with 13 million lines of code was considered a shocking achievement.
A major unsolved problem in theoretical computer science, asking whether every problem whose solution can be quickly verified can also be quickly solved.
A cellular automaton devised by John Horton Conway, used as an analogy for simple rules leading to complex emergent behavior, and for its nature as a universal Turing machine.
A fundamental problem in theoretical computer science, asking if there exists a variable assignment that makes a given Boolean formula true; a canonical NP-complete problem.
An algorithm that computes the discrete Fourier transform of a sequence, mentioned as having seen improvements via AI-generated proofs.
A generalization of the Riemann zeta function, whose zeros on the critical line were part of the AI's proof of the quasi-Riemann hypothesis.
A multiplicative function in number theory, whose twisted by sextic characters was part of the methodology used in the AI's proof of the quasi-Riemann hypothesis.
A theorem describing the asymptotic distribution of prime numbers, related to the zero-free region of the Riemann zeta function.
A set of four partial differential equations that, together with the Lorentz force law, form the foundation of classical electromagnetism, cited as an example of general equations whose phenomena are hard to predict.
A researcher whose work implies optimal hardness of approximation for many constraint satisfaction problems, under the assumption of the Unique Games Conjecture.
The mathematician who introduced the Unique Games Conjecture, mentioned in a personal anecdote about his reaction to its potential solution by AI.
The mathematician famous for proving Fermat's Last Theorem, whose solitary 7-year effort is contrasted with AI's rapid theorem proving.
An MIT faculty member who advised Ankur Moitra to 'pretend you already have tenure' to achieve tenure, highlighting intrinsic motivation.
A mathematician and number theorist at Rutgers, discussing the Riemann Hypothesis, Lean, and automated proof verification.
A colleague of Ankur Moitra at MIT and IAS, who contributed to a weaker version of the Unique Games Conjecture, the two-to-one games theorem.
An American singer-songwriter, quoted for his line, 'I like my town with a little drop of poison,' used as an analogy for embracing a bit of 'weirdness' in life.
A mathematician famous for proving the Poincaré conjecture, and for his unusual behavior of declining prestigious awards like the Fields Medal and Clay Prize.
A mathematician and fluid dynamics expert at the University of Washington, co-host of the discussion, offering insights on Navier-Stokes and applied mathematics.
A co-author with Virgi Vasilievska Williams on the result about the threesome problem.
A mathematician and theoretical computer scientist at MIT, visiting the Institute for Advanced Study, contributing to the discussion on complexity theory and AI's learning mathematics.
A colleague of Alex Kontorovich who axiomatized an argument for what the future of mathematics might look like with AI.
A Fields Medal winner and prominent mathematician, whose quote about AI casually disrupting research efforts and causing a 'loss of meaning' for graduate students is discussed.
A mathematician whose analogy of 'semi-group' is used to describe AI's ability to combine known mathematical moves, contrasting with the 'convex hull' metaphor.
A colleague of Ankur Moitra at MIT, co-authored a result on the threesome problem, impacting fine-grained complexity theory.
Physicist whose theory of general relativity is mentioned in contrast to Newtonian physics, where equations broke down and led to new discoveries like black holes.
A physicist and AI researcher, mentioned for his advice on balancing mainstream work with 'renegade' and 'weird' endeavors in a career.
A mathematician who had a grant to formalize Fermat's Last Theorem over five years, highlighting the unexpected speed of AI's achievement in this area.
A Dutch graphic artist known for his mathematically inspired woodcuts, lithographs, and mezzotints, whose exhibit was opening at the Museum of Mathematics.
Cited for his quote, 'If my laptop can't find it, then neither can the market,' emphasizing the practical limits of computation in economic models.
Physicist who found a 'blowup' solution to Einstein's general relativity equations, which later led to the discovery of black holes, illustrating how equation breakdowns can lead to new discoveries.
A Nobel Prize-winning physicist, quoted for his observation that technology often solves problems radically differently from nature, applied to AI's mathematical proofs vs. human methods.
A colleague of Ankur Moitra at MIT and IAS, who contributed to a weaker version of the Unique Games Conjecture, the two-to-one games theorem.
The institution that offers the Millennium Prize Problems, including Navier-Stokes and the Riemann Hypothesis, with a million-dollar prize for their solution.
The institution where Steven Brunton is a mathematician and fluid dynamics expert.
The institution where Ankur Moitra is a mathematician and theoretical computer scientist.
The institution Ankur Moitra is visiting, indicating a high-level academic environment.
The institution where Alex Kontorovich is a mathematician and number theorist.
Mentioned in the context of mathematical results potentially impacting national security, especially regarding cryptography.
A large-scale physics experiment, cited as an example of a collaborative project involving code, data, and reusable components as academic contributions, rather than just PDF papers.
An interactive theorem prover that allows mathematicians to write formal mathematical proofs and verify their correctness.
A large, well-maintained library of formalized mathematics for the Lean interactive theorem prover, aiming to make mathematics operate more like a science with explicit specifications.
An AI-powered search engine used by Lex Fridman to quickly define the 'doorman fallacy' during the discussion.
An AI research company, mentioned as having released the formalization of Fermat's Last Theorem, among other proofs.
The artificial intelligence research company that released proofs for several hard mathematical problems, triggering the discussion.
A platform for software development, mentioned in the context of how open-source collaboration and the 'pull request' model could be a future model for mathematical research, replacing traditional journals.
More from Lex Fridman
View all 557 summaries
91 minVisiting Taiwan w/ Jensen Huang | Lex Fridman
217 minPsychiatry, Insane Asylums, Mental Illness, ECT, Lobotomies, Freud & Jung | Lex Fridman Podcast #502
316 minDHH: Future of Programming, AI, Agentic Engineering, Vibe Coding & Linux | Lex Fridman Podcast #501
193 minKhabib Nurmagomedov: Dagestan, MMA, UFC, Islam, Conor, Fedor & Football | Lex Fridman Podcast #500
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free