Lex Fridman PodcastTerence Tao: Hardest Problems in Mathematics, Physics & the Future of AI | Lex Fridman Podcast #472
CHAPTERS
- 0:00 – 1:27
Hard problems worth pursuing: the 90% solved, 10% missing frontier
Lex asks Terence Tao about the first research-level problem that truly challenged him. Tao contrasts “impossibly hard” famous conjectures with the more interesting boundary problems where existing methods almost work but fail at a critical last step.
- •Difference between arbitrarily hard vs. strategically hard problems
- •Why the ‘last 10%’ is where new ideas are forced
- •Early exposure to major conjectures (Riemann hypothesis, twin primes) vs. workable research targets
- 1:27 – 6:14
Kakeya needle problem: tiny area/volume, big mathematical consequences
Tao explains the Kakeya problem via the puzzle of rotating an idealized needle in minimal area (2D) and minimal volume (3D). He describes how counterintuitive constructions allow arbitrarily small area in 2D, and why the thin-tube 3D version connects broadly across analysis and geometry.
- •2D needle rotation: surprising ‘arbitrarily small area’ constructions
- •3D thin-tube version: how minimal volume scales with thickness δ
- •Connections to PDE, number theory, geometry, combinatorics
- •Wave packets and tube-packing intuition for wave concentration
- 6:14 – 9:17
From wave concentration to Navier–Stokes: can smooth fluids blow up?
The discussion turns to the Navier–Stokes existence and smoothness Millennium Prize problem: whether smooth initial data can develop singularities. Tao motivates why blow-up would be mathematically possible even if it seems physically absent in everyday fluids.
- •Navier–Stokes regularity vs. finite-time singularities
- •Why mathematicians care about 100% guarantees (not 99.99%)
- •Physical intuition: viscosity damping vs. potential concentration mechanisms
- •Status and importance of Clay Millennium problems
- 9:17 – 13:20
Maxwell’s demon analogy and the ‘conspiracy’ challenge in proofs
Tao uses Maxwell’s demon to illustrate why ruling out extreme, unlikely configurations is difficult mathematically. He parallels this with proving the digits of π have no hidden bias and with excluding rare fluid configurations that might trigger blow-up.
- •Improbable ‘conspiracies’ are hard to rule out rigorously
- •Statistical intuition vs. mathematical certainty
- •Energy transport vs. dissipation framing for fluid behavior
- 13:20 – 19:22
Engineering blow-up in an averaged Navier–Stokes: creating obstructions
Tao explains his 2016 result: constructing finite-time blow-up for an averaged variant of 3D Navier–Stokes by selectively modifying interactions while preserving key conservation laws. The purpose is to show certain proof strategies for the real equations cannot work unless they use special structure absent in the averaged model.
- •Why constructing counterexamples guides real-proof strategy
- •Turning off/redirecting nonlinear interactions to force concentration
- •Obstructions: ruling out broad classes of ‘energy method’ approaches
- •Supercriticality as a key organizing concept
- 19:22 – 26:08
Liquid computers and self-replicating blow-up: Turing machines in fluids
Tao sketches a provocative roadmap: if fluid dynamics can implement computation, one could encode a self-similar, self-replicating mechanism that cascades energy to smaller scales and blows up. He connects the idea to engineered delays/gates reminiscent of circuit design.
- •Why naive scale-cascades fail in 3D and need ‘airlock’ delays
- •Circuit-building intuition (gates, clocks) applied to nonlinear PDE
- •Concept of a fluid Turing machine and Von Neumann self-replication
- •Practical obstacles: noise, analog error correction, powering down states
- 26:08 – 30:16
Conway’s Game of Life: emergence, logic gates, and computability parallels
To justify computation-like behavior arising from simple rules, Tao discusses cellular automata—especially the Game of Life—where gliders, glider guns, and logic gates can be built. The key caveat: such complexity usually requires carefully engineered initial conditions.
- •Gliders as traveling localized structures; glider guns as sources
- •Constructing AND/OR gates and eventually Turing machines in Life
- •Self-replicating structures discovered via community effort
- •Engineered structure vs. typical random initial states
- 30:16 – 35:11
Structure vs randomness, Szemerédi, and the infinite monkey intuition
Tao introduces the recurring mathematical dichotomy between structured objects and random-looking ones. He uses Szemerédi’s theorem and the infinite monkey theorem to explain why patterns inevitably emerge in sufficiently large or dense sets, while emphasizing the difference between existence and quantitative bounds.
- •Most mathematical objects ‘look random’; structure is rare but powerful
- •Inverse/structure theorems: diagnosing hidden structure
- •Szemerédi’s theorem: dense sets contain long arithmetic progressions
- •Infinity vs. finitization: qualitative truth vs. quantitative rates
- 35:11 – 38:07
Infinity as idealization: limits, pitfalls, and why finitizing is messier
Lex presses on how humans handle infinity; Tao frames it as a clean abstraction of ‘very large’ or ‘very small.’ He notes classic analytic pitfalls of infinite processes (rearrangements, convergence) and why turning infinite statements into finite quantitative ones improves intuition but complicates proofs.
- •Infinity/zero as simplifying idealizations (spherical cows)
- •Where infinity breaks naive reasoning (series rearrangements)
- •Epsilon–delta discipline as a safety mechanism
- •Finitization yields intuition and bounds at the cost of complexity
- 38:07 – 47:49
Math vs physics: models, observations, and theories as data compression
Tao describes science as interplay among reality, observations, and mental models, with mathematics exploring consequences inside models. He discusses theory–experiment feedback loops, the rise of experimental mathematics, and why good theories act as powerful compression of enormous datasets.
- •Math studies models; science proposes/validates models via observations
- •Experimental math historically (Gauss) and today (computers, AI)
- •Combinatorial explosion limits brute-force exploration
- •A theory as compression: few parameters explaining petabytes of data
- 47:49 – 51:52
Universality: why simple laws emerge—and when they fail (finance example)
Tao offers universality as a partial explanation for mathematics’ effectiveness: macroscopic behavior often depends on few parameters despite enormous microscopic complexity. He highlights central limit theorem as a canonical universality mechanism and warns about correlated/systemic effects where Gaussian assumptions break down.
- •Universality reduces dependence on microscopic details
- •Central limit theorem and the ubiquity of Gaussians
- •Failure modes: correlations and systemic shocks
- •2008 financial crisis as a cautionary tale about pretty models
- 51:52 – 1:00:43
Unifying threads in mathematics, fox vs hedgehog styles, and proof craftsmanship
Lex asks about deep structure across mathematics; Tao emphasizes progress via unexpected connections (e.g., numbers and geometry via Descartes). He discusses personal style as a ‘fox’ who transfers tools across fields, and shares Conway’s idea of ‘extreme proofs’ and proof-writing as craftsmanship akin to good code.
- •Unification as a driver of mathematical progress (analytic/algebraic geometry)
- •Fox vs hedgehog: breadth via analogies vs depth via scholarship
- •Reproving results with preferred tools to understand other methods
- •Conway’s ‘space of proofs’ and optimizing proofs for elegance/clarity
- 1:00:43 – 1:07:09
Beauty of Euler’s identity and the importance of choosing the right concepts
Tao explains why Euler’s identity is beautiful beyond ‘famous constants,’ focusing on the unification of growth (exponential) and rotation (imaginary exponent). He expands to how physics evolves by elevating the right fundamental quantities, such as energy/Hamiltonians, enabling cross-domain transfer of ideas.
- •Euler identity as a bridge between growth/decay and rotation
- •Notation collisions as signals of having the right conceptual objects
- •Energy/Hamiltonian reformulation clarifying classical mechanics
- •Noether’s theorem: symmetry ↔ conservation in classical and quantum settings
- 1:07:09 – 1:13:10
Toward a theory of everything: why unification is hard but plausible
The conversation turns to unifying quantum mechanics and general relativity. Tao is optimistic given physics’ history of unification, but notes the lack of experimental leverage and the difficulty of finding the right mathematical objects to replace naive spacetime coordinates at tiny scales.
- •Historical precedents: Maxwell, Newton, and unification patterns
- •Why GR + QM are ‘too successful’ to easily falsify in accessible regimes
- •Need for new fundamental mathematical language/objects
- •Mathematics as pre-adaptation: tools often exist before physics needs them
- 1:13:10 – 1:16:34
General relativity adjacent work: wave maps, gauges, and visual intuition
Tao describes his work on critical wave maps (sigma models), studying whether energy can concentrate to form singularities. He highlights gauge transformations as a way to ‘stabilize’ nonlinear dynamics and shares a personal story of physically visualizing vector fields to find the right coordinate change.
- •Wave maps as a nonlinear wave constrained to a manifold (sphere)
- •Critical vs supercritical behavior and implications for blow-up
- •Gauge transformations reducing apparent nonlinearity
- •Embodied/visual reasoning as a legitimate mathematical tool
- 1:16:34 – 1:20:01
How Tao tackles difficult problems: strategic ‘cheats,’ decomposition, and tools
Asked about his thinking process, Tao emphasizes simplifying a problem by turning off most difficulties, solving pieces, then recombining them. He describes heavy use of blackboards, diagrams, and increasing use of computers for exploration—accelerated by AI-assisted coding.
- •Strategic simplification: remove 9 of 10 difficulties first
- •Iterative recombination: solve sub-difficulties then merge
- •Blackboards and bespoke doodles for global situational awareness
- •AI lowers friction for quick computational experiments
- 1:20:01 – 1:27:40
Lean proof assistant: formal verification, pedantry, and why it can be worth it
Tao introduces Lean as a programming language that produces proof certificates, enabling machine-checked mathematical correctness. He compares Lean to explaining proofs to a maximally pedantic colleague, but shows how the payoff can be huge: easy refactoring, localized failures, and improved collaboration.
- •Lean outputs proofs/certificates, not just results
- •Type discipline and explicitness force clarity about objects
- •Formalization cost ~10× handwritten proof (currently)
- •Benefits: refactoring constants, compiler highlights exact breakpoints
- 1:27:40 – 1:43:15
Scaling collaboration: blueprints, trustless math, and the Equational Theories Project
Tao explains how Lean enables distributed, fine-grained collaboration—down to a few stuck lines—because all context is machine-traceable. He highlights blueprint-style decomposition and describes his Equational Theories Project: 22 million implication/counterexample problems in algebra solved with ~50 contributors.
- •Hover-to-define and traceability make proofs easier to navigate than papers
- •Atomic collaboration: share a local snippet + context for help
- •Blueprints as supply chains for proof construction
- •Equational Theories Project: 22M problems, near completion, 50 authors
- 1:43:15 – 1:47:45
AlphaProof and AI theorem proving: what’s impressive, what doesn’t scale yet
Discussing DeepMind’s AlphaProof, Tao praises the demonstration of what is possible but emphasizes current inefficiencies: enormous compute and limited scalability to longer proofs. They also discuss why natural-language-to-formal translation is hard and why reinforcement learning signals are cleaner for numeric answers than long-form proofs.
- •Compute cost and exponential difficulty with proof length
- •Natural language ↔ formal language translation remains extremely hard
- •Formal proofs enable automatic verification; informal proofs need human grading
- •Why RL works better for short/closed-form answers than full proofs
- 1:47:45 – 2:04:47
What humans still contribute: mathematical ‘smell,’ strategy evaluation, and phase shifts
Tao argues current AI struggles with recognizing wrong turns and generating reliable, globally coherent proof strategies—what he calls a human ‘sense of smell.’ He predicts phase shifts as tooling improves (formalization cost dropping below handwritten), analogous to LaTeX adoption, changing publishing and refereeing norms.
- •Key human edge: viability assessment of approaches and decompositions
- •AI outputs can be ‘odorless’—plausible yet subtly wrong
- •Phase shift prediction: formalization becomes the default when cost < 1×
- •Potential future: journals prioritize certified correctness, focus referees on significance
- 2:04:47 – 2:11:54
Perelman and the Poincaré conjecture: Ricci flow, singularities, and new invariants
Tao explains the Poincaré conjecture via simple connectivity and classification of 3D manifolds, then outlines Hamilton’s Ricci flow approach. He describes Perelman’s breakthrough: introducing new scale-invariant quantities (entropy/reduced volume) to make the analysis effectively critical and to classify/manage singularities via surgery.
- •Simply connected surfaces vs 3D manifolds; why 3D is uniquely hard
- •Ricci flow as ‘inflating/smoothing’ geometry toward canonical shapes
- •Main obstacle: classifying and handling all possible singularities
- •Perelman’s new monotone quantities that tame supercritical behavior
- 2:11:54 – 3:14:33
Failure modes, persistence, and mathematical psychology—plus primes as a haunted frontier
Tao discusses how mathematicians handle dead ends, including his own tendency to switch problems (fox style) and the usefulness of strategic assumptions. He then transitions to prime-number problems: why Riemann is currently out of reach, why additive–multiplicative interactions are hard, and how twin primes resist ‘robust’ methods despite progress on bounded gaps.
- •Tactics for setbacks: switch problems, add temporary assumptions, forward reconnaissance
- •Near-solution mistakes can fuel eventual breakthroughs (missing ‘13th term’ story)
- •Primes as additive vs multiplicative structures; mixing them causes deep difficulty
- •Twin primes vs arithmetic progressions: fragile vs robust patterns; bounded gaps via pigeonhole-style weighting and the parity barrier
