Terence Tao: Hardest Problems in Mathematics, Physics & the Future of AI | Lex Fridman Podcast #472
Summary
Lean could become the trust layer that turns unreliable AI into scalable mathematical infrastructure. Formalizing a proof currently takes Tao roughly 10 times longer than writing it by hand, but certificates enable “trustless mathematics,” atomic collaboration, and safe refactoring: changing a theorem’s constant from 12 to 11 left 90% of thousands of lines intact and exposed only the broken dependencies. His phase-change threshold is below 1×, when formal-first papers, faster refereeing, and a potentially exponentially growing mathlib become the default workflow.
Near-term AI value lies in workflow leverage, not autonomous discovery. AlphaProof’s silver-medal-equivalent IMO performance was impressive but required human formalization and roughly three days of Google server time for one high-school problem; proof search still deteriorates exponentially with length. Current tools can supply perhaps “30, 40%” of mathematical skills—coding, calculation, search, autocomplete—but lack the human “sense of smell” that detects a beautiful-looking argument built on a stupid error.
Formal verification already lets research operate at a scale conventional publishing cannot support. Tao’s Equational Theories Project generated about 22 million implication problems among roughly 4,000 algebraic laws; around 50 contributors settled all but two, and a pen-and-paper proof for one of the remaining cases was being formalized. The emerging model resembles a modern supply chain: a blueprint decomposes one theorem into independently verifiable nodes, opening research to distributed specialists, students, programmers, and eventually AI agents.
Navier–Stokes is fundamentally a tail-risk problem: ordinary water behaves well, but mathematics must eliminate every engineered catastrophe. Tao’s averaged equation demonstrates a finite-time energy cascade by selectively closing interaction channels, proving that conservation of energy and viscosity alone cannot establish regularity. His more speculative route is a self-replicating “water-punk” computer that transfers its energy into progressively smaller copies—physically unbuilt, error-prone, but not obviously forbidden by the equations.
The episode’s most investable modeling lesson is that elegant averages fail when correlations become systemic. Universality compresses roughly (10^{23}) gas particles into a handful of variables, and Gaussian laws work when many inputs are sufficiently independent; 2008 showed what happens when mortgage defaults move together instead. Tao’s test is blunt: a model with 10 parameters explaining 10 observations is useless, while a compact theory explaining petabytes of observations earns credibility—but only within its stated assumptions.
The famous number-theory problems sit at sharply different distances from available tools. Bounded-gap methods prove infinitely many prime pairs separated by at most 246, but twin primes require reaching the 50% “parity barrier”; Tao expects substantially closer partial results within 10 years, not necessarily a complete proof. He sees the Riemann hypothesis as needing something “out of left field,” Collatz as vulnerable to one engineered exception despite 99%-type results, and P versus NP as leaning toward inequality while carrying unusually many no-go theorems.
Human advantage remains problem selection, conceptual compression, and productive collaboration across styles. Tao identifies as a fox who imports tools between fields, while hedgehogs command one domain deeply; the best teams combine both. His durable career advice follows the same logic: learn transferable abstraction and problem-solving, try something even when no standard method applies, and treat failures as information—because future tools will automate routines faster than they automate judgment.
Deep dive
1. Kakeya’s remaining 10% made a toy problem consequential
Tao distinguishes impossibly famous questions from problems “on the boundary” of technique, where existing methods complete “90% of the job” and invention supplies the last 10%. Kakeya caught his attention as a PhD student, became a major part of his early research, and, he says, had “just got solved.”
Sōichi Kakeya’s roughly 1918 puzzle asks how little planar area an ideal needle needs to reverse direction. Rotating around its center sweeps area (\pi/4), while a three-point turn uses (\pi/8); Besicovitch’s surprising construction showed that sufficiently elaborate back-and-forth motion can make the required area arbitrarily small.
In three dimensions, Tao imagines a zero-thickness Hubble Space Telescope pointing toward every star. The real question gives the telescope thickness (\delta) and asks how its minimum swept volume shrinks as (\delta) approaches zero; the conjectured answer was that it declines only very slowly, roughly logarithmically.
2. Geometric tube packing controls whether waves can concentrate
Kakeya matters because localized wave packets trace tubes through space-time. A dispersed wave can focus at one point and then defocus, just as the time reversal of a pebble’s expanding pond ripples creates converging waves and a final splash; the underlying wave equations permit that reversal.
Had directionally varied tubes packed far more efficiently than conjectured, waves could sustain many concentrations across space-time rather than one isolated focus. Their amplitude might then enter a regime where linear wave laws fail and nonlinear effects produce a singularity or “blowup.”
Tao is careful about the implication: Kakeya is not a direct solution to Navier–Stokes. It sharpens understanding of tube geometry and wave concentration, which “would indirectly probably help us understand” singularity formation in harder nonlinear equations.
3. Navier–Stokes asks mathematics to eliminate its Maxwell’s demon
The Clay problem concerns incompressible Navier–Stokes: can a smooth initial velocity field develop infinite velocity or another singularity in finite time? Only one of the seven Millennium Prize problems—the Poincaré conjecture—has been solved, so this remains the literal “million-dollar question.”
Lex stresses that fluids are practically consequential, not merely abstract. Tao separates incompressible equations for water from compressible equations for air, noting that weather prediction combines large data-gathering systems with repeated approximate solutions of fluid equations.
Mathematicians cannot settle for water behaving safely 99.99% of the time. Tao invokes Maxwell’s demon: molecular collisions almost surely mix gases, yet a fantastically coordinated sequence might separate them; similarly, the digits of (\pi) look unbiased, but present methods cannot exclude a hidden “conspiracy” favoring one digit.
Real water calms because viscosity dissipates dispersed energy, yet Tao says recent opinion has “drifted” toward believing that carefully prepared configurations might form singularities. The claim remains unproved, and ordinary bathtub behavior says little about the engineered exceptional case.
4. A finite-time cascade must outrun viscosity without dispersing energy
The dangerous scenario repeatedly transfers most energy into a smaller, faster eddy. If every stage takes perhaps half as long as the previous one, infinitely many contractions can fit into finite time, concentrating energy at one point in a self-similar blowup.
Normal turbulence frustrates that plot: one large eddy divides into perhaps three or four smaller eddies, each divides again, and the energy disperses until viscosity dominates. Blowup requires keeping the energy unusually coherent while accelerating the cascade faster than dissipation can respond.
Earlier regularity arguments tried to combine conservation of energy with viscosity, but purported proofs repeatedly hid sign errors or subtle gaps. Tao’s response was diagnostic: build a nearby equation that conserves energy yet unmistakably blows up, revealing which broad proof strategies cannot possibly suffice.
In his averaged three-dimensional Navier–Stokes model, Tao selectively turned off interactions that sent energy into unwanted eddies and kept the channel driving it downward in scale. “I basically engineer a blowup by changing the laws of physics,” creating an obstruction: any true regularity proof must exploit structure removed by his modification.
5. Supercriticality explains why fine scales defeat aggregate models
Navier–Stokes is a tug-of-war between linear viscosity, which smooths motion, and nonlinear transport, which moves energy. In a supercritical equation, transport grows relatively stronger at smaller scales, precisely where viscosity needs to regain control.
In two dimensions, Ladyzhenskaya proved in the 1960s that blowup does not occur. Tao describes the two-dimensional equation as critical—the competing effects retain comparable strength across scales—while three-dimensional Navier–Stokes is supercritical and therefore beyond the strongest regularity technology.
Follow-on research has produced many blowups in other supercritical equations. Tao treats the critical/subcritical/supercritical classification as a qualitative dividing line between systems that remain controllable and systems where “all kinds of bad things” can occur.
Planetary motion tolerates aggregation: the Moon or Mars can often be approximated as a point mass. Weather cannot be summarized by one Los Angeles temperature and wind speed because fine-scale information matters; Tao ties that supercritical sensitivity to why forecasts fail beyond roughly two weeks.
6. A blowup might require a self-replicating liquid computer
Simply pushing energy downscale as fast as possible works in five or more dimensions, Tao says, but fails in three. Energy occupies many scales simultaneously, becomes too dispersed, and gives viscosity enough leverage to damp the cascade.
His averaged model instead uses “air locks”: energy enters one scale, waits until the previous reservoir has emptied, and only then opens the next gate. Inspired by his electrical-engineer wife, Tao assembled mathematical analogues of capacitors, resistors, clocks, and gates into a “Rube Goldberg type machine.”
The speculative extension is a hydraulic Turing machine whose bits are water configurations and whose collisions implement logic gates. A von Neumann-style liquid robot would construct a smaller dormant copy, transfer all its energy into it, power down, and let scaling symmetry repeat the operation faster and smaller until blowup.
Tao calls this a “pipe dream.” He cannot yet build the fluid logic gates, vortex rings are only candidates, analog errors demand correction, and the larger machine may not shut down cleanly; the point is merely that the program “doesn’t contradict any of the laws of physics.”
7. Conway’s Game of Life shows why simple rules are not simple systems
Conway’s Game of Life gave Tao precedent for computational emergence. Its few local rules generate gliders, glider guns, streams implementing AND and OR gates, Turing machines, and enormous self-replicating structures assembled through what looks like “steampunk calculations.”
Much of that construction was crowdsourced by amateur mathematicians. Once basic logic gates existed, ordinary software-engineering composition could build highly elaborate machines, inspiring Tao to ask whether continuous fluid equations might hide analogous components.
Lex’s emergence framing receives an important qualification: random initial cells do not normally produce glider guns or self-replicators. Complexity appears under carefully engineered conditions, just as a Navier–Stokes singularity—if possible—may require an exceptionally designed initial fluid state.
8. Structure-versus-randomness turns partial patterns into usable dichotomies
Tao’s inverse theorems ask why an object exhibits a pattern. The map taking (n) to the integer part of (n\sqrt2) is almost additive: rounding can alter a sum by one, but a nearby completely structured function explains that approximate law.
The broader dichotomy says an object is either sufficiently random or related to something structured; either conclusion supplies leverage. What remains hard is proving that one explicit object—such as the digits of (\pi)—does not conceal a peculiar pattern, even though almost every random sequence behaves normally.
Szemerédi’s theorem says every positive-density set of integers contains arithmetic progressions of any desired finite length. Structured sets such as the odd numbers contain them transparently; random subsets contain them through fluctuation, giving distinct mechanisms for the same unavoidable pattern.
“Infinity absorbs a lot of sins,” Tao jokes about the infinite-monkey theorem. Finitization restores intuition by asking how many monkeys and how much time are needed; generating a prescribed text takes time exponential in its length, explaining why random typing may yield a four-letter word but not Hamlet.
9. Mathematics studies models while science negotiates with reality
Tao separates three layers: reality, imperfect observations, and mental or mathematical models. Science gathers observations and proposes models; mathematics begins with a model’s axioms and asks what consequences and predictions follow within it.
Most disciplines are conclusion-driven—build a bridge, forecast weather, make money—whereas mathematics also explores “suppose I did this, what would happen?” Neither top-down theory nor bottom-up experiment suffices: anomalies in one tell the other where to search.
Mathematics itself has historically been about “99%” theoretical, but experimentation has deep roots. Gauss used extensive prime tables, partly produced by human computers whose job was arithmetic, to conjecture the prime number theorem before a proof existed.
Computation still meets combinatorial explosion: 1,000 elements have (2^{1000}) subsets, and chess has too many positions for exhaustive enumeration. Chess engines nevertheless explore selectively, overturning conventional opening wisdom; Tao hopes AI will similarly enlarge experimental mathematics without supplying immediate formal explanations.
10. A good theory compresses data until correlation breaks the codec
Tao describes physical theory as data compression. A 10-parameter model explaining 10 observations is overfit and useless; a model with roughly 14 parameters explaining petabytes of astronomical data is powerful because a short specification reproduces far more information than it contains.
Universality makes such compression possible. A gas containing around (10^{23}) particles can often be modeled through temperature, pressure, volume, and perhaps five or six parameters, because macroscopic laws forget almost all microscopic detail.
The central limit theorem explains why Gaussian bell curves repeatedly emerge when many sufficiently independent inputs are averaged. Universality is conditional, however: systematic correlations can generate distributions radically unlike a bell curve.
Tao uses 2008 as the warning specimen. Mortgage-default models treated a large population as sufficiently decorrelated for Gaussian risk management, but a systemic shock pushed many borrowers toward default together; beautiful mathematics did not rescue a model whose independence assumptions failed.
11. Foxes create value by moving tools between mathematical silos
Mathematics repeatedly advances by connecting previously separate subjects. Descartes linked geometry to numbers through coordinates; later fields fused algebra with geometry, and today’s obvious (x,y) representation was once a conceptual unification.
Using the fox-and-hedgehog distinction, Tao identifies primarily as a fox rather than a hedgehog. He likes mathematical “arbitrage”: learn one field’s tricks, transport them into a field thought unrelated, and give its deep specialists tools they would not naturally reach for.
His exploratory method is to reprove an attractive theorem using familiar tools, even when the replacement proof is worse. The exercise reveals what the original proof was doing; hedgehogs contribute fuller history, sharper calculations, and precise knowledge of a technique’s limits, making mixed collaborations stronger.
12. Proofs are crafted objects, not disposable certificates
At Princeton, John Conway’s talk on “extreme proofs” changed Tao’s view of mathematical work. Conway imagined all proofs of a theorem occupying a space with axes such as length, elegance, and elementarity, then searched for boundary points: the shortest, simplest, or most unusual proof.
Undergraduate homework rewards any correct argument, but influential mathematics must also be motivated, readable, adaptable, and “a pleasure to read.” Tao compares a technically valid but unusable proof to spaghetti code that performs one task yet invites bugs and resists extension.
Lex’s code-golf analogy preserves Conway’s deeper point: optimizing for an artificial extreme can look frivolous, yet stress-testing proofs, notation, or programming languages exposes ideas ordinary solution-seeking would miss.
For Tao, Euler’s (e^{i\pi}=-1) is beautiful because it connects mechanisms, not merely famous symbols. Exponentiation models growth and contraction; multiplying the exponent by (i) turns motion into a right-angle change, so evolving for time (\pi) produces a half-rotation.
13. The right organizing object can unify apparently incompatible physics
The transcript’s later physics discussion treats notational collisions as possible confirmation that concepts have been chosen well. Early mechanics centered directly measurable mass, acceleration, and force in (F=ma), while energy and momentum emerged later as conserved quantities that deserved more fundamental status.
Hamiltonian mechanics elevated energy into the object governing a classical system’s full dynamics. Quantum mechanics looks completely different—waves rather than classical particles—but its Hamiltonian operator likewise determines evolution through the Schrödinger equation, allowing intuition to transfer between theories.
Noether’s theorem expresses the shared structure: spatial translation symmetry yields momentum conservation, rotational symmetry yields angular momentum, and time translation yields energy conservation. In both classical and quantum settings, symmetries of the Hamiltonian generate conservation laws.
Tao believes quantum mechanics and general relativity should ultimately unify, as electricity and magnetism once did. The blockage is conceptual: Cartesian space-time coordinates are unlikely to survive quantum fluctuations, yet physics lacks the replacement “analog Hamiltonian” that organizes the combined theory.
14. Physics’ success deprives unification of the anomalies it needs
General relativity and quantum mechanics together cover “99.9%” of accessible observations, in Tao’s rough formulation. Evidence distinguishing a unified theory appears only at extreme accelerator energies, in the early universe, or in other regimes difficult to observe.
His confidence comes from history rather than a proposed mechanism: Newton unified terrestrial and celestial motion, Maxwell unified electricity and magnetism, and Einstein found that Riemannian geometry already supplied mathematics for curved space-time.
Tao says string theory was the leading candidate for decades but is “slowly falling out of fashion” because it has not matched experiment. That assessment remains tentative; his larger point is that a beautiful model must rejoin the observational layer.
Analogy lets finite human intuition travel beyond familiar scales. A basketball, golf ball, and light source can reproduce eclipses and lunar phases; reconstructing how ancient Greeks estimated astronomical distances provides “intellectual travel” from flat-looking local experience to a round Earth moving through space.
15. Wave maps yielded when Tao changed the observer’s coordinates
Tao’s work on wave maps, or a sigma model, concerned fields living on space-time rather than gravity itself. He pictures arrows constrained to a sphere and propagating like wheat bending across a field, then asks whether their energy can concentrate into a singularity.
The equation was critical, with comparable behavior at every scale. Around 2000, Tao “barely” proved global regularity by showing that energy must disperse slightly; once dispersion began, the solution could not concentrate enough to blow up.
A curvature-driven nonlinear term initially appeared larger than the stabilizing linear behavior. Tao devised a gauge transformation resembling tiny cameras that move with much of the flow, making the scene look more stationary and exposing a more linear equation.
His discovery process was physical: unable to compute or manipulate the fields effectively, he lay on his aunt’s floor in Australia, closed his eyes, imagined being the vector field, and rolled around seeking better coordinates. Her interruption received the only practical answer: “It’s complicated.”
16. Strategic cheating isolates one difficulty before recombining ten
Tao’s core problem-solving method is “cheating strategically.” If 10 features make a problem hard, he installs nine conceptual cheat codes—set the dimension to one, remove an error term, impose a spherical-cow assumption—and solves the surviving difficulty in isolation.
After treating each obstruction separately, he turns several back on and learns how their mechanisms interact. This differs from “Iron Man mode,” where one attacks the maximally difficult formulation without first identifying which feature causes which failure.
His Hong Kong action-film analogy is exact: a hero defeats 100 attackers because choreography presents them one at a time. If the villains swarmed intelligently they would win, but sequentializing the obstacles makes both cinema and mathematical learning possible.
Tao still works largely with pen, paper, four large blackboards, drawings, and private doodles. AI has nevertheless cut a moderately complicated plotting task from perhaps two hours of recalling and debugging Python to “10, 15 minutes,” so computers increasingly support his exploratory stage.
17. Lean trades convenience for machine-checkable certainty
Lean resembles an ordinary programming language but can output a proof certificate alongside an answer. Each command composes earlier certified statements, and a deliberately small kernel checks the result; several compilers are available for Lean.
Writing Lean feels like explaining a proof to an “extremely pedantic colleague.” Every object needs a type, implicit edge cases become explicit, and a statement that seems obvious on paper may trigger questions about whether a variable is real, natural, functional, or possibly zero.
Most type inference uses “good old-fashioned AI” and tree matching, not large language models. LLMs sit above Lean to search mathlib’s tens of thousands of results or suggest tactics without compromising the underlying deterministic verification.
Tao’s Copilot estimate is deliberately unglamorous: a suggestion works exactly perhaps 25% of the time, is repairably close another 10–15%, and is “complete rubbish” about half the time. Overall, he estimates formalization still costs around 10 times the effort of an informal proof.
18. Formal code makes mathematical refactoring dramatically safer
Tao’s sharpest specimen is a theorem whose final constant was 12. When later work improved 12 to 11, changing the headline left about 90% of thousands of lines compiling and highlighted only the dependencies that genuinely required repair.
On paper, every line might secretly rely on a special property of 12, requiring laborious rereading. In Lean, well-structured abstractions localize the damage, and the repaired formal proof was ready within a day or two.
Formal proofs are longer but locally easier to inspect. Hovering over an object reveals its type, origin, and dependencies, whereas opening page 27 of a conventional paper may require reconstructing definitions scattered across the previous 10 pages.
That local context enables “trustless mathematics.” Tao can send three failing lines to a distant collaborator he has never met, accept a correction from someone whose reputation he cannot evaluate, and still receive a certificate that the resulting argument is valid.
19. Collaboration begins as improvisation before becoming a supply chain
Difficult research cannot start with clean divide-and-conquer because nobody yet knows the viable route. Tao describes an opening “jam session” where collaborators grant themselves unlimited budgets, remove hostile cases, and search for a skeleton that can later be made realistic.
In the Green–Tao theorem, Ben Green possessed number-theoretic control for three-term progressions while Tao had tools influenced by ergodic theory for longer patterns. Tao requested a randomness hypothesis Green could not prove; Green offered a weaker one Tao could not use, and iteration found the property satisfying both constraints.
Once a human proof exists, a formal blueprint can decompose it into a graph of self-contained lemmas with explicit dependencies. Contributors need not understand the whole theorem, much as a specialist in an iPhone supply chain can turn incoming widgets into one larger component.
Conventional experimental mathematics often produces bespoke Python written by nonprofessional programmers; one buggy module poisons the calculation and deters collaboration. Lean’s compatibility and verification let experimentation scale from datasets to proofs themselves.
20. Twenty-two million algebra problems tested industrialized proof
Tao’s Equational Theories Project generated roughly 4,000 candidate laws for a binary operation and examined about 22 million implication questions. Does one identity force another? A yes needs a formal derivation; a no needs a counterexample operation.
Most questions were suitable for an undergraduate algebra student, while roughly 100 were genuinely hard. At the time of the conversation, all but two had been settled; a pen-and-paper proof for one of the remaining cases existed and was being formalized.
Around 50 people participated—an enormous author list by mathematical standards. Lean made that scale credible because no human had to inspect 22 million arguments individually, and every accepted fragment could be checked against the same kernel.
GitHub automatically records activity, but Tao distrusts crude leaderboards: once a metric becomes an incentive, Goodhart’s law says it will be gamed. The project instead uses self-reported contribution categories, while listing everyone as an author and supplying a matrix describing coding, validation, resources, concepts, and other work.
21. Crowdsourcing fails when credit collapses onto one famous name
Earlier Polymath projects required human moderators to validate every contribution, creating a bottleneck formal proof can remove. Their papers sometimes appeared under the collective pseudonym D. H. J. Polymath, modeled on the Bourbaki tradition.
The pseudonym protected equality but harmed junior contributors who could not use the work for tenure because they lacked formal authorship. Public retellings also collapsed the collective into “Tim Gowers’s project” or “Terence Tao’s project,” erasing less famous participants.
The new experiment keeps alphabetical mathematical authorship while documenting roles. Tao values the discipline’s tradition of equal author status, but concedes that equality without contribution metadata stops scaling once collaborations reach dozens or potentially thousands of people.
22. AlphaProof’s medal exposed both capability and brutal scaling costs
Long proofs amplify small error rates: if each of 20 steps has a 10% chance of taking a wrong direction, reaching the end becomes unlikely. Complexity then grows combinatorially as the system must explore and reject branches.
Translation is itself unsolved. Natural language tolerates grammatical gaps and implied context, while one malformed symbol can invalidate a formal statement; even converting among Lean, Coq, Isabelle, and other formal languages remains difficult.
DeepMind’s AlphaProof achieved an IMO silver-medal-equivalent score, but humans first helped formalize the problems, the system exceeded the human time allowance, and Tao cites roughly three days of Google server time for one high-school problem. “This is not a scalable prospect” for graduate research as currently implemented.
Numerical-answer competitions are easier because reinforcement learning receives an immediate right-or-wrong signal. Tao supports a future AI Mathematical Olympiad where systems receive the same problems and time as humans and submit natural-language solutions to the same judges, but says performance was not yet ready for the next IMO.
23. Mathematical smell matters more than fluent-looking proof text
Weak human proofs usually announce themselves through elementary mistakes and “code smell.” AI-generated mathematics can look superficially flawless because training rewards resemblance to correct exposition; the decisive error is often subtle in placement yet embarrassingly stupid once found.
Humans also sense whether a proposed reduction improves a problem. Random transformations usually create two subproblems harder than the original; a skilled mathematician smells when both are simpler, plausible, and worth pursuing.
AlphaZero effectively learned that evaluative sense for chess and Go positions without articulating it. Tao thinks mathematical AI becomes genuinely competitive when it can similarly judge the viability of a proof strategy, not merely generate locally plausible next lines.
Asked for an oracle, Tao wants verification, proof generation, computation, and new representations. Present conversations instead feel like “herding cats”: he repeatedly railroads a model toward a proof he already knows, checks its seductive errors, and spends more energy than solving the problem unaided.
24. Falling below the 1× formalization threshold would reorder publishing
Tooling is already moving Tao’s estimated overhead from 10 toward nine, eight, or seven times informal work. Those increments feel modest, but “one day it will drop below one,” making it rational to develop a theorem formally before turning it into prose.
Journals could then expedite review: the referee evaluates significance, novelty, exposition, and literature while the proof checker certifies correctness. That matters as mathematical papers become longer and qualified reviewers become harder to recruit.
Tao compares the transition with LaTeX. Mathematicians once used word processors, typewriters, and other tools; once LaTeX became easier than its alternatives, adoption crossed a threshold and swept the discipline within a few years.
The feedback loop is potentially virtuous: easier formalization grows mathlib, a larger library makes subsequent formalization easier, AI gains more reliable training and search material, and progressively more proofs become composable infrastructure.
25. AI may join research soon, but discovery lacks its negative training data
Tao’s published forecast was that by 2026 research-level collaborations would include AI involvement. He says versions already existed: a result may depend on an AI suggesting computations or candidate moves, even when authorship credit cannot be cleanly separated.
Current systems can reproduce a “non-trivial percentage,” perhaps 30–40%, of the skills used in mathematics. They reduce friction in Python, routine calculation, and verification, but do not independently supply the full chain of taste, strategy, checking, and exposition.
Literature review is a near-term opportunity with a poor current signal-to-noise ratio: six suggested papers might include one relevant source, one real but irrelevant source, and four hallucinations. It works best when the mathematician already half-remembers the literature and can recognize the valid result.
The missing dataset is mathematical “negative space.” Publications record successful conjectures and polished proofs, not the promising false starts, embarrassing errors, and advisor corrections that teach judgment; Tao jokes AI may need to attend graduate school, submit assignments, visit office hours, and learn from failure.
26. Perelman made Ricci flow critical enough to classify its failures
The Poincaré conjecture asks whether every bounded, simply connected three-dimensional space is topologically a three-dimensional sphere. On an ordinary sphere every loop contracts to a point; on a torus, a loop around the hole cannot.
Richard Hamilton’s Ricci flow smooths curvature like inflating a crumpled balloon. In two dimensions the flow rounds a simply connected surface, but in three dimensions singularities may form—neck pinches, knotted concentrations, or other failures requiring classification and surgical repair.
The three-dimensional problem behaved supercritically: curvature could concentrate on progressively finer scales. Perelman introduced reduced volume and entropy, new scale-invariant quantities analogous to energy, converting the central analysis into a critical problem whose nonlinearities became less threatening.
He then classified the possible singularities and showed how surgery could continue the flow, completing a chain of “really ambitious steps.” Tao contrasts that judgment with current LLMs: a model might list the right idea among 100 suggestions, but 99 dead ends could each consume months before being rejected.
27. False dawns sometimes sustain the endurance that real proofs require
Tao’s fox response to blockage is to switch problems or temporarily assume the bad case away. If multiple failures remain, abandon the route; if everything works except one obstruction, forward reconnaissance can justify fighting that obstruction longer.
In one collaboration, the team believed after two months that it had solved a hard PDE problem. While writing, a co-author noticed an expansion with 13 terms: notes controlled 12, while the omitted 13th was worse than all the others combined.
Months of attempted repairs failed, but the sunk intellectual investment pushed the collaborators toward increasingly unconventional ideas. After roughly two years they found a substantially different approach that avoided the bad term and solved the problem; without the initial “false dawn,” they likely would have quit by month two.
Tao warns against “black holes,” famous problems that latch onto researchers for years while careers deteriorate around the promised eventual triumph. Perelman’s seven-year period of mostly working alone worked spectacularly, but Tao does not recommend that emotional concentration without exceptional fortitude.
28. Prime patterns survive when they are structurally indestructible
Natural numbers are easy to generate additively—start at one and repeatedly add one—or multiplicatively—multiply primes. Problems become extraordinarily rich when addition and multiplication interact, as in asking whether shifting a prime by two produces another prime.
The Green–Tao theorem proves that primes contain arithmetic progressions of every finite length. Tao’s deeper explanation is robustness: progressions appear in structured sets for explicit reasons and in random sets through fluctuation, so the structure/randomness dichotomy wins either way.
They are “like cockroaches”: one may remove 99% of the primes and still retain arbitrarily long progressions under the theorem’s appropriate density formulation. Twin primes are fragile by comparison; carefully deleting perhaps 0.1% of primes could destroy every twin while leaving aggregate statistics looking authentic.
That fragility says any twin-prime proof must exploit a delicate property of the actual primes that edited pseudo-primes lack. Prime randomness is not merely aesthetic: the conjecture tests whether mathematicians can prove with zero error probability what statistical models predict overwhelmingly.
29. The parity barrier blocks twin primes even after the gap fell to 246
Tao states the current bounded-gap result as infinitely many prime pairs differing by at most 246. It does not identify which gap repeats: twin primes differ by two, cousin primes by four, and the less consequentially named “sexy primes” by six.
The mechanism resembles the pigeonhole principle after replacing primes with better-understood almost-primes. Within a carefully weighted almost-prime set, true primes may achieve enough relative density that some bounded pair must occur.
Twin primes require pushing that density to at least 50%, but sieve methods cannot cross the “parity barrier.” Lex compares the obstruction to passing the speed of light; breaking it could unlock the twin-prime and Goldbach conjectures along with multiple neighboring questions.
His forecast is measured: in 10 years he expects “many more much closer results,” though perhaps not the full conjecture. When an attempted proof reaches Beijing from New York without visibly crossing an ocean, his mathematical smell says the route was too easy and a hidden mistake remains.
30. Riemann may need an accident, while P versus NP accumulates no-go results
The Riemann hypothesis formalizes square-root cancellation: multiplicative statistics of the primes should fluctuate as little as genuinely random data. Sampling more voters reduces error like the square root of sample size; Riemann demands an analogous, nearly optimal randomness from primes.
Existing techniques leave too much error and survive modifications of the primes that destroy the hypothesis. The true proof must use an exceptionally delicate feature and “come out of left field”; Tao sees no serious current proposal and says it may have to happen “by accident.”
A disproof would shock number theory and cast doubt on cryptographic intuition. Encryption aims to turn meaningful text into output indistinguishable from random noise; if primes contain a major unsuspected pattern, other supposedly random number-theoretic constructions deserve re-examination.
Tao calls P versus NP the problem with the broadest potential ripple effects. Evidence leans “slightly” toward (P\ne NP), but computer science has also proved unusually many obstruction and no-go theorems against proposed approaches; he leaves open even the possibility that the statement could be undecidable.
31. Collatz shows why proving 99% can leave the entire mystery intact
Collatz repeatedly divides an even integer by two or maps an odd integer to (3n+1). Starting from 13 gives 40, 20, 10, 5, 16, 8, 4, 2, 1, after which the cycle (1,4,2,1) repeats.
Typical trajectories resemble Brownian motion or a stock chart with downward drift—like repeatedly gambling at slightly unfavorable odds. Tao proved, roughly speaking, that around 99% of inputs eventually fall far below where they began, though not necessarily all the way to one.
Probability cannot remove the exceptional trajectory that keeps winning. One number might encode a “heavier-than-air flying machine,” a self-sustaining computational structure that grows forever even while most inputs descend.
Conway showed that richer Collatz-like iterations can encode Turing machines through his FRACTRAN language, making generalized versions undecidable. The result does not settle ordinary Collatz, but explains why a simple rule may inherit cellular automata’s computational depth and why “100% of all inputs” is categorically harder than a statistical theorem.
32. Recognition simplifies collective work but can distort what gets valued
Perelman declined both the Fields Medal and the million-dollar Millennium Prize, saying correctness required no further recognition. Tao has never met him and avoids diagnosing his withdrawal, describing him only as an outlier who became disillusioned and chose not to engage.
Winning the Fields Medal made Tao “part of the establishment”: people suddenly requested opinions, and casual remarks carried new weight. The medal did not solve any research problem, but seniority brought a social contract of mentoring, administration, outreach, and shaping the field after years spent “in the trenches.”
Tao accepts famous individuals as first approximations—Steve Jobs for Apple, or the final solver of a theorem—but stresses that the last step often rests on decades or centuries of invisible work. Human minds comprehend stories through a few representatives, yet that shorthand can erase whole teams.
Andrew Wiles represents a style opposite Tao’s: years of secret, concentrated work on Fermat’s Last Theorem. Kevin Buzzard’s five-year grant aims to formalize the proof back to results known by 1980, exposing the deep tower of algebraic objects beneath a headline normally attached to one person.
33. Mathematics needs more than one native language of thought
Tao argues that evolution supplied no dedicated mathematics center. People repurpose vision, language, symbolic reasoning, gaming, or puzzle-solving systems, producing mathematicians who reach the same conclusions through genuinely different internal routes.
Mass education cannot easily teach 30 students in 30 styles, so many learners never discover their “native math language” before a poor classroom experience drives them away. Tao recommends alternative entry points: YouTube, puzzles, popular books, poker probability, chess, baseball statistics, and other communities where mathematics serves a concrete interest.
Formalization may extend citizen science into mathematics. High-school students may already be contributing to some formalization projects, and programmers can find an entry point through Lean without understanding an entire research program, while the kernel relieves professional mathematicians of manually validating every public submission.
His career advice emphasizes transferable capacity over one language or narrow technique: abstract reasoning, adaptation, and recovering when a plan fails. Tao modeled that shift by learning Lean after realizing he could not merely predict an AI–formal-proof synthesis from the authority of a Fields Medal; he had to “walk the walk.”
34. Collective intelligence keeps impossible tasks moving toward homework
Hilbert’s 23 problems demonstrate the power of declaring what deserves attack: without a target, bystander paralysis prevails. Tao’s undergraduate remedy is to “try anything,” preferably something obviously flawed, because the precise way it fails reveals which unused hypothesis matters.
Psychology is part of the machinery. Strategic cheating makes a problem feel feasible; “structured procrastination” gets an unwanted task done by placing an even worse task behind it. Marathoners and mathematicians alike need techniques that preserve motivation, not just technical preparation.
The isolated unaided human is already a fiction: language, pen and paper, blackboards, software, and institutions are cognitive technologies. The mathematical community is a “super-intelligent entity” beyond any member, visible when MathOverflow rapidly assembles answers from specialists with complementary knowledge.
Tao’s closing source of hope is historical compression of difficulty. Navigation once cost lives and fortunes; a pocket device now solves it automatically. Healthy infrastructure lets younger generations turn today’s impossible research into tomorrow’s homework, even though people rapidly normalize each advance—from voice-capable AI to robotics—the moment it arrives.