What this covers

This interview with mathematician Terence Tao starts from the publisher's question of whether AI has ended the era of the individual math problem. It then ranges well beyond it. Tao describes a move from proof scarcity to proof abundance: the seven Millennium Prize Problems once marked hard-won landmarks, while machines can now brute-force many older ones. He also argues that AI-generated papers lose the talks, collaborations, and trained students that came with a human paper.

The middle of the conversation covers foundations and history. Tao discusses Kurt Gödel's incompleteness result and what it meant for David Hilbert's program, mathematics as a "multiverse" of alternative rule systems, and the slow arrival of zero, negatives, and infinite limits. He explains that the Riemann hypothesis quantifies the apparent randomness of the primes. He also says statistical evidence cannot rule out one exceptional trajectory, so a proof is still needed. On the applied side, he credits information theory with showing when to stop improving a system, then turns to economics. He covers Black-Scholes pricing, its dependence on Gaussian fluctuations, and the 2009 crisis, along with the role quantitative finance plays for mathematicians academia cannot absorb.

The more contested turns are philosophical. Tao says Newcomb's paradox and the repeated prisoner's dilemma show that precommitment can beat flexible rationality. He calls Nick Bostrom's simulation argument a good attempt that still needs humility about unimagined models. He also criticizes string theory as too flexible, favors gradual technological transitions, and doubts that small classroom experiments survive scaling to many schools.

Sharpest takeaway

Tao argues that AI is changing mathematics from a field where proofs were scarce and understanding was the goal into one where proofs are abundant and understanding must be deliberately recovered, which requires new scales of collaboration, communication, and exposition.

  • Proofs have decoupled from understanding: AI can produce answers, such as a new Riemann hypothesis idea, that humans do not yet understand
  • Individual problems and papers lose their role as the unit of progress, so math must move to large datasets, networked literature, and human-to-human exposition
  • Mathematical models and axioms are tools with limits, so a humble, pluralistic, model-aware approach is needed across math, economics, and finance

The argument · threads12 threads · 60 claims
0.71

AI-generated answers decouple proof from understanding, reshaping research incentives and the human role.

6 pointscentrality 5/5
  • Historically the only way to make progress on a hard problem was to understand it and communicate that understanding, but AI machines now give answers while humans must work much harder to extract useful insight from them.

    “Now we have these machines that can give us answers, but we have to really work much harder than normal to extract out the useful insights from them.”

  • Two days before the interview, a significant advance on the Riemann hypothesis was made, but the new ideas were generated by AI and mathematicians do not yet understand them, which Tao describes as getting answers without understanding.

    “Unfortunately, the ideas were generated by AI, and we don’t yet understand them.”

  • The shift from human-generated to AI-generated proofs is changing the philosophy of mathematics, because the notions of proving something and understanding something, once almost synonymous, have been decoupled quite severely.

    “As I said, these two statements used to be almost synonymous, and now they have been decoupled quite severely.”

  • Papers will evolve to show more of the process, as behind-the-scenes content enriches movies, and the literature will become networked databases of thousands of papers that can be queried and updated in real time, with collaborative projects combining AIs and humans.

    “the literature itself would be a much more networked, living, collaborative entity where both AIs and humans can collaborate. I think this is the only way forward.”

  • Mathematics is a playground where failure is cheap, unlike bridge building or surgery, and the most valuable skill it teaches is how to set up an unknown problem, start, and avoid paralysis, which matters more than specific tools that can be looked up or delegated to AI.

    “I view math as a good playground to think where the cost of failure is cheap.”

  • Researchers racing to post papers minutes before others who are using AI to generate nearly identical results, while cutting corners and making papers harder to read, shows that the priority metric has become over-optimized and counterproductive.

    “People are just racing to dump out a paper three minutes before someone else with an AI is generating almost exactly the same paper”

0.57

Models and social-science evidence demand scrutiny for hidden assumptions, testing, and humility rather than trust.

11 pointscentrality 3/5
  • String theory has strongly influenced geometry, but it has no striking experimental confirmation, and one major objection is that it is too flexible, explaining trillions of alternate universes without a reason we live in this one.

    “One of the objections that I’ve heard is that it’s actually too flexible a theory.”

  • Quantum mechanics requires accepting that objects live in superpositions most of the time, such as a cat being a mixture of alive and dead, which is unintuitive but mathematically works and explains the world better at the molecular level.

    “you have to accept that, in fact, most of the time, objects live in superpositions”

  • When quantities become very large, it is often easier to send the number to infinity and study the limit, as with modeling fluids as continuous objects with pressure and volume rather than trillions of particles.

    “Often, the mathematics becomes simpler.”

  • Some models of the universe can be refuted by pure reasoning, for example prophecy and fate are incompatible with free will and predictable natural laws, and similarly, a simulation may conflict with certain theological or creator-based ideas.

    “There are certain models for the universe that you can refute by pure reasoning.”

  • Black-Scholes option pricing works because of an algebraic miracle that holds when price fluctuations are Gaussian, so pricing depends on knowing risk; without Gaussian fluctuations the model is dangerous, especially when compounded into derivatives, which Tao links to the 2009 crisis.

    “If you don’t have Gaussian fluctuations, then actually Black-Scholes is somewhat dangerous to use, especially if you compound it into derivatives over and over again.”

  • Small enthusiastic classroom experiments, such as flipped classrooms or tech bans, often yield great results but fail when deployed to 100 schools, because the prototype-to-scale transition and typical school realities are unsolved problems.

    “Then you try to deploy them to 100 schools, and suddenly it doesn’t work.”

  • Economics is a mixed case where models sometimes predict reality well and sometimes fail, unlike physics where accurate models can be explored freely, so economists must understand when models break down and how to adjust them.

    “Economics is a much more mixed situation where you have models, and sometimes they predict the real world very well, but sometimes they don’t.”

  • Bayesian reasoning is subtle because one may weigh only a hypothesis and a null hypothesis while the true explanation is a third model one has not imagined, so humility about one's own imagination is required; this applies to Bostrom's simulation argument, which Tao calls a good attempt.

    “there could be a third hypothesis that you’re not considering at all, which is actually the true explanation.”

  • Tao holds his philosophy of mathematics with humility, keeping an open mind and updating views on new events and data rather than adopting a dogmatic view of what mathematics or science is.

    “You don't have a dogmatic view of what science is or isn't. You update it based on events and new data”

  • Economics is somewhere between the hard and social sciences because it has enough mathematical models to apply math, yet real-world messiness and subjectivity must be grappled with too.

    “It's somewhere between the hard sciences and the social sciences”

  • Quantitative studies of which pedagogies work are a very soft science, and some of their findings are controversial, so educational policy should rest on many experiments rather than single studies.

    “but it is a very soft science. I think some of the findings are controversial.”

0.55

Research value lies in human-communicated process and collaboration, not just finished papers.

5 pointscentrality 4/5
  • Exposition has been undervalued in mathematics, and the bottleneck is now what results can be explained human to human, because papers are a poor communication medium and talks with questions, body language, and visual aids will increasingly become primary research outputs.

    “Papers are actually a very terrible way to communicate.”

  • AI-generated papers separate the PDF from the social and intangible value a human paper carried, such as the author giving talks, meeting collaborators, training students, and discovering new questions, and AI typically does not communicate the process behind its results.

    “They have unbundled the PDF, the actual physical paper, from the penumbra of more social or intangible values of that paper.”

  • Trusting junior collaborators with ambitious tasks often yields results that exceed expectations, and Tao's blog moved from about 99 percent of articles written by him to about 95 percent written by community members who offered viewpoints he would not have had.

    “I have a blog where, for almost the entire history of the blog, I wrote like 99 percent of the articles. Recently, I’ve turned over my blog to the community, and now 95 percent of the articles are written by other people”

  • Papers should move toward showing the process of discovery, much as films gain from behind-the-scenes documentaries, since the finished product alone has been overemphasized in the past.

    “We have focused a bit too much in the past on just showing the finished product”

  • Delegating ambitious tasks to bright, motivated junior people, and giving AI a role in complex work, lets junior people take the lead and exceed expectations, a shift from doing everything oneself that AI makes possible.

    “if you give someone an ambitious task, and they’re bright and they’re motivated, they often can exceed your expectations”

0.52

Mathematics is shifting from scarce landmark problems to abundant, industrialized, AI-scaled problem sets.

4 pointscentrality 4/5
  • Mathematics should shift from individual landmark problems to huge datasets of thousands or millions of problems swept out at scale, where researchers report success rates such as handling 30 percent of problems with AI and 10 percent with a human contribution.

    “we may start creating huge datasets of thousands or millions of problems and try to sweep them out.”

  • Math academia is still pre-industrial and will be forced almost at gunpoint into an industrial or post-industrial mode of mass production, requiring major cultural change, and even Millennium-level individual problems may disappear.

    “We are forced, almost at gunpoint right now, to switch to an industrial or post-industrial economy”

  • Mathematics has moved from an era of proof scarcity, in which hard proofs were valuable landmarks like the seven Millennium Prize Problems, to proof abundance, in which machines can brute-force many old landmark problems and the currency of individual open problems has been hyperinflated.

    “in one sentence, we have moved from an era of proof scarcity to proof abundance.”

  • Instant access to any music through the internet makes Tao listen to it less than when he needed CDs, showing that the more abundant something is, the less people may use it.

    “The more abundant something is, sometimes the less you use it.”

0.52

Rigorous mathematics requires proof over probability, careful handling of infinity, and respects hard limits on provability.

9 pointscentrality 3/5
  • Knowing the mathematical capacity limit tells you how close you already are to perfection; a wireless encoder at 98 percent of theoretical capacity justifies stopping further development rather than spending another $1 million.

    “when to stop trying to develop an even better wireless encoding protocol because you’re already at 98 percent of mathematical capacity”

  • Shannon's information theory from the 1940s became essential to the modern economy only when cellphones and big data arrived decades later, showing that mathematics can take a long time to become useful.

    “Shannon’s information theory came out in the ’40s, but it’s only when cellphones came along and big data that it really became so essential to the modern economy.”

  • Gödel's incompleteness theorem showed that no fixed set of rules of mathematics can settle every statement, so Hilbert's dream of solving all open problems can never be finished, though the undecidable statements Gödel built are rarely ones that arise in practice.

    “It said that Hilbert’s program would never be finished in some sense.”

  • The Riemann hypothesis precisely quantifies the apparently random distribution of prime numbers, and many results in number theory are conditional on it, so their strength increases if it is true and they say nothing otherwise.

    “many, many consequences. Any part of mathematics that touches the primes, if you know the Riemann hypothesis, normally your results get stronger”

  • Mathematics is the soil of the STEM ecosystem: other sciences show what is possible, while mathematics proves limits, such as the maximum bits per second a wireless channel can carry, which cannot be discovered by engineering alone.

    “If you want to show that something’s impossible, that there’s a limit, for example, if you want to communicate using a wireless spectrum”

  • The Collatz conjecture is true in Tao's view, every number eventually reaches 1, and the first trillion numbers tested confirm it, but no current mathematical idea can rule out a rare exceptional trajectory that escapes to infinity or loops, so it may not be proven even with AI.

    “It’s true, but we have no idea how to prove it, even with our fancy AIs.”

  • Greek paradoxes showed early mathematicians could not handle infinity, and it took centuries to make infinity safe to use, since not all algebra laws hold for infinite quantities and it must be handled with care like radioactive material.

    “It took centuries to work out how to deal with infinity safely”

  • Almost-all statistical results and billions of tested cases cannot exclude one special exceptional trajectory, just as a gambler's long run of losses cannot exclude a rare winning streak, so the conjecture needs a proof rather than probability.

    “For Collatz, we can't rule out that one exceptionally lucky hailstone”

  • Real analysis arose because large-scale problems are easier to model in the infinite limit, as with treating fluid particles as a continuum, but infinity does not obey all algebra rules and must be handled with care, which took centuries to achieve.

    “often, the mathematics becomes simpler. It’s like if you want to model a whole bunch of particles of a fluid”

0.33

Gradual transitions are preferable to abrupt disruption because sudden shocks destroy careers.

2 pointscentrality 3/5
  • Gradual transitions are better than abrupt ones because if the Industrial Revolution had happened within one month, entire careers would have been cut short with no chance to transition halfway, so a medium path that keeps benefits and reduces disruption is preferable.

    “If the Industrial Revolution happened within one month, it would have been much more traumatic, entire careers cut short with no ability to transition halfway through.”

  • Tao is excited by possible new mathematics but horrified by the disruption of the AI transition, which he thinks is much larger than necessary, and a gradual transition would preserve benefits while reducing career-destroying shocks.

    “I'm also actually quite horrified by the disruption, which I think was much more than necessary”

0.30

Personal struggle and intuitive framing shape how mathematics is learned, taught, and practiced.

6 pointscentrality 2/5
  • Intimidating analysis statements like 'for every epsilon there exists a delta' become tractable when framed as a game in which an opponent supplies epsilon and you counter with delta, letting you reason about strategies and who is ahead.

    “if you phrase it as a game, your opponent is giving you an ε, you need to counter that ε with a δ of your own.”

  • Ricardo's law of comparative advantage is a useful principle that Tao uses to decide what to spend his time on, because the choice is about which tasks he is relatively best suited to.

    “Ricardo’s law of comparative advantage, I use that to run my life.”

  • Euler's identity e^(iπ) = −1 is most beautiful because it shows that rotating a quantity by π radians is a form of multiplication that flips it to its negative, connecting geometry, dynamics, and algebra.

    “what it’s really saying is that rotation is actually a type of multiplication”

  • Being at a small university where everyone knew each other, rather than in a super-competitive elite environment, let Tao feel on par with peers, while at Princeton he was shocked at how much more math his fellow graduate students knew.

    “I think actually not being in a super competitive environment at the time, but in a small place where everyone knew everyone else, I felt on par with my other students, that worked well for me.”

  • At age 51, Tao is often exhausted and has no choice but to delegate to humans and AI, accepting that others can do some things better, while still finding meaning in choosing daily work.

    “This is another reason why I’m learning to delegate. I have no choice.”

  • Tao failed his oral qualifying exam at Princeton because he could not convince his adviser he knew the subject, and this failure changed his life by making him study harder and put real effort into explaining his work.

    “I actually learned I was terrible at oral exams. I could not convince my adviser that I knew the subject I was studying”

0.25

Mathematics is pluralistic and best appreciated through its history, stories, and lost craft practices.

4 pointscentrality 2/5
  • Mathematics has become pluralistic: there is a standard foundation but also many exotic alternative systems with more or fewer rules, analogous to chess variants, forming a multiverse of mathematics to explore.

    “We’ve moved to a more pluralistic view of mathematics, where there isn’t just one true mathematics”

  • Zero was a late conceptual leap because absence is invisible; humans only named what they could see, and later discovered zero, negatives, fractions, and square roots by following the rules of arithmetic operations and trusting them to predict new numbers.

    “Just by following the rules of what these operations were trying to do, they were naturally led to consider new numbers”

  • Mathematics is more about stories than normally presented, and classes strip out historical backstory due to time pressure, so reading history, such as Galois writing group theory the night before his fatal duel, makes the subject come alive.

    “I think mathematics is about stories a lot more than is normally presented.”

  • Preserving obsolete craft skills has cultural value and strategic value because after a disaster such as losing the power grid they may become relevant again, and history of obsolete practices can inform modern skill development.

    “from a strategic point of view, there’s actually some value in preserving what you might think of as obsolete skills.”

0.25

Predictable, committed strategies can outperform free, rational deliberation in strategic interactions.

4 pointscentrality 2/5
  • Newcomb's paradox shows that more free will and rational deliberation can hurt you, because a predictor rewards those who precommit by removing options, such as lobotomizing oneself to take only one box, so being too smart can be penalized.

    “Newcomb’s paradox is actually an environment where you are penalized for thinking too much.”

  • Interacting with a highly rational or predictable agent can be safer than with an intelligent but unpredictable one, and in the game of chicken, visibly disabling one's steering wheel can force the rational opponent to swerve.

    “I think the game of chicken is another game where one of the strategies to win chicken is to actually disable your steering wheel”

  • In the repeated prisoner's dilemma, publicly committing to a simple automated strategy, such as cooperating unless the opponent defected, removes free choice and signals trust, which can produce better outcomes than flexible rational play.

    “If you signal in advance what you do and you remove your own free will, then you can actually do better in the game because people will see that you’re trusted.”

  • Many women reportedly said they would feel safer with a bear than a man, which Tao attributes to the bear being more predictable, so interacting with a highly intelligent, unpredictable agent can add risk.

    “Many women said they feel safer with a bear”

0.22

Math readiness and curriculum should shift toward statistics, with enrichment kept optional.

3 pointscentrality 2/5
  • A UC San Diego report raised concern about the basic math readiness of incoming students, with possible causes including selection system changes, the debate over reinstating the SAT, COVID, and social media's effect on attention.

    “There was a report from UC San Diego in particular that was quite concerning about the basic readiness of many of the students.”

  • The college math syllabus has traditionally emphasized calculus for physics and engineering, but in the big-data era statistics, machine learning, linear algebra, and probability are becoming more important, so curriculum tweaks toward them are warranted.

    “In the era of big data, statistics and now machine learning and linear algebra and probability are becoming more important.”

  • Advanced math enrichment such as AP exams, clubs, and competitions should be optional because forcing a child into them causes hatred of mathematics, and they scale only with qualified teachers, supportive administration, and parental choice among school options.

    “They shouldn’t be mandatory. I think if a kid doesn’t want to do these things and you push them into this, they will just have a horrible time and end up hating mathematics.”

0.22

Quantitative finance absorbs mathematicians academia cannot employ, but optimizing wrong metrics risks crashes.

3 pointscentrality 2/5
  • Mathematicians who succeed in finance can give back to pure mathematics, as Jim Simons did by founding the Simons Foundation after making billions at Renaissance Technologies, which now sponsors pure mathematics, physics, astronomy, and computer science.

    “Jim Simons is one example. He founded Renaissance Technologies.”

  • Quantitative finance supplies needed jobs for mathematicians whom academia cannot absorb and can price things and provide liquidity, but optimizing the wrong metrics can destabilize economies and create crashes.

    “If you pursue the wrong metrics too much, you can also destabilize economies and create crashes.”

  • Academia is a fairly small sector of the economy and trains many more mathematics students than can obtain academic jobs, so other industries like finance must value math PhDs, and quant finance often pays better than academia.

    “Academia is a fairly small sector of the economy. We train many more students than can end up having academic jobs”

0.08

Tao's personal background and tastes are incidental anecdotes about schooling, geometry, and film.

3 pointscentrality 1/5
  • Tao's parents designed a track that let him skip five grades in math and science, and he enrolled at Flinders University around age 13 while living at home, taking university math while still in high school.

    “In fact, in math and science, I skipped five grades.”

  • Tao finds geometry beautiful partly because it takes effort for him, since he cannot rotate shapes mentally and must write things out and use visual aids, while his native field is analysis about what is big or small and when waves interfere.

    “Maybe because of that effort, the ones that I do understand, I actually find quite beautiful”

  • Tao's favorite Hong Kong action film is Kung Fu Hustle, a parody that works best in the original Cantonese but holds up in English.

    “My favorite is called Kung Fu Hustle by Stephen Chow.”