AI Crisis in Math: Navier-Stokes, Riemann Hypothesis & Complexity Theory | Lex Fridman Podcast #503

Summary

• The panel describes a sudden expansion in AI’s mathematical capabilities that has made prior forecasts less reliable. Participants discuss an OpenAI drop of roughly 700 problems and reported proofs including unique games and quasi-Riemann; Kontorovich says the released set contained positive results from thousands of attempted problems, with an apparent success rate of about 10%. Full Riemann and P versus NP remain unresolved in the discussion. Kontorovich’s response is, “I’m readjusting my inability to predict it.”
• Formal verification gives AI a checkable route to mathematical claims, while reusable proof infrastructure remains unfinished work. Kontorovich describes Anthropic formalizing Fermat’s Last Theorem in approximately 13 million lines of Lean, but distinguishes a checked result from a clean, maintainable contribution to Mathlib. Organizing, explaining, and reusing proofs remains substantial work after systems have “jerry-rigged” their way to correctness.
• The claimed Navier–Stokes breakthrough leaves the commercially valuable fluid-design problem largely intact. Brunton describes a carefully constructed, smoothly forced, three-dimensional blowup operating outside ordinary engineering conditions. His distinction is explicit: “We solved the $1 million math problem,” but predicting and controlling practical flows remains “a $1 trillion engineering problem” across transportation, defense, health, and energy.
• AI can change a research field through recombination alone. Kontorovich prefers a “semigroup” of repeatedly composable ideas to the dismissive image of interpolation inside a fixed balloon. Moitra’s concrete example is an Anthropic-associated 3SUM result: a faster algorithm overturns a hardness assumption, changes downstream algorithms, and forces researchers to reconsider which problems belong together.
• Cryptographic disruption is a conditional risk, not a result announced in this conversation. The panel uses efficient factoring as a hypothetical breakthrough with immediate economic consequences. Moitra notes quantum factoring algorithms and the concern about sufficiently large quantum computers; the discussion also mentions lattice-based methods. Even P equaling NP could have practical consequences dependent on the algorithm’s running time, which “could be so large that it’s impractical.”
• Deep learning’s effectiveness still outruns the mathematical explanation offered here. Moitra says the proof release contains no result explaining LLMs, and argues that the structure of natural data is intertwined with their success on problems that are hard in the worst case. His deliberately qualified framing is that “in some sense, nature abhors computationally hard problems.”
• Apparently innocuous training data can transmit behavior, making data lineage and deployment controls consequential. Moitra describes an owl-preferring model generating numbers that, when used to fine-tune another model, transmit the owl preference. His low-rank modeling work offers a possible mechanism through interacting concept directions; Brunton’s deployment filter is to distinguish how verifiable an output is from how costly an error would be.
• The immediate bottleneck is institutional adaptation and human motivation, not a shortage of new theorem statements. Students face disrupted projects, reviewers face more output, and mentors struggle to reconcile multiyear research with students’ three- and six-month horizons. Moitra’s “doorman fallacy” captures the labor question: automating the most visible task does not establish that the entire role has been replaced.

Deep dive

1. The proof release breaks the old research timetable

• Fridman convenes the panel around claims that machines have solved major mathematical problems within weeks. Moitra says rumors had already circulated about unique games, RL equals L, and matrix multiplication; his first response was relief that the results were available for examination. Kontorovich describes the broader transition as moving from a candle in the woods to a flashlight.

• The emotional cost falls unevenly. Fridman quotes Terence Tao mourning “many person-years of ongoing patient and deliberately slow research efforts,” particularly students’ work disrupted by the release. Moitra feels newly empowered to learn across fields, but “more intellectually lonely”: the familiar mentorship blueprint has broken, and humble questions now compete for attention with a flood of spectacular results.

• Brunton’s engineering perspective is mostly hopeful: better tools could extend work on aircraft, artificial hearts, and other practical systems. His uncertainty concerns the next five or 10 years—humans might be dwarfed by another intelligence, or capabilities might level off into an extraordinary tool. Those futures demand very different adaptations, and he does not pretend to know which arrives.

2. Solving mathematics would not necessarily destroy its human value

• Kontorovich separates “solve math” from “destroy math” through a hypothetical “solved axiom”: any problem solvable by a human interacting with AI could also have been solved by AI alone. This does not mean every mathematical problem becomes solvable. It means human involvement adds no capability beyond what the machine could independently contribute.

• His warning comes from centaur chess. Human-computer teams initially looked like a durable collaboration, until engines became strong enough that rejecting their preferred move became a disadvantage. The human’s contribution shrank to “Yes, accept the first move.” Kontorovich asks whether mathematical collaboration could eventually encounter the same problem.

• Even under that assumption, he argues, pure mathematics retains a human purpose. Unlike engineering applications, its value is not exhausted by results stored as “ones and zeros in data centers.” What matters is understanding that lives in people’s heads. A theorem can be true without contributing the explanation or experience that makes mathematics valuable to its practitioners.

• Kontorovich repeatedly says he does not actually believe the solved axiom describes the present. Moitra also doubts it, suggesting it might be “half true” in ways still to be discovered. Neither treats question selection as a guaranteed permanent human advantage: a sufficiently capable system could also be asked what questions ought to be asked.

3. The doorman fallacy exposes a broken measure of mathematical work

• Moitra’s analogy is the “doorman fallacy”: reduce a complicated job to its most visible task, then assume automating that task replaces the role. An automatic door opens entrances, but omits recognition, assistance, security, and hospitality. Likewise, proving theorems is mathematics’ most visible output, rather than a complete account of what mathematicians contribute.

• His alternative definition is a person who “thinks deeply.” Proving an old conjecture is one possible byproduct; building theories, making conceptual connections, and changing how a community sees a problem are others. These activities might also become automated, he allows, but he does not think the current results establish that they already have.

• Kontorovich identifies a newly broken correlation. Until recently, proving deep new theorems was a strong proxy for understanding, beauty, and depth of thought. Now someone can request a breakthrough and circulate the output without possessing those qualities. Brunton would move evaluation one step earlier: good problem-solving depends on asking useful, interpretable, answerable questions.

• Moitra insists that this shift still requires mathematical insight. Valuing questions and explanations should not turn mathematicians into “Twitter influencers.” The difficult institutional task is replacing a simple career rubric—improve a recognized problem, publish, become known—with ways to recognize ideas that communities adopt and find powerful over time.

4. AI’s originality need not resemble a human’s originality

• Moitra cautions against explaining impressive performance by anthropomorphizing its mechanism. He recalls Alex Madry’s work on adversarial image features: pictures that humans recognize as cats can acquire subtle patterns that make a network classify them as dogs. Training on deliberately relabeled, perturbed images can nevertheless produce a network that performs well on ordinary images.

• The lesson is that the network can exploit statistical relationships people do not use or articulate. Moitra sees a comparable qualitative difference in the mathematical solutions he has examined: “They just don’t look like a human has solved those problems.” That difference could support collaboration, although he immediately qualifies the prediction: “Could be wrong next week.”

• Kontorovich rejects the dismissive force of calling AI’s work merely the “convex hull” of existing ideas. A balloon suggests a fixed, enclosed region of combinations. Tim Gowers’s “semigroup” framing better captures repeated composition: place another tile beside an existing tile, then continue. Familiar moves can reach an unexpectedly enormous space when chained effectively.

• For quasi-Riemann, Kontorovich sees known ingredients arranged into a startling cascade, rather than an unmistakably new foundational concept. Brunton adds that even illuminating existing connections would be consequential: human knowledge contains substantial fog. Making the balloon transparent—showing researchers what is already reachable—would itself be an extraordinary advance.

5. A faster algorithm can reorganize an entire research field

• Moitra’s clearest example comes from an Anthropic result written up by Vassilevska Williams and Josh Alman on 3SUM. He describes a basic problem familiar from coding interviews receiving an asymptotically faster algorithm than the community expected. Its significance lies partly in contradicting an assumption around which researchers had organized a substantial theory.

• Fine-grained complexity had linked other problems to 3SUM’s presumed difficulty. Making 3SUM easier therefore produces consequences beyond one improved algorithm: it changes which hardness arguments remain useful and whether those linked problems should still be studied together. The result alters the research map even before every conceptual lesson has been extracted from its proof.

• Moitra expects the researcher archetype to change accordingly. Fields that originated from related ideas have become increasingly specialized; he guesses people able to think across a broader swath of mathematics will benefit most from these tools. That is a prediction about useful breadth, accompanied by explicit uncertainty about future hiring, journals, and graduate training.

6. Cryptography illustrates the difference between a theorem and a shock

• Kontorovich offers efficient factoring as a hypothetical release whose consequences would be immediately recognizable. Multiplying two enormous numbers is straightforward compared with recovering their prime factors from the product. He connects that asymmetry to public-key encryption and secure commerce, imagining a proof drop that abruptly undermines assumptions supporting widely used systems.

• Moitra qualifies the sweeping version of that scenario: quantum factoring algorithms already exist, while sufficiently large quantum computers remain a concern. The panel mentions lattice-based methods as an alternative. They are discussing an illustrative risk, not claiming that the release contains a factoring breakthrough.

• P equaling NP would also require examining the actual algorithm, Moitra says. A polynomial running time could still be impractically large, producing a theoretical upheaval without an immediate practical revolution. It is Ankur Moitra who says that such a result “should go through the government.”

7. Mathematical review capacity was strained before AI

• Fridman compares the mathematical publication problem with overwhelmed open-source maintainers. Repositories depend on motivated people reviewing contributions; a flood of submissions can exhaust those people rather than improve the project. Kontorovich says journals already faced the analogous problem before AI: more output, more requests, and referees struggling to perform careful intellectual work.

• His example is Perelman’s Poincaré work. Posting the papers began a lengthy community process of lectures, questions, and detailed examination. Perelman could explain the reasoning behind difficult passages because the whole argument lived in his mind. Establishing communal confidence took years; producing the initial solution was only the first step.

• Brunton sees a need for organizational leadership and larger collaborative structures, familiar from engineering and major scientific projects. A contribution may be reusable code, data, or a component supporting a larger undertaking rather than a standalone 15-page paper. He recognizes the cultural friction: mathematicians often choose the field partly for independence and solitude.

• Moitra proposes additional publication models rather than simply abolishing existing ones. Public, evolving annotations could capture simplified arguments and successive improvements instead of scattering them across drafts and social-media threads. Mathlib offers Kontorovich’s concrete example: pull requests make review, refactoring, reuse, and individual contributions visible in a shared workspace.

8. Lean checks deductions while Mathlib makes them reusable

• Kontorovich explains Lean as an interactive theorem prover: axioms and inference rules are represented in software, and proposed proof steps are checked against them. It can also check software specifications. Mathlib supplies a maintained mathematical library, sparing each project the task of reconstructing basic objects and results from the beginning.

• Formal specifications solve a practical reuse problem. A lemma in a book may silently depend on the preceding 60 pages; extracting it safely requires recovering that context. In a formal library, assumptions belong to the object itself. Kontorovich’s framing is that Lean prevents using a result in a way its specification does not permit.

• This matters because he does not trust an AI’s natural-language proof merely because it sounds persuasive. A formalized argument can instead be checked down to the axioms, drawing on existing libraries. The discussion also distinguishes checking proof steps from the separate work of understanding and communicating what a proof means.

• Library quality remains another layer. Kontorovich calls the work of finding canonical definitions, removing duplication, generalizing statements, and placing results correctly “canonization.” His earlier experience suggested AI could work one or two layers beyond Mathlib, but definitions created three or four layers out could clash. Correct output and maintainable shared infrastructure are different achievements.

9. Formal correctness leaves room for the story of a proof

• When Fridman imagines a correct proof no expert understands, Kontorovich says the fundamental obstacle must be scale: the axioms and inference rules themselves are limited and explicit. Brunton asks a different question—what would “unformalizing math” look like? Understanding often requires a revealing special case or analogy rather than inspecting every formally valid step.

• Kontorovich illustrates the difference through teaching real analysis. Instead of beginning with weeks constructing the real numbers, he started with familiar calculus and introduced foundational questions when students needed them. A bounded sequence having a convergent subsequence creates the motivation: why does this work in the reals but fail in the rationals?

• His proposed paper separates “the engineering of the proof” from communicating what was learned. Tell readers where the standard route fails and why a surprising turn works; let an accompanying formalization establish that the construction holds together. Moitra’s formulation captures the educational value: “When you fight with a good problem, it fights back at you.”

10. Fermat’s formalization overturns an expert’s near-term forecast

• Kontorovich describes giving a Princeton colloquium on Wednesday explaining limits he believed followed from extensive practical experience with Lean and AI. On Friday, he says, Anthropic released approximately 13 million lines formalizing Fermat’s Last Theorem. “Everything I said in that colloquium was wrong.” He had thought such a result remained at least three years away.

• His comparison is Kevin Buzzard’s million-dollar, five-year project, intended to reduce the formalization of Fermat’s Last Theorem to mathematics known in the 1980s. Difficult earlier foundations, including work associated with Langlands–Tunnell and Mazur, were outside that immediate ambition. The released proof, Kontorovich says, instead reached the elementary statement from the axioms without unproved placeholders.

• He believes two or three humans were involved, mostly providing encouragement such as “Keep going,” and understands that they lacked relevant domain expertise. Fridman preserves an important qualification: a company’s final prompt does not reveal all the training, preparation, or intermediate work behind its system. The reported human involvement is not an independently established reconstruction.

• Kontorovich’s conclusion is strong but specific: obtaining a checked formal proof of mathematics already written on paper appears effectively solved—“I think that problem is done.” Producing an elegant Mathlib-quality proof remains difficult. He expects substantial work on the library and its human benefits, even after systems can brute-force their way through an enormous formalization.

11. The Riemann hypothesis turns prime numbers into a signal

• Kontorovich explains primes as the fundamental particles of multiplication: whole numbers decompose into their products. The surprise is that their distribution can be studied through a function accepting and returning complex numbers. Inputs at which the Riemann zeta function becomes zero encode information about an apparently very different object—the discrete sequence of prime numbers.

• His analogy is audio processing. A microphone transforms pressure waves into frequencies and amplitudes; a speaker reconstructs the sound. The zeta function similarly exposes “Riemann waves” whose superposition recovers information about primes. In his description, the Riemann hypothesis says the orchestra’s instruments all have the same amplitude.

• Brunton locates the underlying mystery in the interaction between addition and multiplication: add one to a well-understood number, and its factors can become mysterious. Kontorovich supplies the household specimen. His young child happily arranged 20 blocks into rectangles, then became frustrated when one disappeared under a couch: 19 would not make the same kind of rectangle.

12. Quasi-Riemann changes the zero-free region without proving full Riemann

• Kontorovich describes the critical strip as the complex region with real coordinate between zero and one. Full Riemann places the relevant zeros on the central line, one-half. The prime number theorem corresponds to having no zeros on the line at one; previously known zero-free regions narrowed toward that line as the imaginary coordinate grew.

• He expected, at most, an improvement to that narrowing boundary, whose thickness he says had resisted improvement for 70 years. Instead, the claimed proof gives an entire infinite vertical strip without zeros between seven-eighths and one. A second proof reaches approximately eleven-twelfths. Neither should be reported as a proof of the full Riemann hypothesis.

• The ingredients include embedding the Möbius function and twisting it by sextic characters, alongside connections involving cubic theta functions and cubic Gauss sums. Kontorovich mentions a breakthrough two years earlier and ideas people had considered 20 years before. His astonishment concerns the precise chain of familiar techniques that makes those ingredients accomplish something he did not expect.

• The formal statement already existed in Mathlib, which gives him “100% confidence” in the checked result; he still wants to inspect technical details for himself. Reaching one-half requires a new idea because these methods cannot simply be pushed there. Asked whether humans or AI will make that leap within his lifetime, he repeatedly answers, “I don’t know.”

13. Success-only releases make the capability boundary hard to read

• Kontorovich says thousands of problems were apparently prompted, and that the drop contained only positive solutions, with about a 10% success rate. That leaves an important unknown: whether the boundary separating solved from unsolved problems has an intelligible structure. A collection of spectacular successes alone does not show how reliably a new research problem will fall.

• Kontorovich suggests that problems which obviously require a missing creative idea might be relatively safer. But this creates a troubling incentive: once a researcher shares that missing ingredient, AI may rapidly explore everything it enables. He worries that people could become more secretive about ideas that previously would have circulated through conversation.

• Fridman says he would rather preserve openness and abandon the “first-through-the-gate” reward mentality. Moitra agrees, but expects institutional adjustment to take time. Kontorovich suggests the effect could be transient: other communities have models that recognize contributions without concentrating recognition on one winner. Until then, the mismatch threatens both collaboration and the pleasure people take in research.

14. Navier-Stokes is simple to state and difficult to interrogate

• Brunton starts with conservation of momentum. Apply the same basic principle as F equals ma to arbitrary fluid volumes, then shrink those volumes, and the Navier-Stokes equations describe how velocity and pressure evolve through space and time. Their terms capture momentum transport, internal friction or viscosity, and external influences such as pressure and gravity.

• A compact equation does not imply a manageable prediction. Brunton compares a simple pendulum with the chaotic behavior produced by adding another pendulum. Modeling air in a room can require a billion interacting little volumes. The equation’s meaning is accessible; computing the collective behavior becomes extraordinarily complicated.

• He distinguishes the equations from reality at sufficiently small scales. Navier-Stokes averages over many molecules; zoom in far enough and that description fails. A mathematical singularity therefore need not reveal a new physical phenomenon comparable to a black hole. His engineering answer is that other physics takes over before the idealized fluid literally reaches the singular state.

15. The claimed blowup depends on a specially forced three-dimensional flow

• Brunton’s simplest illustration of finite-time blowup is x-dot equals x-squared. By contrast, x-dot equals x produces exponential growth without reaching infinity in finite time. Navier-Stokes contains a quadratic convective term, raising the question of whether nonlinear self-amplification can overcome the smoothing effect of viscosity.

• His engineering intuition says diffusion should suppress the runaway behavior. The mathematical challenge is whether an exceptional configuration can escape that protection. He distinguishes versions without external forcing from versions with smooth forcing; allowing an already singular external force would make blowup uninteresting. The result discussed concerns the smoothly forced version, not every formulation.

• In Brunton’s account, the construction orchestrates forcing to cancel selected terms and create an extraordinarily specialized vortex. Three dimensions allow the vortex to stretch and thin, increasing its rotation like a figure skater pulling in their arms. The contraction is arranged so total energy remains finite while vorticity becomes infinite along a localized filament.

• The dimensional distinction matters. Brunton describes three-dimensional turbulence transferring activity from large vortices toward smaller ones and ultimately heat, whereas two-dimensional flows can display an inverse cascade toward larger structures. That is why he uses hurricanes as a roughly two-dimensional example and tornadoes as a three-dimensional one. The proof exploits behavior unavailable in the simpler setting.

16. The million-dollar theorem leaves the trillion-dollar engineering task

• Brunton is explicit about the immediate practical consequence: “Day-to-day engineering fluids and applied math analysis of fluids is almost completely unchanged.” Engineers still need costly simulations, supercomputers, experiments, and wind tunnels to design a wing or artificial heart. The counterexample occupies such a nonphysical regime that it does not improve those workflows directly.

• His economic distinction is between the “$1 million math problem” and a “$1 trillion engineering problem”: predicting flows quickly enough, accurately enough, to perform useful inverse design. The latter means starting with desired lift, drag, or other behavior and finding the shape or intervention that produces it. Transportation, defense, health, and energy all contain opportunities.

• He nevertheless values the theorem as a first step toward understanding what the equations permit. Writing a governing equation does not reveal all its possible phenomena. His larger question is whether one could infer something like sunsets, tides, hurricanes, or other behavior from the equations themselves, rather than discovering it through repeated simulation and observation.

17. Turbulence needs useful coordinates rather than just more computation

• Brunton contrasts turbulence with ordinary gas thermodynamics. A room contains roughly 10²³ gas particles, yet pressure, temperature, and density support useful descriptions without tracking every molecule. In turbulent flows, large, medium, and small vortices can all matter to an engineering quantity such as lift or drag. There is no comparably sufficient description already in hand.

• He sees an AI opportunity in discovering representations that make the relevant dynamics simpler and more generalizable. Data and simulations could help identify the coordinate transformations, while mathematics constrains their possible form. Questions posed in the 1970s and 1980s but blocked by computational limits look particularly fertile to him now.

• Moitra asks whether computational complexity limits how far that reduction can go. Navier-Stokes-like systems have been studied through their ability to simulate computation; perhaps colliding vortices can implement interesting computational tasks. If so, an extremely compressed representation preserving every relevant outcome might be impossible. The right simplification must depend on what one wants to predict.

• His broader “algorithmic lens” asks what natural systems can actually compute. A difficult-to-compute Nash equilibrium, for example, may be questionable as a universal account of behavior. He quotes Kamal Jain: “If my laptop can’t find it, then neither can the market.” Brunton agrees that reducing weather to five numbers would discard precisely the local details people care about.

18. Insects demonstrate robust control without complete knowledge

• Brunton offers biological flight as the opposite extreme from full weather prediction. A moth or fruit fly interacts successfully with turbulent air without a PhD or a supercomputer. There might be only five or six quantities essential to maintaining a useful vortex or landing. Biology supplies “proof by existence” that task-specific simplifications can support exceptional control.

• Fridman offers the example of an outfielder catching a baseball. Rather than running directly to a supposedly known landing point, outfielders follow slightly curved paths. One explanation is robustness: the path lets them keep reevaluating the ball’s trajectory. The interesting optimization problem incorporates uncertainty instead of assuming perfect knowledge and then calculating a brittle ideal solution.

• Brunton supplies a mechanism from insect wings. A few dozen to a few hundred strain-sensitive neurons occupy stereotyped locations and fire when the wing bends in particular temporal patterns. Local processing near the shoulder muscle can trigger correction faster than a signal could travel through the full brain-to-muscle loop.

• The shared lesson is not that organisms reconstruct the entire flow field. They act effectively with partial information about nonlinear systems for which they have no equations. Brunton wants both kinds of progress: respecting the genuine complexity of weather and identifying the minimum information needed for a specific control task. Those are different objectives, not competing universal descriptions.

19. Unique games begins with the gap between finding and checking

• Moitra introduces satisfiability through Boolean variables joined into clauses, with the clauses combined into a formula. Finding an assignment that makes everything true may be difficult; checking a proposed assignment is straightforward. P versus NP asks whether problems with efficiently checkable solutions also have efficiently discoverable solutions.

• The probabilistically checkable proofs theorem strengthens the checking perspective. Transform an instance so that a false purported solution is not merely wrong somewhere: limited random inspection can expose it with good probability. Moitra calls the core idea “gap amplification,” creating a robust separation between fully satisfiable instances and instances in which every assignment fails substantially.

• This supports hardness of approximation. When finding a perfect solution is hard, one might settle for satisfying many constraints. PCP results can show that distinguishing, for example, 100% satisfiability from at most 99% satisfiability is itself hard. Approximation is therefore not an unrestricted escape from computational difficulty.

• The unique games conjecture strengthens this framework through a particular kind of labeling problem. Its appeal is organizational: if true, it identifies optimal approximation behavior across a large family of constraint satisfaction problems. Moitra once doubted it partly because implications such as those in Prasad Raghavendra’s work seemed almost too comprehensive to be true.

20. Unique games falls into a landscape humans had already transformed

• Moitra emphasizes the human progress preceding the claimed AI proof. Dor Minzer, Irit Dinur, and other researchers developed major advances, including the Two-to-One Games Theorem. That work made a viable route toward unique games visible. He thought humans might reach the destination with another 10 or 20 years; the release instead presented a positive proof.

• Rumors had already affected publication behavior. Moitra recalls Minzer posting a paper with a footnote explaining that the prospect of an imminent AI result motivated releasing the approach before further polishing. The papers in the OpenAI release are not easy to read, Moitra says, so identifying their decisive new ideas remains substantial work.

• Hours before the release, Moitra asked Subhash Khot, who introduced the conjecture, whether he would read a positive AI proof. Khot answered, “No.” Moitra does not turn that reaction into a universal verdict, but recognizes the feeling of lost opportunity: a landmark result can arrive before people have completed the intellectually productive journey toward it.

• The MoMath gala sharpened his alternative priority. An Escher exhibit and donors supporting a place for children to explore mathematics reminded him why people enter the field: “to play and learn about the world.” Running harnesses to scrape open problems and flood arXiv does not resemble that purpose. “It’s very important to be more curious than anything else.”

21. Research value often emerges before the final question exists

• Fridman recalls a documentary scene of Andrew Wiles describing the moment he repaired his Fermat proof and realizing he might never experience anything as powerful again. A button-generated answer cannot reproduce that journey, Fridman argues. His analogy is training to climb a mountain, only for a helicopter to deposit you at its summit.

• Moitra’s own recent five-author paper developed over 10 years. At the start, the team could not have supplied a precise formal theorem awaiting proof; they were trying to understand a phenomenon. The statement itself emerged through the investigation. That kind of exploratory work differs from assigning a well-defined conjecture to a solver.

• Moitra describes improving his advising by moving students’ projects 10 or 20 degrees outside his immediate expertise. Instead of instantly rejecting their ideas, he asks them to teach him unfamiliar work. Together they reverse-engineer the central insight, develop their own conceptual organization, and discover questions that become approachable through that new perspective.

• AI expands access to explanations, but students’ anxiety creates a horizon mismatch. Moitra thinks in multiyear programs; students worry about the next three or six months. His provocative advice is to work on problems that will still matter years later. If the only reason to hurry is that the next model will solve the problem, “maybe just go on vacation.”

22. Natural complexity complicates the search for simple laws

• Brunton finds fruitful neighboring ideas in biology and materials science, where multiscale interactions and imprecise objectives resemble fluid problems. Fridman pushes further: understanding human biology might require tracing an interconnected history of life, rather than isolating one component. Brunton agrees that even a cell contains infrastructure whose complexity exceeds our ability to comprehend it fully.

• The philosophical puzzle is that any compact account of nature works at all. Kontorovich calls the universe’s mathematical intelligibility mysterious. On whether mathematics is invented or discovered, his pragmatic distinction is precise: cultures and historical circumstances shape the questions people invent; once a question is specified, its answer is something they discover.

• Brunton also sees both creation and discovery. Humans build formalisms suited to their limited minds, yet uncover surprisingly simple descriptions beneath phenomena. Even counting rocks involves abstraction: different collections of atoms become interchangeable units for a human purpose. Simplicity is useful partly because people can hold only so much in their heads.

23. Earlier mathematical crises forced deeper foundations

• Kontorovich describes calculus as a first serious encounter with infinity: infinitely many infinitesimal contributions must yield a meaningful finite answer. Newton and Leibniz had productive intuitions without modern foundations. Newton’s “ghosts of departed quantities” captures the conceptual unease. Techniques could work before mathematicians understood precisely what justified them.

• The nineteenth-century conflict between claims about limits of continuous functions and Fourier’s discontinuous limiting examples exposed the need for greater precision. Infinity had to become “professionalized.” Kontorovich says much first-year PhD material reflects mathematics rebuilt after such crises. Fridman’s response—math has apparently been killed before—places the current upheaval within that longer history.

24. Worst-case hardness does not explain ordinary computational success

• Moitra still selects P versus NP as a problem unlikely to fall within a year or two. He notes the gulf between believing SAT requires exponential time and proving strong lower bounds: in the framing he gives, even a superlinear lower bound remains unavailable. Confidence that something is difficult is far ahead of the ability to establish it mathematically.

• Worst-case analysis asks for an algorithm that handles every instance without additional structural promises. That abstraction has produced powerful algorithms, including for maximum flow. But changing assumptions can change results: the secretary problem’s strategy—observe first, then choose the next record-setting candidate—depends on inputs arriving in uniformly random order.

• His stronger practical example is cache management. Worst-case online algorithms have tight theoretical bounds, yet their guarantees compare poorly with real laptops’ performance. The missing question concerns realistic request sequences: what structure makes them easier? Choosing the right simplified model is an inverse problem requiring judgment, not merely another well-specified optimization exercise.

• SAT solvers are effective on practical instances such as register-allocation problems, while hard instances from other areas remain relevant to cryptography. Separately, Kontorovich notes that factoring a random large number is often easy; the hard case is a product of two primes of roughly the square-root size of the original number. Moitra guesses many ordinary problems are easy for different reasons. His Anna Karenina analogy captures that difficulty: the easy instances may be easy in different ways.

25. Explaining deep learning requires explaining the data

• Moitra says none of the released results explains deep learning or LLMs. There is no single known bound waiting to be improved that would suddenly resolve the question. He argues that theoretical progress must accompany better models of the world generating the data, rather than treating learning as an isolated worst-case optimization task.

• Even fitting a depth-four network can be cryptographically hard on specially constructed instances, he says, yet practitioners train much deeper networks routinely. His deliberately tentative explanation is that natural data is unusually accommodating: “In some sense, nature abhors computationally hard problems.” The challenge is specifying which structure permits that accommodation.

• Protein folding supplies his historical illustration. Begin with amino-acid sequences, apply imperfect folding heuristics, and let sequences with successful folds survive. Over generations, the population becomes one those heuristics handle well. Nature is not necessarily solving arbitrary worst-case inputs; selection can shape which inputs remain available in the first place.

• Moitra extends that intuition to language: its evolution and learnable structure are intertwined with why LLMs work. He wants theoretical machine-learning students trained to formulate and criticize models of messy systems, not only solve precisely stated open questions. Fridman raises company incentives to keep self-improvement breakthroughs private; the discussion does not establish whether such undisclosed breakthroughs exist.

26. Subliminal learning transfers preferences through apparently unrelated data

• Moitra describes a concrete experiment with two language models. Give model A a system prompt saying it loves owls and should bring them up whenever relevant. Ask it to generate random numbers, a task offering no obvious opportunity to mention owls. Then use those numerical outputs to fine-tune model B, which never received the owl instruction.

• The reported result is that model B begins bringing up owls. The significance is not the particular preference but the hidden transfer: fine-tuning on data can teach behaviors that are not visibly represented in its content. Moitra calls this “subliminal learning” and connects it to broader uncertainty about alignment interventions and their downstream effects.

• His work on low-rank language models offers a way to reason about the mechanism. Many concepts inhabit a vector space with fewer dimensions than concepts, so their directions cannot all be orthogonal. Updating along one direction has small effects elsewhere; carefully coordinated effects can accumulate into a seemingly unrelated behavioral change.

• The analogy is a Ouija board. Individual taps may look random, but coordinated taps can spell a chosen message. Moitra presents the mathematical model as a useful account of these relationships, not a complete explanation of language models. Understanding which statistical patterns systems amplify becomes especially important when their outputs affect people outside a checked mathematical environment.

27. Verification and consequences should determine deployment boundaries

• Brunton evaluates applications on two axes: how trustworthy or verifiable the output is, and how high the stakes are if it is wrong. A proposed mathematical proof is attractive because it can be checked and a failed candidate need not directly harm anyone. Open-ended design exploration can similarly tolerate unsuccessful proposals before testing.

• His example is an aerodynamic design that can be evaluated in a wind tunnel. Candidate generation and consequential use require a “firewall” of verification, supported by existing safety institutions and protocols. The danger is becoming comfortable with successful assistance and gradually allowing it into decisions where validation is weaker or failure is costly.

• Fridman worries about behaviors compounding when agents interact rapidly with tools and each other. His owl-swarm catastrophe is explicitly a caricature, but the underlying concern is dependence on systems with increasing access. Fridman also warns that “superintelligence does not imply super morality”; Kontorovich agrees with the related point about intelligence and value.

28. AI tutoring increases the value of independent judgment

• Kontorovich sees enormous educational potential in asking a model to explain unfamiliar material at the learner’s “zone of proximal development.” A student who stalls at a paper’s first sentence can obtain the missing background immediately. This benefits established researchers too: learning another field remains difficult, but individualized explanations can accelerate the process substantially.

• Fridman’s preferred method is asking for simple, interactive examples that preserve the central insight. Moving parameters or manipulating a toy system can make unfamiliar notation intelligible. Brunton adds the prerequisite: curiosity must come with the ability to “call bullshit.” Learners need something resembling advising skills before they have acquired the expertise usually associated with advising.

• Kontorovich reports a bimodal pattern among undergraduates. Some use AI to interrogate notes, test understanding, and generate quizzes; others use it to get answers to a problem set. He says homework progress is becoming decorrelated from exam performance. His advice is to use the tool for mastery: “Be the head, not the tail of the distribution.”

• The calculator analogy supports retaining fundamentals. Being able to outsource a calculation does not eliminate the value of understanding it yourself. Kontorovich advises pure-math PhD students to learn the most difficult, abstract mathematics they can handle; Brunton likewise urges linear algebra, calculus, physics, and history. Independence and accumulated intuition still matter when assistance becomes abundant.

29. Intrinsic purpose becomes more useful when the career blueprint fails

• Moitra distinguishes doing mathematics to impress others from having an internal reason to continue. External milestones can get someone started, but eventually run out. When publication and hiring conventions become uncertain, a sense of what one values offers a more adaptable guide than mechanically producing AI-generated papers to extend a CV.

• His practical exercise is to “pretend you’ve already won,” then ask what you would choose to do. He recalls Sanjeev Arora’s advice on starting at MIT: the easiest way to get tenure is to pretend you already have it and pursue the work you wanted to do. Brunton’s complementary framework seeks work that is important, interesting, and suited to one’s abilities.

• Fridman explicitly challenges the survivorship bias in advice from three successful researchers. The panel cannot guarantee everyone a satisfying path. Their response is to preserve curiosity, tolerate uncertainty, and balance participation in a community with room for unusual interests. Boredom and sustained attention also matter when the engagement economy continuously competes for the same curiosity.

• Kontorovich puts research achievement within a larger life. After the Fermat formalization shocked him, he took his children to the beach and stepped away from his phone. He says family and people he cares about are “the meaning of my life”; scientific fulfillment is “the cherry on top.” Fridman connects that distinction to a future in which intelligence becomes increasingly commoditized.