Hamkins, Joel David
▶ Watch on YouTubeCore ideas
Hamkins builds the episode as a guided tour from Cantor's original crisis of infinity up through his own controversial "multiverse" philosophy of set theory. He opens with Galileo's paradox (perfect squares can be matched one-to-one with all natural numbers, seemingly making the whole equal to a part) as the ancestor of Cantor's insight, then walks through Hilbert's Hotel to build intuition for why countable infinity is closed under nearly everything you throw at it (adding one guest, a busload, even an infinite train of infinite train-cars all still fit), before showing why the real numbers break this pattern via Cantor's diagonal argument — the same underlying "diagonalization" logic he shows recurs, structurally identically, in Russell's paradox (no set of all sets), the undecidability of the halting problem, and Gödel's incompleteness theorems. He narrates Frege's devastation upon receiving Russell's letter just as his life's work on reducing mathematics to logic was going to press, and frames Zermelo-Fraenkel set theory (ZFC) — including the once-fiercely-contested axiom of choice — as the eventual, largely successful fulfillment of the logicist project once contradictions forced axiomatization.
A long, careful section separates truth from proof (a distinction he argues wasn't even clearly drawn before Gödel and Tarski) and reconstructs Hilbert's program — a strong theory to answer every mathematical question, backed by a finitary proof that the theory is safe — as the thing Gödel's two incompleteness theorems decisively demolish: no sufficiently powerful, computably axiomatized theory can be both complete and able to prove its own consistency. Hamkins insists this isn't "traumatic" but clarifying, and walks through an elegant halting-problem-based proof of incompleteness (diagonalizing a hypothetical halting-decider against itself) as, in his view, the simplest route to Gödel's result. This sets up the episode's philosophical centerpiece: the continuum hypothesis, first posed by Cantor and never resolved despite decades of partial progress (Cantor's own proofs for open, closed, and eventually — with enough large cardinal axioms — projective sets of reals), until Gödel (1938) showed it can't be refuted and Cohen (1963, via the technique of "forcing") showed it can't be proved from ZFC either. Rather than treating this independence as a defect to be patched by ever-stronger axioms (the "universe view," associated with his own former advisor Hugh Woodin's program for "Ultimate L"), Hamkins defends the "multiverse view": there is no single true set-theoretic universe, but a vast interconnected landscape of set-theoretic worlds, related by forcing, in which statements like the continuum hypothesis are simply true in some and false in others — independence isn't a failure to find the right question but evidence of "carving nature at its joints."
The conversation ranges widely through Hamkins's other work and stated tastes: the surreal numbers (John Conway's system that generates every number — reals, ordinals, and infinitesimals alike — from a single recursive left-set/right-set rule, which Conway reportedly considered underappreciated as his greatest disappointment); a proof that "almost all" Turing machines provably never halt for trivial structural reasons (a "black hole" in the halting problem that lets you computably decide the vast majority of instances even though the general problem is undecidable); infinite chess, where board positions can have transfinite game-values (a forced win for white in some ordinal number of moves, with black able to choose how large a finite countdown to make it, but never a fixed number of moves) — developed with chess-master collaborator Corey Evans; and reflections on mathematical realism (Platonism), arguing that abstract mathematical objects have a clearer, less mysterious mode of existence than physical objects do, since physics only becomes stranger with deeper investigation while mathematical understanding only becomes clearer. He closes on AI skepticism specific to mathematics: current LLMs, in his experience, produce arguments optimized to look like proofs rather than grounded in actual mathematical understanding, which he finds genuinely dangerous for anyone insufficiently critical (drawing an analogy to his own undergraduate mistake of trusting beautifully-typeset LaTeX homework as a proxy for correctness) — while remaining open that this may change, and that his own collaborative, MathOverflow-driven, playful style of mathematics (as opposed to solitary multi-year grinds like Andrew Wiles's) has been unusually well suited to social exchange of ideas.
Excerpts
On why the halting problem being undecidable doesn't mean nothing can be proven about it, illustrated by his "black hole" result:
“It turns out we proved that there is a black hole... there's a computer procedure that decides correctly almost every instance of the halting problem... the asymptotic density of the programs is one... the probability-one behavior of a random Turing machine is that the head falls off the tape before it repeats a state. And that is the stupid proof that shows how to solve the halting problem" for almost all cases, even though it remains impossible to solve in general.”
On why he rejects the "universe view" that the continuum hypothesis must have one true answer, in favor of his pluralist multiverse view:
“The developments of set theory over the past half-century or more as evidence that there isn't such a unique set-theoretic reality. Rather, what we've been doing for decades now is producing more and more alternative set-theoretic universes in which the fundamental truths differ from one to the other. And that is the answer to the continuum hypothesis question: the fact that given any model of set theory, there's a forcing extension where the continuum hypothesis is true, and another one where it's false. You can sort of turn it on and off like a light switch.”
On why he distrusts current LLMs for mathematical work, and the danger of mistaking polish for correctness:
“The AI is trying to give me an argument that sounds like a proof rather than an argument that is a proof. The motivation is misplaced... they're not designed to make a logically correct argument. They're designed to make something that looks like a logically correct argument. And it's easy to get fooled if you're not skeptical." He compares this to his own undergraduate error of trusting beautifully LaTeX-typeset homework as if its typography guaranteed its correctness.”