Terence Tao: Hardest Problems in Mathematics, Physics & the Future of AI | Lex Fridman Podcast #472

Tao emphasizes that the most interesting problems are those where existing techniques do 80–90% of the work but fail on a crucial remaining piece, like Kakeya or Navier–Stokes; these boundary cases expose where our methods and intuitions truly break down.

In equations like 3D Navier–Stokes, nonlinear transport dominates dissipative effects at small scales, allowing energy to cascade into finer structures and potentially blow up; this same supercriticality underlies why we can forecast planetary motion far ahead but not weather beyond about two weeks.

Tao’s averaged Navier–Stokes blowup construction shows that many tempting proof strategies for global regularity must fail, because slight variants of the equation already blow up; such “obstructions” prune whole families of doomed approaches and sharpen what a successful proof must exploit.

The central limit theorem and related universality principles show why bell curves and simple macroscopic laws emerge from vast micro‑complexity, yet Tao notes that when hidden correlations or systemic shocks exist (e.g., in finance), assuming Gaussian behavior leads to catastrophic mispricing of risk.

Results like Szemerédi’s theorem, Tao’s work on primes in arithmetic progressions, and inverse theorems show that objects are either highly random or secretly structured (and thus near a simpler model); leveraging this dichotomy lets mathematicians prove robust patterns in primes and dense sets.

Lean can turn proofs into checkable certificates and support large‑scale, trustless collaboration (e.g., 22 million small algebra problems solved by ~50 people), while AI already acts as a powerful autocomplete and search aid; Tao expects a phase transition when formalization becomes faster than handwritten proofs.

Tao stresses that humans excel at sensing when an approach is wrong, breaking hard problems into selectively simplified sub‑problems (“installing cheats”), and switching problems before obsession becomes a “mathematical disease”; current AIs can generate plausible text and code but lack reliable meta‑judgment about their own dead ends.