Dwarkesh Podcast

Terence Tao on Dwarkesh Patel: How Erdős Problems Exposed AI

Tycho Brahe data let Kepler derive orbital laws by regression on six points; AI solved 50 Erdős problems fast then stalled on cumulative partial progress.

DPDwarkesh PatelhostTTTerence Taoguest

Mar 20, 2026·1h 23m·Watch on YouTube ↗

📖 CHAPTERS

1. 1

   Kepler’s data-driven leap: from Platonic solids to elliptical orbits

   Tao retells how Kepler, inspired by geometric beauty, tried to fit planetary spacings with Platonic solids before Tycho Brahe’s precise observations forced a rethink. Kepler’s breakthrough came from years of iterative model-testing against high-quality data, culminating in his laws of planetary motion without an underlying explanation.

2. 2

   “Kepler as a high-temperature LLM”: idea generation vs verification

   Dwarkesh proposes that Kepler resembles an LLM generating many speculative hypotheses, with validation filtering signal from “slop.” Tao agrees the generative phase matters, but stresses that verification and data quality are equally essential—and increasingly the real bottleneck.

3. 3

   From classic scientific method to big-data-first discovery—and regression pitfalls

   Tao contrasts traditional “hypothesis then test” science with modern workflows that start from massive datasets and infer laws statistically. Kepler’s third law is reframed as early regression on very few datapoints, illustrating both the power and fragility of pattern-finding from limited evidence.

4. 4

   AI slop at scale: the new bottleneck is evaluation and filtering

   With AI driving the marginal cost of idea generation toward zero, Tao argues scientific institutions face overload. The challenge becomes how to validate, prioritize, and decide which outputs are genuine progress amid floods of low-quality submissions.

5. 5

   Spotting new unifying concepts: why “importance” is context-dependent

   Dwarkesh asks how we’d recognize a new concept (like the “bit”) amid oceans of mediocre work. Tao argues significance often only emerges after adoption, applications, and cultural path dependence; many ideas look unimpressive or wrong initially.

6. 6

   When progress looks worse: partial theories, deleted assumptions, and persuasion

   They discuss how historically correct theories often began less accurate or more mysterious than rivals (Copernicus vs Ptolemy; Newton’s action-at-a-distance). Tao emphasizes the narrative and communication layer—Darwin’s persuasive exposition vs Newton’s Latin and secrecy—as a key driver of uptake and progress.

7. 7

   Deductive overhang and extracting more signal: astronomy’s “Sherlock” culture

   Dwarkesh brings up “deductive overhang”—how much could be inferred from existing signals with better methods. Tao notes astronomy’s extreme data scarcity trained a culture of extracting maximal information, and suggests similar clever inference tricks could be applied broadly, including to science-of-science metrics.

8. 8

   Erdős problems and AI’s plateau: breadth wins, depth still hard

   Tao reviews the burst where AI-assisted systems solved ~50 Erdős problems, followed by a slowdown after low-hanging fruit. He characterizes current models as good at “jumping” to solutions occasionally but poor at cumulative partial progress and identifying meaningful intermediate milestones.

9. 9

   Selection bias in AI ‘discoveries’: why systematic evaluation matters

   Tao warns that public perception is skewed by highlighting rare wins while ignoring the many failures. Systematic sweeps show low per-problem success rates, motivating standardized benchmarks rather than relying on company-reported or social-media-amplified successes.

10. 10

    AI makes math papers richer and broader—but not deeper (yet)

    Dwarkesh asks about productivity gains; Tao says AI drastically reduces “secondary” work (plots, code, formatting, literature scanning), changing what papers contain. But it hasn’t meaningfully sped up the hardest creative steps of proving new results, so output becomes broader rather than deeper.

11. 11

    Artificial cleverness vs intelligence: the missing ingredient is cumulative learning

    Tao distinguishes current AI’s “cleverness” (trial-and-error, recombination) from human-like intelligence (adaptive refinement, building on partial progress). He notes current sessions don’t retain new skills; progress doesn’t accumulate within the model the way it does within a collaborating human pair.

12. 12

    If AI solves big theorems, can humans extract understanding from formal proofs?

    They explore whether future breakthroughs could arrive as inscrutable Lean code or brute-force case analyses (like the four-color theorem). Tao argues that once a formal proof artifact exists, it can be analyzed, summarized, refactored, and “ablated,” potentially enabling humans (and other AIs) to uncover the key ideas.

13. 13

    Toward a semi-formal language for strategies, plausibility, and scientific argument

    Tao proposes the need for languages that capture not just deductive proof but the semi-formal reasoning scientists use: plausibility, heuristics, evidence, and narrative. He illustrates with the primes: data-driven conjectures, probabilistic heuristics, and community belief systems that guide research despite lacking formal proof.

14. 14

    How Tao learns new fields—and how he’d spend time: obsession, writing, serendipity, and hybrids

    Tao describes his learning process: collaboration, an “obsessive completionist” drive to understand techniques, and writing to prevent knowledge decay. He argues that optimized schedules can kill serendipity, and predicts human–AI hybrids will dominate mathematics for a long time, shifting what mathematicians do rather than simply replacing them.