Most important take away

AI has driven the cost of hypothesis generation in science down to nearly zero, but this shifts the bottleneck to verification, validation, and assessment of which ideas actually constitute progress. The future of mathematics and science lies in human-AI complementarity: AI excels at breadth (trying thousands of approaches at scale) while humans excel at depth (cumulative reasoning, building on partial progress, and recognizing which results matter), and we must redesign how we do science to leverage both.

Summary

Key Themes:

• AI as breadth, humans as depth. Tao draws an analogy between AI and Kepler’s brute-force approach to testing hypotheses against Brahe’s data. Current AI tools can sweep across hundreds of problems applying known techniques, succeeding roughly 1-2% of the time on any given problem, but at scale that still yields impressive results. However, they lack the ability to build cumulatively on partial progress the way human collaborators do.

• The shifting bottleneck in science. Just as the internet drove communication costs to near zero, AI has made idea generation almost free. The new bottleneck is verification and evaluation. Scientific institutions like peer review are already being overwhelmed by AI-generated submissions, and new paradigms are needed to filter signal from noise at scale.

• AI progress in mathematics: simultaneously amazing and disappointing. About 50 of 1,100 problems on the Erdos problem website have been solved with AI assistance, but the low-hanging fruit has been picked and progress has plateaued. Most solved problems were ones that had received little human attention and yielded to combinations of existing techniques. Pure AI one-shot solutions have largely stopped.

• Artificial cleverness vs. artificial intelligence. Tao distinguishes between AI’s current “cleverness” (trial-and-error jumping that either succeeds or fails) and true intelligence (the cumulative, adaptive process of building understanding through conversation and partial progress). Current models lack persistent learning across sessions.

• The need for formal languages for mathematical strategy. Lean has formalized proofs effectively, but there is no equivalent framework for formalizing mathematical strategies, conjectures, or plausibility arguments. Creating such a framework could unlock AI’s ability to contribute to the more subjective, narrative aspects of science.

• Doing mathematics at scale is in its infancy. Tao envisions a future where AI enables an experimental side of math, running thousands of problems through systematic studies to discover what techniques work broadly, rather than the traditional approach of deep focus on individual problems.

• Tao’s personal productivity with AI. His papers now include more code, plots, and deeper literature searches. Tasks that would have taken hours now take minutes. The core difficulty of solving the hardest part of a math problem has not changed, but auxiliary tasks are dramatically faster. He estimates papers of the kind he now writes would take 5x longer without AI, though he would not have written them the same way.

Actionable Insights:

• Embrace complementarity. Whether in math, science, or engineering, pair AI’s breadth capabilities with human depth. Use AI to map out problem spaces and clear easy observations, then focus human expertise on the identified hard spots.

• Invest in verification infrastructure. As AI generates hypotheses at scale, the ability to evaluate and validate those ideas becomes the critical differentiator. Build systems and benchmarks for standardized evaluation rather than relying on cherry-picked success stories.

• Write things down. Tao credits his blog with preserving insights he would have otherwise lost. Recording what you learn, even informally, compounds over time.

• Leave room for serendipity. Over-optimizing schedules eliminates the accidental discoveries that come from unplanned interactions. Tao finds that some distraction and randomness is essential for sustained inspiration.

Career Advice:

• Tao advises those early in a math career to adopt an adaptable mindset. Traditional education still matters, but non-traditional opportunities now exist where even high school students can contribute to frontier research using AI tools and formal proof systems like Lean. Embrace change and be open to radically different ways of doing science.

No specific stocks or investments were mentioned in this episode.

Chapter Summaries

Kepler, Brahe, and the origins of data-driven science — Tao recounts how Kepler’s discovery of planetary motion laws emerged from years of testing hypotheses against Brahe’s meticulous observational data. Kepler’s beautiful but wrong Platonic solids theory eventually gave way to elliptical orbits, illustrating how data-driven iteration, not just genius insight, drives scientific progress.

AI as high-temperature hypothesis generation — Dwarkesh and Tao discuss the analogy between LLMs and Kepler’s brute-force approach. Tao argues AI has made idea generation nearly free, but the bottleneck has shifted to verification. Science needs new structures to evaluate the flood of AI-generated hypotheses.

Why correct theories initially look worse — The conversation explores how paradigm-shifting theories (Copernicus, Newton, Darwin) often appeared less accurate or made puzzling implications when first proposed. Assessing whether a new idea constitutes progress requires future context, making it nearly impossible to automate such evaluation.

Darwin, Newton, and the role of communication in science — Tao highlights that scientific progress depends not just on correctness but on persuasive communication. Darwin’s plain English writing helped his ideas spread, while Newton’s secretive Latin publication delayed adoption of his work.

The deductive overhang and extracting information from data — Astronomy pioneered squeezing every drop of insight from limited data. Tao suggests there may be far more information extractable from existing data than we realize, and clever metrics could help detect which scientific developments represent real progress.

AI progress on the Erdos problems: a plateau after initial successes — About 50 problems were solved by AI, mostly low-hanging fruit that had received little human attention. Large-scale systematic sweeps show a 1-2% success rate per problem. AI tools struggle with creating partial progress or identifying intermediate steps.

Breadth vs. depth: redesigning science — AI excels at breadth while humans excel at depth. Tao envisions complementary science where AI maps out entire fields, clears easy problems, and identifies islands of difficulty for human experts. This paradigm of “mathematics at scale” is brand new.

Artificial cleverness vs. artificial intelligence — Tao distinguishes current AI capabilities (trial-and-error jumping) from true intelligence (cumulative, adaptive reasoning built through interaction). Models lack persistent learning and cannot build on partial progress across sessions.

Formal languages for mathematical strategy — Lean has formalized proofs, but there is no equivalent for strategies, conjectures, or plausibility arguments. Tao wishes for a semi-formal framework that could help automate the subjective, narrative aspects of mathematical reasoning.

Tao’s personal productivity and learning process — AI has enriched Tao’s papers with more code, plots, and literature searches, making auxiliary tasks dramatically faster. The core difficulty of deep mathematical problem-solving remains unchanged. He learns new fields through collaboration, obsessive curiosity, and writing on his blog.

Serendipity and the value of unoptimized time — Tao emphasizes the importance of leaving room for unplanned interactions and accidental discoveries. Over-optimization of schedules, accelerated by remote work and AI, risks eliminating the randomness that fuels inspiration.

Career advice for aspiring mathematicians — In an era of rapid change, Tao recommends an adaptable mindset. Traditional education remains valuable, but non-traditional paths to contributing at the frontier are now possible through AI tools and formal proof systems. It is a scary but exciting time.