Dwarkesh Podcast• Mar 20, 2026• 1:23:44Interview

Terence Tao – How the world’s top mathematician uses AI

From Dwarkesh Podcast

Terence Tao•Mathematician

Executive Summary

The historical process of scientific discovery, exemplified by Kepler using Brahe's data, is shifting.
AI has made hypothesis generation nearly free, making rigorous data analysis and verification the new bottleneck.
Terence Tao characterizes current AI as a powerful tool for mathematical 'breadth'—applying known techniques at scale—but it lacks the 'depth' for novel conceptual breakthroughs, with a success rate on truly hard problems remaining very low (1-2%).
Tao predicts a future dominated by hybrid human-AI collaboration.
AI will automate routine work and enable a new 'experimental' side of mathematics, while humans will focus on strategy, intuition, and creating new frameworks.
He stands by his 2026 prediction of AI as a 'trustworthy co-author' but believes a full replacement of human mathematicians is not imminent, citing AI's fundamental inability to build cumulatively on partial progress.

11 quotes

Concerns Raised

Current AI models cannot build on partial progress, limiting their ability to solve complex, multi-step problems.
The flood of AI-generated ideas and papers makes verification and quality control the new scientific bottleneck.
Over-reliance on AI for answers could inhibit the development of deep intuition and problem-solving skills in researchers.

Opportunities Identified

AI will function as a 'trustworthy co-author' by 2026, handling routine tasks and checking for errors.
AI will enable a new paradigm of 'experimental mathematics,' allowing for large-scale data analysis of mathematical structures and techniques.
Hybrid human-AI teams will dominate research by combining AI's breadth and speed with human depth and intuition.

Key Themes

The Shifting Bottleneck in Scientific Discovery

Using Kepler's discovery of planetary motion as a historical analog, the conversation posits that AI has drastically lowered the cost of generating hypotheses. This shifts the primary challenge in science from creative ideation to the rigorous verification and evaluation of a flood of new, often low-quality, ideas.

This reframes the core work of scientists and researchers, emphasizing skills in experimental design, data analysis, and critical evaluation over pure 'eureka moment' ideation.

AI as a 'Breadth' vs. Human 'Depth' Engine

Terence Tao frames AI's current strength as its ability to explore a vast 'breadth' of possibilities, such as systematically applying every known mathematical technique to a problem. However, it lacks the human expert's 'depth' of understanding, intuition, and ability to invent entirely new concepts when existing methods fail.

This highlights the complementary nature of AI and human intelligence, suggesting that the most effective approach for the foreseeable future will be hybrid teams that leverage the strengths of both.

The Rise of Experimental Mathematics

Mathematics has traditionally been an almost purely theoretical discipline. Tao predicts AI tools will revolutionize the field by enabling an 'experimental' side, allowing for large-scale data gathering and analysis to identify patterns and test the efficacy of different problem-solving approaches at scale.

This signals a fundamental paradigm shift in mathematical research, creating new avenues for discovery and changing how mathematicians are trained and what skills are valued.

The Limits of Current AI Architectures

Despite impressive successes on specific problems (like some Erdős problems), current AI models have a very low success rate (1-2%) on truly hard, open problems. Tao points to a key limitation: their inability to build cumulatively on partial progress within a single problem-solving session, as they lack a persistent state of understanding.

This provides a sober assessment of AI's capabilities, indicating that fundamental architectural breakthroughs are likely needed before AI can autonomously solve the most profound scientific and mathematical challenges.

Get started free

Topics

AI in Science Mathematics Terence Tao History of Science Johannes Kepler Tycho Brahe Planetary Motion Hypothesis Generation Scientific Method Large Language Models (LLMs)Experimental Mathematics Erdős Problems Human-AI Collaboration Future of Work Problem Solving

Processed Apr 3, 2026 yt-dlp + mlx-whisper + Gemini