Dwarkesh Podcast: Terence Tao – Kepler, Newton, and the True Nature of Mathematical Discovery
Date: March 20, 2026
Host: Dwarkesh Patel
Guest: Terence Tao, mathematician
Theme: How historical scientific discovery parallels modern AI’s contributions to mathematics, scientific bottlenecks, and the evolving role of AI in math research.
Episode Overview
Dwarkesh Patel interviews Terence Tao about the origins of mathematical discovery, using the story of Kepler and Newton as a lens through which to examine AI’s place in modern math. They discuss the shifting bottlenecks of science, the dynamics between breadth and depth in problem-solving, and how mathematicians (and future AIs) might be most productively used. The conversation spans historical anecdotes, philosophical questions about progress, and concrete observations about the current AI-for-math landscape.
Key Discussion Points & Insights
1. Kepler, Newton, and the Nature of Mathematical Discovery
-
Kepler’s Data-Driven Approach:
- Kepler built on Copernicus' heliocentric model, initially believing in elegant, geometric regularity (Platonic solids) between planet orbits.
- After access to Tycho Brahe’s high-quality data, Kepler abandoned his original beautiful theory because it didn’t quite fit the observations. He painstakingly analyzed the data, leading to the discovery that planetary orbits are elliptical (not circular) and formulating his three laws of planetary motion.
- Quote:
"He worked on this problem for years…eventually he figured out how to use the data to work out the actual orbits of the planets...he eventually worked out that they [orbits] are ellipses, not circles, which was shocking for him." (B, 03:15)
- Newton later explained Kepler’s laws with his theory of gravity.
-
Parallel with AI:
- Dwarkesh draws a parallel between Kepler's process—trying numerous hypotheses until finding one that fits the data—and modern large language models (LLMs), which can generate and test massive numbers of possible theories thanks to computational scale.
“Kepler was a high temperature LLM…through his career he’s just trying random relationships…In there is the cube square law...I feel like LLMs can do the kind of thing of like 20 years. Let's try random relationships, some of which make no sense, as long as there's a verifiable data bank.” (A, 04:09)
- Dwarkesh draws a parallel between Kepler's process—trying numerous hypotheses until finding one that fits the data—and modern large language models (LLMs), which can generate and test massive numbers of possible theories thanks to computational scale.
2. Shifting Bottlenecks in Scientific Discovery
- Past vs. Present Bottlenecks:
- Historically, the bottleneck was in "hypothesis generation," the ‘eureka’ moments that led to scientific progress.
“Idea generation has always been kind of the prestige part of science…these eureka, genius moments of idea generation.” (B, 05:44)
- Now, with AI making idea generation cheap and fast, the new bottleneck is verification, validation, and selection:
“AI has basically driven the cost of idea generation down to almost zero…the rest of the aspects of science have to catch up.” (B, 12:17)
- Systems for evaluating, verifying, and socially propagating promising ideas are overwhelmed by this abundance.
- Historically, the bottleneck was in "hypothesis generation," the ‘eureka’ moments that led to scientific progress.
3. Scientific Progress: Social, Historical, and Communicative Factors
-
Test of Time and Reception:
- Scientific breakthroughs are often only recognized as such retrospectively, dependent on not only inherent merit but also cultural context, communication, and subsequent applications.
- Example: Deep learning as a formerly niche field, or the bit as a cross-disciplinary abstraction.
"One reason why it's hard to assess whether a given idea is going to be fruitful is that it depends on the future. It depends also on the culture and society, which ones get adopted, which ones don't." (B, 15:13)
- Scientific breakthroughs are often only recognized as such retrospectively, dependent on not only inherent merit but also cultural context, communication, and subsequent applications.
-
Progress Can Appear Worse Before Better:
- New paradigms (e.g., Copernicus’ heliocentrism) can be less accurate initially than mature but flawed predecessors.
“Copernicus theory was a lot simpler, but…much less accurate. It was only Kepler that made it more accurate than Thomley's theory.” (B, 18:48)
- Removing faulty assumptions (e.g., earth at rest; species as static) is integral to this leap.
- New paradigms (e.g., Copernicus’ heliocentrism) can be less accurate initially than mature but flawed predecessors.
-
Communication Matters:
- Darwin’s theory, despite lacking a mechanistic basis, gained rapid acceptance in large part due to articulate, accessible writing—unlike Newton, who wrote in Latin and withheld discoveries.
“The art of exposition and making a case and creating a narrative is also a very important part of science.” (B, 23:01)
- Darwin’s theory, despite lacking a mechanistic basis, gained rapid acceptance in large part due to articulate, accessible writing—unlike Newton, who wrote in Latin and withheld discoveries.
4. Breadth vs. Depth: Human and AI Strengths
-
AI’s Breadth, Human’s Depth:
- AI’s main strength is in breadth—brute-forcing through problem landscapes and clearing out ‘low-hanging fruit’ at scale.
“They excel at breadth and humans excel at depth...” (B, 35:21)
- New paradigms for science could leverage this: AIs quickly clear easy parts, map new fields, humans focus on hard ‘islands.’
- AI’s main strength is in breadth—brute-forcing through problem landscapes and clearing out ‘low-hanging fruit’ at scale.
-
Math Case Study – Erdos Problems:
- A burst of rapid progress: AI-assisted or AI-driven solutions solved ~50 out of 1100 Erdos problems until hitting a plateau. Now, the "low walls" have been climbed, but higher walls still require human (and perhaps future AI) depth.
“Someone might use AI to generate a possible proof strategy and then another person will use a separate AI tool to critique it...” (B, 31:03) “These tools, they either succeed or they fail. And they've been really bad at creating sort of partial progress or identifying intermediate stages that you should focus on first.” (B, 33:14)
- A burst of rapid progress: AI-assisted or AI-driven solutions solved ~50 out of 1100 Erdos problems until hitting a plateau. Now, the "low walls" have been climbed, but higher walls still require human (and perhaps future AI) depth.
-
Exploitation of Breadth in the Future:
“We have to redesign the way we do science to take full advantage of this breadth capability…” (B, 35:21)
5. AI for Math: Today’s Limits and Promise
-
State of AI Tools:
- AI is already boosting productivity by automating auxiliary tasks, literature searches, generating figures, and combining existing techniques—but fundamental, creative breakthroughs remain largely human.
“The core of what I do, actually solving the most difficult part of a math problem—that hasn’t changed too much. I still use pen and paper for that.” (B, 48:01)
- Most progress comes not from radical new AI techniques, but combining neglected, known methods with search and scale.
- AI is already boosting productivity by automating auxiliary tasks, literature searches, generating figures, and combining existing techniques—but fundamental, creative breakthroughs remain largely human.
-
Brute Force vs. Insight:
- Some math (e.g., Four Color Theorem) can be settled by brute force, but big conjectures (e.g., Riemann Hypothesis) probably require conceptual breakthrough—and it’s unclear if AI alone can provide ‘insight.’
“Some problems have been basically solved by pure brute force...But part of the reason we we prize problems like the hypothesis that we're pretty sure that… a new type of mathematics has to be created.” (B, 53:39)
- Some math (e.g., Four Color Theorem) can be settled by brute force, but big conjectures (e.g., Riemann Hypothesis) probably require conceptual breakthrough—and it’s unclear if AI alone can provide ‘insight.’
-
Formalization & Abstraction:
- Discussion on the importance and challenge of building not just formal proof systems, but also languages for mathematical strategies and heuristics—the “softer” parts of discovery.
“If there was some semi-formal framework where this could be done semi-automatically in a way that isn’t sort of easily hackable…” (B, 59:34)
- Discussion on the importance and challenge of building not just formal proof systems, but also languages for mathematical strategies and heuristics—the “softer” parts of discovery.
6. How Humans Learn and Advance in the Face of AI
-
Tao’s Approach to Learning:
- He describes himself as a “fox” (breadth-focused), often learning from obsession, collaboration, and writing.
“If there’s something which I read about, which I feel like I should understand, I have the capability to understand this, but I don’t understand why it works...then I want to find out what was their trick.” (B, 70:06)
- Writing blog posts as a form of cementing knowledge.
- He describes himself as a “fox” (breadth-focused), often learning from obsession, collaboration, and writing.
-
Importance of Serendipity:
- “Accidental” discovery is valuable. Hyper-optimization (e.g., perfect scheduling, remote meetings) often kills serendipity and hinders unexpected learning and creativity.
"A lot of events that I kind of reluctantly went to...I often find interactions...I would learn interesting things...I do believe a lot in serendipity." (B, 73:28)
- “Accidental” discovery is valuable. Hyper-optimization (e.g., perfect scheduling, remote meetings) often kills serendipity and hinders unexpected learning and creativity.
Notable Quotes and Memorable Moments
-
“He had this theory which he thought was absolutely beautiful…And then he worked out, to kind of his disappointment, that his beautiful theory didn't quite work.”
(Tao, 01:45) -
On scientific progress and abundance:
“AI has driven the cost of idea generation down to almost zero, …but it doesn't create abundance by itself.” (B, 12:17)
-
On human vs. AI productivity in math:
“The papers that go in for the top journals are usually ones where the existing methods can kind of solve 80% of the problem, but then this is 20% which is resistant, and a new technique has to be invented to fill in the gaps.” (B, 41:35)
-
On the possible impact of AI on the Riemann Hypothesis:
“It could be a collaboration of a type that doesn't exist yet.” (B, 54:36)
-
On the unpredictable path of science:
“Maybe, by somehow destroying serendipity, we actually inhibit certain types of progress. Anything is possible really at this point.” (B, 80:19)
-
Advice for young mathematicians:
“We live in a time of change…one just has to embrace that there's going to be a lot of change... But you should also be open to very, very different ways of doing science, some of which don't exist yet. So it's a scary time, but also very exciting.” (B, 81:28)
Timestamps for Important Segments
| Timestamp | Segment | | --------- | ------- | | 00:15–04:09 | Kepler’s journey: How he discovered his laws, interplay with Platonic ideas and Brahe’s data | | 04:09–08:48 | Analogy to AI and LLMs: Data-driven discovery, validation, and the “LLM as Kepler” analogy | | 12:17–15:13 | Idea generation abundance & new bottlenecks | | 18:48–21:35 | Apparent regressions in new scientific paradigms | | 26:10–28:51 | Deductive overhang and value of extracting more from existing data (cosmic distance ladder) | | 31:03–34:19 | AI’s burst in solving Erdos problems; limits at higher difficulty | | 35:21–36:55 | Breadth vs. depth; complementarity of human and AI efforts | | 41:35–45:16 | What AI can and can’t do in frontier mathematics today | | 48:01–49:36 | AI’s effect on mathematician productivity (auxiliary tasks vs. deep breakthroughs) | | 49:36–51:23 | Tao on artificial cleverness vs. true intelligence | | 53:39–56:25 | Could AI prove deep conjectures with no insight? The brute force hypothetical | | 59:34–62:35 | Desire for a formal language of strategies, not just proofs | | 70:06–73:06 | Tao’s approach to learning new math fields; value of obsession, collaboration, documentation | | 73:28–77:04 | The role of serendipity and “inefficiency” in scientific creativity | | 81:28–83:40 | Career advice in an age of AI transformation |
Flow and Tone
The dialogue is reflective, layered with historical context, philosophical musings, and the sober optimism of a world-class mathematician. Dwarkesh intersperses analogies and contemporary parallels to AI, while Terence brings nuance, drawing from deep technical knowledge, historical perspective, and his own current experience working with AI tools. The conversation is engaging, candid, and intentionally speculative, aiming to understand both where science has been and where it is heading.
Conclusion
This episode offers an intellectually rich exploration of the nature of discovery—past, present, and future—arguing that while AI has dramatically shifted the landscape of mathematics and science, core questions of insight, validation, communication, and serendipity remain deeply human. The conversation closes on an encouraging, if cautionary, note for young scientists: embrace change, nurture adaptability, and never underestimate the value of curiosity and accident in genuine progress.