Terence Tao: Hardest Problems in Mathematics, Physics & the Future of AI | Lex Fridman Podcast #472
Executive Summary
Tao is drawn to problems on the edge of existing methods, where familiar techniques nearly work but leave an important gap. He separates reality, observations and models, discusses Lean as a strict proof-checking colleague, and sees room for software to expand collaborative experimental mathematics. His advice emphasizes transferable reasoning and trying something when stuck. His AI forecasts are views from the June 2025 discussion.
Key Takeaways
Look near the boundary of what methods can do
Tao distinguishes hard problems from problems that are hopeless with current tools. He finds interest in cases where existing methods do most of the work but leave a remaining gap. This is a way to choose useful research questions, not a claim that the final gap is easy or that any particular percentage can be measured precisely.
Keep reality, observations and models distinct
He describes actual reality, what we observe and the models we build as different things. They can move closer over time, but this often requires discarding an initial intuition. The distinction encourages checking which part of an argument is observed and which part comes from the explanatory model rather than treating a familiar picture as reality itself.
A correct argument should also be usable by others
Tao says writing that is correct is not enough if a result is meant to be read and built on. He draws an analogy with code that works for one task but is hard to reuse. Clear motivation and adaptable structure matter because mathematics is also a collaborative practice, not only an answer handed in for a mark.
Lean exposes details informal reasoning can leave implicit
He compares Lean with an extremely pedantic colleague who asks what objects mean and how steps are justified. The software can automate some tasks, but explicit types and definitions remain important. His account explains why translating a mathematical argument into checked code can reveal work hidden by informal notation.
Software could expand experimental collaboration
Tao sees Lean and collaboration tools as a possible way to scale experimental mathematics. He contrasts that with small bespoke programs where a bug can undermine an entire computation. This is his expectation for better shared work, not a statement that collaboration removes all error or makes every conjecture accessible.
Promising proof strategies need judgment
He uses a metaphorical mathematical sense of smell to describe recognizing whether a strategy is likely to help. Splitting a problem into two tasks can make it harder rather than easier. His comparison with game-playing AI is an intuition about a research challenge, not proof that an AI system has the full judgment of a mathematician.
Build skills that travel when tools change
In his advice to young people, Tao emphasizes adaptability, abstract reasoning and solving problems when something goes wrong. A single language or narrow technique may not transfer as well. He also urges students to try an approach when stuck because a failed attempt can reveal which assumptions matter.
The Conversation Breakdown
Trying a first approach
⏱ 3:06:17Tao describes the paralysis students can feel when familiar techniques do not obviously apply. He encourages attempting even a simple or unlikely approach and studying how it fails. A failure may point to a hypothesis that has not yet been used, turning a stalled problem into a question with a next step.
Core Frameworks & Ideas
Boundary of existing methods
Choose a question where familiar tools reveal useful structure and an identifiable unresolved gap.
Reality, observation and model
Distinguish the world, the evidence collected from it and the explanation built around that evidence.
Learning from a failed attempt
Try an approach, inspect the specific failure and use it to decide what assumption or method to examine next.
Actionable Next Steps
- ✓When stuck, write down a simple approach and inspect why it fails.
- ✓Identify which statements are observations and which come from a model.
- ✓Make an explanation understandable and reusable rather than only technically correct.
- ✓Practice reasoning and problem solving alongside learning a particular tool.
Notable Quotes
“What's really interesting are the problems just on the boundary between what we can do rather easily and what are hopeless”
“the process of getting closer often means that you have to discard your initial intuitions.”
“the one thing that actually you have to teach undergraduates in mathematics is that you should always try something.”
The One Thing to Remember
When a problem leaves you stuck, try something simple and learn from the way it fails.