Summary of "Terry Tao - Machine assistance and the future of research mathematics - IPAM at UCLA"

Overview

Terry Tao argues that recent advances in AI combined with formal verification are beginning to change how research mathematics is done — enabling scale, broader participation, and new workflows — after a long period in which mathematics has been relatively conservative in practice. He uses the “Erdős problems” (a ~1,000 problem collection curated by Thomas Bloom) as a concrete dataset to show how AI + human collaboration + verification can sweep many medium-difficulty problems, reveal tool strengths and weaknesses, and produce formally checked results.


Main ideas, concepts, and lessons


Workflows and methodology

Creating a productive large-scale math project

Using AI + humans + verification to solve problems (step-by-step)

  1. Survey: use AI-enabled semantic literature search to find relevant results (including overlooked or non-English papers).
  2. Experimentation: generate code with AI, run numerical experiments, and visualize patterns quickly.
  3. Conjecture formation: humans synthesize numerical/AI suggestions into conjectures and iteratively simplify them.
  4. AI-assisted solution generation: feed conjectures into AI tools to produce sketches or candidate proofs (or transform problems).
  5. Human vetting: experts inspect and clean AI output, translating it into a form suitable for formalization.
  6. Formalization: use a proof assistant / formalizer (e.g., Aristotle + Lean) to convert and check the proof; iterate until it compiles.
  7. Publication/record: record solved problems, mark formalized solutions, and update the dataset and metrics.

Moderation and quality control

Measuring progress


Case studies and concrete examples

The Erdős problems data set

Literature-discovery wins

Problem 367

Problems 126 and 1024 (coin-splitting / packing examples)

Productivity gains


Tools, terms, and resources


Takeaway conclusions


Speakers and sources mentioned

Category ?

Educational


Share this summary


Is the summary off?

If you think the summary is inaccurate, you can reprocess it with the latest model.

Video