Terence Tao: Nobody Understands Why AI Actually Works

Scientific concepts, discoveries, and nature/mathematical phenomena

Statistical patterns and “detecting nature vs human-made data”

• Benford’s law: In many naturally occurring datasets, the leading digit is disproportionately likely to be 1 (about 30%).
• Human-generated/artificial data often fails to follow Benford’s law and may look more uniform (e.g., from “fudging accounting books”).
• Practical implication: Benford-type digit distributions can help distinguish natural processes from manipulated numerical data.

Limits of mathematical induction

Mathematical induction (domino principle):

1. Prove the statement for 1
2. Prove “if true for n, then true for n + 1”
3. Conclude the statement holds for all natural numbers

Core theme: Induction-type reasoning only works under strict hypotheses—i.e., no “domino” can fail.

Minimal surfaces and the Simons cone

• Minimal surface idea (soap-film interpretation): a surface spanning a loop that locally minimizes area (analogous to surface tension).
• Dimension-dependent behavior:
  • In dimensions below 8, minimal surfaces are smooth (no knots/singularities in the soap-film analogy).
  • Starting at 8, singularities can occur.
• Simons cone:
  • An explicit minimal “cone-like” example in higher dimensions where a singularity can persist.
  • In contrast to lower dimensions—where a cone tip can often be modified (“surgery”) to reduce area—in higher dimensions the analogous smoothing fails, indicating a breakdown of naive low-dimensional intuition.

High-dimensional geometry and “geometry of data”

• Space-filling collapses in high dimensions:
  • A circle in a square occupies a large fraction of area; a ball in a cube is still sizable.
  • But in very high dimensions, the inscribed ball occupies an extremely tiny fraction of the cube’s volume.
• Implication for data: error models depend strongly on whether one uses “ball-like” (average/RMS) versus “cube-like” (worst-case/max) error regions.

Proof techniques and cross-area connections

• Proof by contradiction:
  • Assume the conclusion is false, then derive a logical inconsistency.
  • Used in examples like the “no largest natural number” argument (e.g., “add 1”).
• “Unexpected connections” as a valued proof style:
  • Linking discrete and continuous math
  • Linking physics intuition and combinatorics
  • Example cited: classical irrationality proof of √2 (Euclid / Pythagoreans mentioned)

Number theory: primes, randomness, and the twin prime conjecture

• Primes and the fundamental theorem of arithmetic (unique prime factorization).
• Twin prime conjecture:
  • Twin primes are prime pairs differing by 2 (e.g., 11 and 13).
  • They’ve been verified for enormous ranges, but it’s not proven they are infinite.
• Pseudo-randomness of primes:
  • While primes have some predictable features (e.g., “mostly odd”), they are believed to behave statistically like random sets once obvious biases are removed.
  • Heuristics based on this randomness motivate the belief that twin primes should occur infinitely often.
• Cryptographic relevance:
  • Many cryptosystems rely on primes and related constructions having no exploitable pattern.
  • Finding hidden structure in primes (or other mathematical objects) could weaken cryptography.

Quantum computing and its interface with mathematics

• Why quantum computers can matter for math:
  • Quantum states are modeled by wavefunctions
  • Dynamics are built from linear operations (matrix operations) plus time-reversible constraints
• Exponential speedups in principle for tasks like:
  • Factoring (classic application)
  • Quantum chemistry (also mentioned)
• Computational complexity as a driver of new math:
  • Not just “is it true/false,” but how efficiently it can be computed (polynomial vs exponential time)
  • Different proofs can reflect different algorithmic efficiencies, refining notions of “strength.”

AI for mathematics: verification, hallucinations, and new conjectures

• LLMs and mathematical reasoning:
  • Can produce plausible reasoning but may be ungrounded, producing errors (“hallucinations”).
• Mitigation approach:
  • Use LLMs as suggestion engines within a rigorous verification system.
  • Prefer outputs in formal languages that can be checked by proof verifiers.
• “Contamination”:
  • Sometimes models repeat known textbook proofs rather than discovering new ones.
• Neural-net example domain (knot theory):
  • A neural network learned correlations between knot invariants (not previously believed).
  • Researchers then perturbed inputs to identify which invariants actually matter, producing a formal prediction that was later provable.
• AI “new prediction” standard:
  • A proposed “Keating test” (Turing-test-like) for math/physics:
    • AI should generate new, previously unknown, verifiable predictions, not just recalled results.

AI performance as an unresolved theoretical problem

• Mathematical gap:
  • We have strong theory for:
    • fully random data (probability)
    • highly structured periodic data
  • But LLM training data is partially structured (intermediate), where theory is less developed.
• Analogy to physics:
  • Good theories exist for continuum (averaged) and for atomic/micro-level behavior.
  • Less is known in the mesoscale/emergent regime (e.g., cells), which is harder to model.

Gauge theory as a unifying principle (physics-to-math “currency conversion” metaphor)

• Gauge theory:
  • Describes how descriptions change when you alter units/frames/measurement conventions.
  • Some properties are gauge-invariant (physical), while others are gauge-dependent.
• Curvature-like idea:
  • Transport quantities around loops to get a nontrivial invariant (analogous to curvature), even though local unit choices differ.
• Metaphor: converting money between currencies is like changing gauge/unit conventions; a loop conversion might not return exactly due to holonomy-like effects.

Physics limits, models, and epistemology

Physics models must be:

• Mathematical, sometimes provable within their own formal structure
• Yet still depend on how they interface with reality, including potential breakdown at extremely small scales or at the origins of the universe.

Compressed sensing and medical imaging (MRI acceleration)

• Compressed sensing:
  • Reconstructs signals/images from far fewer measurements than traditional approaches.
  • Applied to MRI: enables scans up to about 10× faster (claim).
• Underlying math angle:
  • Reformulates imaging as solving linear systems, using signal structure/sparsity assumptions.

---

Methods / methodologies mentioned (bullet outline)

Reasoning and proof workflow (mathematical induction)

• Base case: prove statement for n = 1
• Inductive step: prove P(n) ⇒ P(n + 1)
• Conclude P(n) for all natural numbers (assuming no “failed domino”)

AI-assisted mathematical research workflow (as described)

• LLM proposes candidate ideas/approaches
• Humans verify/dismiss incorrect suggestions
• Use automated proof verification where possible
• Prefer formal, checkable outputs to reduce hallucinations
• Check for “contamination” (memorized known solutions) when claiming novelty

Extracting structure from a neural network “black box” (knot invariants example)

• Train a neural net to predict one invariant from many others
• Perturb/adjust inputs (“dials”) to test which features affect outputs
• Identify key inputs driving predictions
• Formulate a formal conjecture/prediction from observed sensitivity
• Prove it mathematically

Compressed sensing (MRI) concept

• Acquire fewer measurements than standard imaging
• Reconstruct using mathematical structure (implicitly: sparsity/low-complexity assumptions)
• Reduce scan time

---

Researchers/sources featured (named or explicitly referenced)

• Terence Tao (host/interview subject)
• Jim Simons (philanthropist and researcher; work referenced on minimal surfaces / Simons cone)
• Edward Witten (named as the only physicist to win the Fields Medal)
• Steve Strogatz
• Wigner (quote: “unreasonable effectiveness of mathematics” referenced)
• Hardy (quoted about mathematicians offering “the entire game”)
• Euclid (irrationality/prime ideas referenced; proof lineage mentioned)
• Pythagoreans (mentioned as earlier source for √2 irrationality)
• Galileo Galilei (referenced; also the “compass geometry” book)
• Evan Watson (Tao’s student mentioned)
• Sergio Kleinerman (asked a question referenced)
• John Preskill (Caltech; quantum computing aside mentioned)
• Keating (humorous name for the “Keating test” concept; itself references “Turing test”)
• Shannon (information theory / Shannon communication limits mentioned)
• JPL (Jet Propulsion Laboratory; mentioned in context of data on Mercury’s orbit)
• Fermat (Fermat’s Last Theorem referenced)
• Navier–Stokes (referenced)
• Fields Medal / UCLA / Trump administration (contextual references; not researchers)
• Eric Weinstein