Terence Tao Discusses AI's Role in Mathematics, the Mystery of Prime Numbers, Proof Techniques, and High-Dimensional Geometry

Dr Brian Keating

Passwords, primes, and why randomness protects your digital life [00:00:00]

• Digital security, including passwords and encrypted messages, relies on the assumption that prime numbers behave randomly and unpredictably, without exploitable patterns.
• Despite extensive supercomputer testing on trillions of cases, a formal mathematical proof for this inherent randomness remains elusive.
• Terence Tao emphasizes that an undiscovered pattern in prime numbers could potentially break existing encryption systems.

Coffee, Erdős, and the inside jokes [00:00:58]

• Paul Erdős famously stated, "Mathematicians are a means for turning coffee into theorems," highlighting the intense dedication to mathematical discovery.

Tao meets Paul Erdős at age 10 [00:01:28]

• Terence Tao met the legendary mathematician Paul Erdős at the age of 10 in Adelaide, Australia.
• Erdős was known for nurturing young talent and treated Tao as an intellectual equal, later sending him a postcard with a math problem (which Tao did not solve at the time).

Erdős number and the Erdős–Bacon number [00:02:21]

• The "Erdős number" measures the collaborative distance to Paul Erdős through co-authored papers, with Erdős having a number of zero.
  • Tao has an Erdős number of two.
• The "Erdős–Bacon number" combines a person's Erdős number with their "Kevin Bacon number" (co-starring with Kevin Bacon in films), usually resulting in an infinite sum due to the unlikelihood of overlap, though a few individuals have finite, small sums.

  

  

Erdős, amphetamines, and productivity lore [00:03:19]

• Paul Erdős was exceptionally prolific, reportedly aided by amphetamines, a practice less common among mathematicians today due to differing views on work-life balance.

Tao explains the Erdős discrepancy problem [00:04:07]

• Discrepancy theory studies the irregularity of sequences, such as +1s and -1s.
• The Erdős discrepancy problem asks whether a sequence of +1s and -1s, no matter how carefully constructed to appear balanced, must eventually exhibit an unbounded discrepancy over arithmetic progressions.
• Tao proved that the discrepancy must indeed diverge, albeit very slowly (logarithmically or double logarithmically), using tools from information theory and number theory.

Randomness, human bias, and cheating detection [00:06:44]

• The difficulty humans have in generating truly random patterns is connected to statistical anomalies like Benford's Law.
• Benford's Law states that roughly 30% of numbers in naturally occurring datasets begin with the digit "1," a pattern often violated in fabricated data, which tends to be too uniformly distributed. This can be used to detect fraud or cheating.

Induction pitfalls, minimal surfaces, and dimension surprises [00:08:14]

• Mathematical induction is a rigorous proof technique to establish a statement for all natural numbers, by proving a base case and an inductive step (if true for n, then true for n+1).
• This is compared to a row of dominoes, where knocking over the first domino and ensuring each domino knocks over the next guarantees all dominoes fall.
• Jim Simons' work on minimal surfaces revealed that while these surfaces (like soap films) are smooth in lower dimensions (up to 7), singularities can form starting in 8 dimensions (Simons' cone), challenging the expectation that patterns always continue infinitely.
• High-dimensional geometry often defies low-dimensional intuition; for example, a 1000-dimensional ball inscribed in a 1000-dimensional cube occupies an extremely tiny fraction of the cube's volume, unlike in 2D or 3D. This has implications for data science and error measurements.

Proof styles and why contradiction is powerful [00:13:45]

• Tao's favorite proof technique is proof by contradiction, where one assumes the opposite of what is to be proven and shows it leads to a logical inconsistency.
• An intuitive example is proving there is no largest natural number: if there were one, adding one to it would yield a larger number, a contradiction.
• Tao particularly appreciates proofs that forge unexpected connections between different mathematical areas (e.g., discrete and continuous math) or leverage physical models for mathematical problems.

Why square roots and i show up everywhere in physics [00:16:42]

• The prevalence of square roots and imaginary numbers (like i = √-1) in physics, especially in quantum mechanics, is due to the desire for "algebraic completeness."
• Real numbers, while useful, have a flaw: not all polynomial equations (like x² + 1 = 0) have roots within them.
• Complex numbers, which are a two-dimensional extension of real numbers, provide this completeness, allowing all polynomial equations to have roots. The "twice as big" nature of complex numbers naturally calls for the inclusion of a square root not previously available.

Transcendentals and why 1 stopped being prime [00:21:01]

• Mathematical definitions evolve to better serve the overarching structure and theorems of the field.
• The number one was historically considered prime but was redefined to not be prime to preserve the "Fundamental Theorem of Arithmetic," which states that every integer greater than one can be uniquely factored into primes. If one were prime, 12 could be 2x2x3 or 1x2x2x3, violating uniqueness.

Twin primes and what we still can’t prove [00:23:05]

• The "twin prime conjecture" posits that there are infinitely many pairs of prime numbers that differ by two (e.g., 11 and 13).
• Despite computational evidence of trillions of such pairs, a formal proof remains elusive, highlighting the enduring challenges in number theory.
• Primes are considered "pseudorandom" – deterministic but statistically indistinguishable from random noise, apart from obvious patterns (like being odd, except for 2).
• This pseudorandomness is fundamental to modern cryptography, as the security of many systems relies on primes having no exploitable patterns. Unexpected patterns could expose vulnerabilities.

  

Quantum computers: powerful and restricted [00:27:00]

• Quantum computers offer exponential speed-ups for specific problems like factoring large numbers (which could break current encryption) and quantum chemistry simulations.
• However, they are also exponentially more limited, primarily to linear and time-reversible operations, posing significant software and error correction challenges.
• While they may enable new large-scale computations, their direct impact on the theoretical foundations of mathematics is yet to be seen.

Complexity theory: truth vs computability [00:29:12]

• Classical complexity theory, born with the advent of computers, expanded mathematics beyond just "true or false" to "how computable" something is.
• It introduces finer-grained notions of truth, distinguishing between problems that are solvable in polynomial time versus exponential time, providing deeper insights into mathematical proofs and algorithms.

AI in math: strengths, hallucinations, and verification [00:31:11]

• AI, particularly Large Language Models (LLMs), is slowly beginning to aid mathematical discovery, though it has a fundamental weakness: a lack of inherent grounding and a propensity for "hallucinations."
• LLMs can generate plausible-looking solutions but often make errors, especially when venturing beyond their training data.
• Used as a component in a rigorous system, AI can make suggestions, help with literature review by recalling overlooked methods, and detect hidden correlations.
• In knot theory, neural networks have identified unexpected correlations between different types of knot invariants, leading to new mathematical predictions that could then be formally proven.

  

Why LLM mechanics are simple but performance prediction is hard [00:38:00]

• The underlying mathematics of LLMs (matrix multiplication, basic calculus) are not inherently complicated, making them accessible to undergraduate math majors.
• The true mystery lies in predicting their performance: why they excel at some tasks and fail at others is not well understood mathematically.
• This difficulty stems from the nature of "natural data" (like human language), which is partially structured—neither perfectly random nor perfectly predictable—a domain for which mathematics currently lacks robust theories.

Is math a language or something more? [01:04:00]

• Mathematics functions as a highly efficient language, honed over centuries to describe underlying mathematical structures.
• Unlike natural languages, which also convey emotion and nuance, mathematical language is optimized for precision and compactness, aiming to describe the universe in its most minimal and elegant form.
• This optimization for efficiency helps mathematicians uncover the fundamental essence of how the universe works, leading to a convergence between mathematical formalisms (like curved space) and physical laws (like Einstein's theory of spacetime).

Is math invented or discovered? Tao’s answer [01:02:21] [00:43:21]

• Terence Tao believes mathematics is both invented and discovered.
• Humans invent mathematical languages and frameworks, which initially may be imperfect.
• Over time, as these languages are refined for efficiency and power, they naturally converge towards an "ideal Platonic ideal" of mathematics, which feels like a discovery of innate, pre-existing structures.

Teaching in the AI era: verification and critique [00:44:18]

• Education must evolve in the AI era, especially concerning cognitive tasks previously exclusive to humans.
• Students need enhanced training in validating information from various sources of variable quality, including AI-generated content.
• Teachers are experimenting with assignments where students critique and correct AI-generated answers, fostering critical thinking over passive knowledge acquisition.

Can AI police itself? Reliability through verification workflows [00:47:12]

• While current AI models are inherently unreliable (prone to "hallucinations"), they can be valuable tools when paired with robust human verification workflows.
• Tao views AI as complementary to human scientists, particularly for tackling the "long tail" of millions of medium-difficulty problems that human researchers often lack the time for.
• AI's scalability allows it to solve a significant number of these problems, even with a moderate success rate, offering a massive advantage in overall progress.

Tao’s current focus: modernizing mathematics and funding [00:48:54]

• Tao's current research prioritizes new workflows to modernize mathematics, making it more collaborative, publicly accessible, and integrated with new tools like AI.
• He aims to transform mathematics from its centuries-old pen-and-paper tradition into a more open, interdisciplinary field, akin to data-driven sciences.
• Securing funding for basic mathematical research, which has become unstable, is another key focus.

Fame, ego, and why proofs keep you honest [00:50:20]

• Winning the Fields Medal in 2006 significantly changed Tao's public life, bringing invitations and requests for his opinion.
• However, the rigorous nature of mathematics provides a natural check on ego: proofs must be 100% correct and fully detailed; reputation alone cannot validate a claim.
• The vast number of unsolved problems constantly reminds mathematicians of the limits of their knowledge, fostering humility.

Do mathematicians peak at 30? Wisdom vs speed [00:52:04]

• The notion that mathematicians do their best work by age 30 is a stereotype, though some, like Tao, accelerate early.
• Other mathematicians start late. With age, mathematicians develop "wisdom" – an ability to draw on vast experience to identify known solutions or effective approaches quickly, often leading to different, but equally productive, ways of doing mathematics than youthful "brute force."

Should mathematicians learn physics? Intuition across fields [00:54:18]

• Mathematicians benefit greatly from knowing other fields, especially physics, which provides rich intuition and examples for concepts like differential equations or fluid dynamics.
• Cross-disciplinary thinking, whether in economics (e.g., thinking about trade-offs and conversions) or game theory (e.g., viewing epsilon-delta proofs as strategic games), offers valuable mindsets for problem-solving.

Galileo’s mathematical compass and computation before calculators [00:57:11]

• Galileo's "geometric and military compass" was an early calculating device, serving as a precursor to slide rules, capable of complex calculations like currency exchange.
• This historical example highlights early efforts to apply mathematical tools to practical problems, long before modern calculators or computers.

Currency exchange as a gauge theory metaphor [00:59:47]

• Currency exchange serves as an intuitive metaphor for "gauge theory" in physics.
• Gauge theory deals with quantities (like wealth or electromagnetic fields) that can be measured by numbers but whose units or frames of reference can change depending on location or choice.
• Conversions between currencies (or frames) are like "gauge transformations." The "curvature" in this metaphor is seen in exchange fees; converting dollars to euros and back results in a loss, showing a non-trivial path dependence, analogous to how fields behave in curved spacetime.

String theory: elegance, flexibility, and evidence [01:03:01]

• Physics continuously re-evaluates its conception of the universe, moving from Copernican to Einsteinian relativity and quantum mechanics.
• Modern physics faces inconsistencies at tiny scales, requiring new theories beyond the smooth manifold of spacetime.
• String theory, an elegant contender, has not yet yielded a unique, canonical theory that perfectly fits observable data, possibly due to its excessive flexibility.

Gödel vs physics: models you can prove vs worlds you test [01:04:33]

• In mathematics, Gödel's incompleteness theorems establish limits on what can be proven within a formal axiomatic system.
• Physics, in contrast, doesn't offer proofs in the same absolute sense. Physical theories are mathematical models that can be rigorously analyzed (e.g., proving properties of Einstein's equations), but their ultimate "truth" is provisional and subject to experimental falsification (as proposed by Popper).
• Physics and mathematics maintain their distinct epistemologies; physics tests models against reality, while mathematics proves statements within models. Older theories like Newtonian gravity remain useful models even if technically superseded by more accurate ones.

Compressed sensing: math breakthrough to faster MRIs [01:06:47]

• Terence Tao's work on "compressed sensing" exemplifies how fundamental mathematical research can lead to significant technological advancements.
• Collaborating with statisticians and engineers, he helped solve an image acquisition problem, enabling the reconstruction of high-quality images from significantly less data than traditional methods.
• This breakthrough has been adopted by major medical MRI machine manufacturers, speeding up MRI scans by a factor of 10.

  

  

Why basic math research pays off in engineering [01:09:25]

• Basic mathematical research, driven by curiosity, is crucial for applied sciences and engineering, even if its practical impact isn't immediately obvious.
• Early theoretical work, like Shannon's communication complexity theory developed decades before the digital revolution, provided foundational limits and principles (e.g., Shannon bound) that later enabled efficient digital communication systems (like cell phones).
• This foundational work, often done with "pen and paper," is far cheaper and more efficient than trying to solve complex engineering problems without theoretical guidance, saving billions in development costs by informing what is possible.