Terence Tao – Kepler, Newton, and the true nature of mathematical discovery

The episode opens with a promise rather than a thesis: the discovery of the solar system's shape is the right test case for what AI will do to science. The teaser names the reason. Tight verification loops are supposed to be where AI makes fastest progress, and the story of how we worked out the orbits shows how long a verification loop can actually run — decades, sometimes millennia.

The claim is stronger than a delay. Across that interval the theory we now know to be better often makes worse predictions — Copernicus's circular orbits fit the observations less accurately than Ptolemy's geocentric model did. What kept heliocentrism alive through its bad decade was some mixture of judgement and heuristics that nobody has articulated well enough to codify, let alone to put into a reinforcement-learning loop.

Terence Tao takes the story from the top, with the caveat that his interest in it is amateur. Kepler was building on Copernicus, who was building on Aristarchus. Copernicus's proposal is the famous inversion: the Sun at the centre, the planets going around it rather than around the Earth.

His orbits were perfect circles, and the theory fit observations that Greek, Arab and Indian astronomers had accumulated over centuries. What Kepler added, as a student, was a reading of the geometry: the ratios of the orbit sizes Copernicus predicted seemed to carry geometric meaning.

The first theory is the one that makes the rest of the episode legible. Enclose the Earth's orbit in a cube, and the sphere that circumscribes the cube almost exactly matches the orbit of Mars; continue outward and the nesting continues. Six planets were known, so there were five gaps between them, and there are exactly five Platonic solids — cube, tetrahedron, icosahedron, octahedron, dodecahedron.

Kepler found this beautiful and, more importantly, found it fitting. The design of the planets seemed to match a mathematical perfection, which is a hypothesis about God rather than about Mars, and it needed data.

There was exactly one. Tycho Brahe, a wealthy and eccentric Danish astronomer, talked the Danish government into funding an observatory so expensive it was an entire island, and spent decades observing every planet on every clear night — with the naked eye. He was the last of the naked-eye astronomers, and his dataset was the only high-quality one in the world.

Tycho was jealous of it and released it in fragments. Kepler eventually copied it, and had to fight Brahe's descendants for the right to keep it. The episode treats the theft lightly, but it is the hinge of the story: the theory was about to meet the only instrument capable of judging it.

The beautiful theory failed. The data was off by something like ten percent, and Kepler spent years moving circles around, trying every fudge he could think of, before working out how to use the data to recover the actual orbits — an act of data analysis that Tao calls an incredibly clever, genius amount of work.

The answer was ellipses, which shocked him, plus the equal-areas law. Ten years later, after wrestling with the outer planets — Saturn and Jupiter were the hardest — he got the third law: the period scales as some power of the distance to the Sun. He had no explanation for any of the three. It was all driven by experiment, and it took Newton, a century later, to produce a theory that explained all three at once.

Dwarkesh offers the episode's central provocation. Kepler was a high-temperature LLM; Newton is the explanation of why the three laws must be true. The discovery of the relative orbits was a work of genius, but across an entire career Kepler is mostly trying random relationships.

The evidence is the book. The third law appears as an aside in *The Harmonics of the World*, a treatise on how the planets have different harmonies, and on why there is so much famine and misery on Earth — because the Earth is mi-fa-mi, that being the note of Earth. It is astrology, and inside it is the cube-square law relating period to distance. Add Newton's F=ma and the expression for centripetal acceleration and the inverse-square law falls out.

The point of the analogy is not that Kepler was stupid but that the search procedure is reproducible. An LLM can try random relationships for twenty years — musical notes, Platonic objects, arbitrary geometries, all held together by a bias that the orbits have some important geometry — provided there is a verifiable data bank like Brahe's to check against. Then one thing works, and an empirical regularity drives real progress.

Tao accepts the analogy and then widens it. The history of science has always awarded prestige to idea generation, but a scientific problem is a dozen steps: identifying a problem, identifying a fruitful one, collecting data, choosing an analysis strategy, forming a hypothesis, validating it, writing it up. What we celebrate are the eureka moments, and Kepler certainly cycled through ideas, several of which failed and many of which he probably never published.

But the generation has to be matched by verification or it is slop. We celebrate Kepler; we should celebrate Brahe, whose data was ten times more precise than any previous observation, and that extra decimal place was what made the result possible. Kepler then used Euclidean geometry — the most advanced mathematics available to him — to match models to data. Data, theory and hypothesis generation all had to be in play.

Tao is not sure hypothesis generation is the bottleneck any more, and offers a short history of paradigms: theory and experiment classically, numerical simulation in the twentieth century, big data and data analysis at its end.

The consequence is an inversion. Progress now often begins with a massive dataset, from which patterns are drawn and hypotheses deduced — the reverse of the classic method, in which a few observations or one out-of-the-blue idea came first and data was collected to test it. Kepler is arguably an early data scientist, but even he did not start from Tycho's dataset; he arrived with preconceived theories. That is less and less how progress is made, because the data is so much larger and more useful.

Dwarkesh pushes back with the chronology. Kepler's ideas date from 1595 and 1596 — the polygons, then the Platonic solids — and they were wrong. Brahe's data arrives years later, and it is twenty years of trying random things before the empirical regularity appears. The dataset functions like a massive bank of simulations: without it, Kepler writes books about harmonics and there is nothing to verify against.

Tao grants the importance of the data and sharpens his own distinction. Traditionally a hypothesis is formed and then tested against data; with machine learning and statistics, you can begin from data and derive laws that were not present before. The disagreement is narrower than it looks — it is about which end of the pipeline the new capacity sits at.