“Terence Tao on the cosmic distance ladder” — 3Blue1Brown

Episode summary

Part 1 of a two-part Tao + Sanderson collaboration on how humanity first measured cosmic distances — from the radius of the Earth all the way out to the orbits of the planets. The 28-minute video walks the lower rungs of the distance ladder: Eratosthenes’ shadow-and-sundial measurement of Earth’s circumference (Alexandria vs. Syene, ~10% accurate with a graduate student pacing a road), Aristotle’s lunar-eclipse argument that Earth is round, Aristarchus’ eclipse-geometry measurements of the Moon’s distance (~60 Earth radii, essentially correct) and size, his bold-but-wrong measurement of the Sun’s distance (off by an order of magnitude due to inability to clock half-Moon to the half-hour, leading him to a heliocentric model 1,800 years before Copernicus), and Kepler’s “step of pure genius” — using Tycho Brahe’s stolen multi-decade Mars-position data sampled at 687-day intervals to triangulate Earth’s orbit shape relative to fixed Mars locations, and discovering that Earth’s orbit is an ellipse, not a circle. Tao’s load-bearing meta-frame: “If you want to measure the distance to x, you can never just look at x. You have to look at y and how x impacts y.” Closes on the cliff-hanger that Kepler had the shapes of all orbits but no absolute distances — “they could draw the exact picture, but they didn’t know the size of the paper” — setting up part 2’s measurement of the astronomical unit. Companion to 2026-04-20-3blue1brown-tao-cosmological-measurements.

Key arguments / segments

• [0:00] Tao introduces the project — astronomy was a childhood love; cosmic distances are a topic where the visual story matters as much as the numbers. The implicit critique of pop-science: too much “the distance to Mars is 18 million miles,” too little “and here’s how we know”
• [1:02] Tao’s load-bearing principle: “If you want to measure the distance to x, you can never just look at x. You have to look at y and how x impacts y.” Each rung uses the previous rung as the reference frame
• [2:02] Rung 1 — Earth’s radius (Eratosthenes, ~240 BC). Aristotle’s lunar-eclipse argument first established Earth is a sphere (the shadow on the Moon is always a circular arc). Eratosthenes: well in Syene reflects the Sun on the summer solstice; vertical gnomon in Alexandria casts a 7° shadow at the same moment; ratio 7°/360° = (Alexandria-Syene distance)/(circumference) → ~10% accurate
• [4:00] Tao’s p-hacking aside: “You can find sources online that claim his estimate was more accurate than this, but if you selectively choose which conversions to use, you can kind of p-hack your way into a better number.” Tao gives Eratosthenes credit for ~10%, not 1%
• [8:00] Rung 2 — Distance to the Moon (Aristarchus, ~270 BC). Lunar-eclipse traverse time (~3.5 hours) vs. lunar month (~28 days) gives Earth-shadow-radius / Moon-orbit-radius ratio. Aristarchus measured ~60 Earth radii. Actual: 58-62. Essentially correct, with no telescopes
• [10:00] Rung 3 — Size of the Moon. Watch a full Moon rise — about 2 minutes from limb-touch to fully visible. The 24-hour-day rotation scans your line of sight over the Moon. 2 minutes / 24 hours = (Moon radius / Moon distance). With distance known, size follows
• [11:01] Rung 4 — Sun (Aristarchus, attempt). The solar-eclipse coincidence: Moon and Sun look the same angular size, so (Moon radius / Moon distance) = (Sun radius / Sun distance). One ratio fixes both Sun unknowns if you can pin down either size or distance. Method: half-Moons do NOT occur halfway between New and Full Moon; the timing offset depends on the Sun’s distance via trigonometry. distance-to-Sun = distance-to-Moon / sin(offset-angle)
• [15:00] Where the Greeks hit a wall. Aristarchus thought the half-Moon was 6 hours off; actual is half an hour. He had no clocks (sundials don’t work in the dark) and no telescopes. The math was sound; the technology wasn’t. He concluded Sun was 20× Moon distance and 7× Earth size; true values are ~370× and 109×. Qualitatively right that the Sun is much bigger than Earth — which led him to propose the heliocentric model ~1,800 years before Copernicus
• [17:00] The Greeks’ rejection of heliocentrism was for sound reasons, not bad ones: they didn’t observe stellar parallax (which they should have, if Earth was orbiting). Their reasoning: stars must be thousands of times farther than current estimates — and that would be absurd. Tao: “Even when you have the math right, you don’t necessarily get to the truth, because the universe is in fact not just thousands of times larger, but actually billions and trillions of times larger.” A clean “right method, wrong scale” failure mode
• [18:00] Rung 5 — Kepler’s “step of pure genius” (early 1600s). Copernicus (1543) had circular heliocentric orbits and accurate orbital periods (Earth: 1 year, Mars: 687 days). Kepler wanted relative sizes of all orbits. Pet theory: nested platonic solids. He needed data — Tycho Brahe had decades of observations on Mars and refused to share. Kepler stole the data
• [21:00] Brahe didn’t record 3D positions — only where each planet appeared in the sky on a given date (constellation + angle). From sequences of angles only (no distances), Kepler had to deduce the shapes of all orbits including Earth’s. Two angles, two unknown orbits — looks unsolvable
• [23:00] The trick: solve a simpler problem first. If Mars were nailed to a fixed point in space, the Earth-Sun and Earth-Mars angles would be enough to triangulate Earth’s position relative to those two fixed points. Many nights of data → Earth’s orbit
• [24:00] The genius step: Mars returns to where it was every 687 days (Martian year). Sample Brahe’s data at 687-day intervals → Mars IS effectively a fixed reference in that subsampled time series. Brahe observed for 10 years — exactly enough for five samples. Then take a different starting day; Mars only moved a little, so the new five-point set is conditioned on a slightly-shifted Mars location. Adjacent five-point sets must fit together coherently — a massive jigsaw puzzle
• [26:00] Kepler pieced the jigsaw into a coherent Earth orbit alongside a coherent Mars orbit — and saw what no one had: Earth’s orbit is an ellipse, not a circle. Equal areas in equal times followed. With Earth’s orbit shape known, every other planet’s orbit is a follow-on triangulation. Einstein called this “an idea of pure genius”
• [27:00] The cliff-hanger: Kepler’s method gives only relative shapes, no absolute distances. “They could draw the exact picture, but they didn’t know the size of the paper.” Part 2 picks up at the hunt for any absolute distance to lock everything in place

Notable claims

• Tao’s load-bearing principle: “If you want to measure the distance to x, you can never just look at x. You have to look at y and how x impacts y.” (Stated explicitly at [01:02].) This is the canonical statement of the indirect-inference move that defines all of observational astronomy — and arguably all measurement that pushes past direct-instrument range
• Eratosthenes’ Earth-circumference measurement was ~10% accurate, not the often-claimed 1% (claims of higher accuracy are p-hacked via selective stadia conversions)
• Aristarchus measured the Moon’s distance at 60 Earth radii in the 3rd century BC. Actual orbit varies between 58 and 62. Essentially correct, with no telescopes, no clocks
• The Greeks rejected heliocentrism because they didn’t observe stellar parallax — and they were right to use that data point. Their failure was underestimating the scale, not flawed reasoning. Tao’s frame: “even when you have the math right, you don’t necessarily get to the truth”
• Kepler stole Tycho Brahe’s data after Brahe refused to share. The platonic-solids theory Kepler wanted to validate was off by a couple percent and couldn’t be made to fit; Copernicus’ circular orbits also couldn’t be made to fit. Throwing out the circular-orbit assumption (forced by data) gave him the ellipse
• Kepler’s 687-day-sampling jigsaw is the canonical example of solving a simpler problem first by exploiting a periodicity in the data — then assembling many simpler solutions into the full one. Each piece of the puzzle is a constraint conditioned on a hypothesized Mars location; coherence between adjacent pieces (one day apart, Mars slightly moved) is what selects the unique consistent orbit pair

Why this is in the vault

This is the founding-text source for measurement-as-indirect-inference — Tao’s own pithy framing (“never look at x, look at y and how x impacts y”) is the cleanest one-line statement the vault holds for the entire epistemological move that defines observational science, telemetry-based engineering, and (by extension) every “we can’t observe the model directly, so we instrument its behavior” pattern in AI systems. The Eratosthenes shadow-trick is the canonical physical instance; the Kepler 687-day-sampling jigsaw is the canonical algorithmic instance — exploit a hidden periodicity to convert an underdetermined problem into a sequence of determined ones. Second reason: the Aristarchus failure mode (“right math, wrong scale”) is the most concrete historical case-study the vault holds of how good methodology can be invalidated by background-assumption error — a frame that maps directly to AI eval failures where the right metric is computed at the wrong distribution scale (e.g., success-rate-on-curated-test vs success-rate-in-production, or hallucination-rate-at-context-length-N vs at-length-10N). Third: Tao’s “p-hacking your way to a better Eratosthenes” is the cleanest in-vault statement of how historiography itself can hallucinate accuracy — and a reminder that “ancient was smarter than you think” can be a confabulation as easily as a celebration.

Mapping against Ray Data Co

• Indirect-inference-as-measurement-discipline — candidate concept (CA-024). Tao’s “never look at x, look at y and how x impacts y” is general enough and recurring enough across the vault (think: Welch Labs probing diffusion models by removing the noise step, Practical Engineering’s “engineer’s perspective” series probing failures by reasoning from observable wreckage backward) to deserve its own concept-candidate slot. Pair this Tao video with 2026-04-20-3blue1brown-tao-cosmological-measurements (which extends the principle to red-shift, parallax, Cepheids, gravitational-wave standard sirens — five more rungs of the same move) and you have a paired-explainer ripe for a concept page. Mapping strength: strong / new candidate concept. Add to CANDIDATES.md as CA-024.
• Kepler 687-day jigsaw — algorithmic exemplar for “exploit hidden periodicity to make underdetermined problems determined.” Direct mapping to RDCO data-engineering: every time-series problem with an underlying period (revenue cycles, user-session cohorts, retention curves) where you can isolate one factor by sampling at the period and assembling many such single-factor solutions. The trick is recognizable in modern guise as “pivot wide on the cyclical dimension first, then triangulate the residual.” Mapping strength: medium / operational-reuse
• Aristarchus failure-mode — frame for AI eval failures at wrong scale. Right metric, wrong distribution → wrong-by-an-order-of-magnitude. The exact failure mode of “we benchmarked against the curated test set and shipped to production where the input distribution is 1000× weirder.” This is a more memorable historical-anchor for the eval-distribution-mismatch problem than the usual “data drift” framing. Newsletter angle: “Aristarchus Was Right About the Sun and Off by an Order of Magnitude on the Same Day” — about how the same can be true of your AI eval. Mapping strength: medium / newsletter-hook
• The “p-hacking historiography” aside — frame for over-curated source-quality narratives. Tao’s casual debunking of the “Eratosthenes was within 1%” myth is a useful prompt for vault-hygiene: when a source claims a historical or anecdotal precision that’s surprisingly tight, look for the conversion-fudge. Maps to internal RDCO discipline: when a benchmark or case study sounds remarkable, check whether it’s been arithmetically curated. Mapping strength: light / discipline-reminder
• Tao’s collaborative-breadth meta-pattern. Tao explicitly says he chose this topic because it would benefit from a strongly visual presentation — and chose Sanderson because Sanderson’s medium fits. The methodology lesson: the best collaborators aren’t the ones with the same tools, they’re the ones whose tools complement yours and whose subject is one your tools illuminate. Mapping strength: light / RDCO collaboration heuristic

Related

• 2026-04-20-3blue1brown-tao-cosmological-measurements — part 2 of this collaboration; extends the same principle (indirect inference rung-by-rung) from the solar system out to the largest observable structures (parallax for nearby stars, Cepheid standard candles, Hubble’s redshift law, gravitational-wave standard sirens). Natural paired-explainer — watch or cite them together
• 2026-04-20-3blue1brown-what-was-euclid-really-doing — Euclidean construction-as-subroutine as the ancient algorithmic-thinking exemplar; Tao’s Eratosthenes story is the matching ancient measurement exemplar
• 2026-04-20-3blue1brown-five-puzzles-thinking-outside-the-box — the canon-tier coordinate-change source (CA-023); Kepler’s jigsaw is structurally a coordinate-change move (sample at the Mars-period to make Mars a constant)
• concepts/CANDIDATES — propose new entry CA-024 (indirect-inference-as-measurement-discipline)