“Terence Tao continuing history’s cleverest cosmological measurements” — 3Blue1Brown

Episode summary

Part 2 of the Tao + Sanderson collaboration. Picks up where part 1 (filed as 2026-04-20-3blue1brown-tao-cosmic-distance-ladder) left off: Kepler had the shapes of all orbits but no absolute distances — “the picture but not the size of the paper.” This 25-minute video walks the upper rungs: the Venus-transit method for nailing the astronomical unit (with the tragicomic Le Gentil sub-story — declared dead, wife remarried, estate plundered while he tried to observe two transits 105 years apart), Roemer’s Io-eclipse timing measurement of the speed of light from observed Jupiter-eclipse offsets across the Earth’s orbit, stellar parallax as the next rung outward (Bessel 1838, only ~1,000 nearby stars within reach, ~1.5 arcsec for Proxima Centauri = a dime at 2.5 km), the Hertzsprung-Russell diagram as a way to deduce stellar absolute brightness from color (and thus distance from observed brightness via inverse-square), Henrietta Swan Leavitt’s Cepheid standard-candle law (period ↔ absolute brightness) extending the ladder to other galaxies, and Hubble’s redshift-distance law for the cosmological-scale rung. Closes on the live tension: gravitational-wave standard-siren measurements of Hubble’s constant cross-check the redshift method to within ~10%, but that 10% is a real and ongoing anomaly — the cosmological principle (laws are the same everywhere) is a faith-statement that has earned its keep but is now under pressure. Astronomy is a living subject. Tao is writing a book on the topic with collaborator Tanya Klowden.

Key arguments / segments

• [0:00] Recap from part 1 — Kepler had orbital shapes but no absolute scale. Astronomers were now hungry for any distance measurement that would lock the whole solar system in place
• [1:01] Venus transit method (the “lock the AU” rung). Two observers on widely-separated parts of Earth (Greenwich and the southern hemisphere) measure Venus’s apparent position against the Sun’s disc during a transit. Parallax angle is ~1 arcminute (1/60 of a degree). The Sun is 32 arcminutes wide. Each observer measures the duration of the transit (knowing how fast Venus and the Sun move through the sky), and from the two duration values you geometrically derive the Venus-Earth distance
• [3:00] Why duration not pure angle: clocks at the time were too poor for “make this measurement at exactly this moment” coordination across hemispheres. Duration is self-clocking — start the timer when Venus’s silhouette appears, stop when it leaves
• [6:00] The Le Gentil sub-story. Halley originally proposed the method but didn’t live long enough to see a transit. Le Gentil was dispatched for the 1761 transit but was delayed by the Seven Years’ War. Next transit 8 years later (1769); the one after that 105 years away. He extended his journey, set up in the Philippines, and on the day of the 1769 transit it was cloudy. Returned to France to find he’d been declared dead, his wife had remarried, and his relatives had plundered his estate. Cost-of-data-acquisition anecdote, vault-grade
• [7:01] One Venus-transit measurement was enough to fix the astronomical unit (AU) — the most important rung. Almost everything beyond the solar system is measured in AU. Roemer’s Io-eclipse method (the speed-of-light bonus rung). Io orbits Jupiter every 42 hours; you can time when Io enters/exits Jupiter’s shadow. Roemer noticed Io was 20 minutes ahead of schedule when Earth was on Jupiter’s side of the Sun and 20 minutes behind when on the opposite side. Realization: light takes 20 minutes to traverse 2 AU. Historically before the Venus-transit measurement, so the absolute speed-of-light estimates from this don’t look impressive — but at the time it wasn’t even obvious that light HAD a finite speed. Set the stage for later, more precise terrestrial speed-of-light experiments
• [10:00] Stellar parallax (next rung outward). Same trick as the Earth-baseline parallax for planets, but use the Earth’s orbit as the baseline — 2 AU between June and December. Time-lapse of Proxima Centauri over 6 months shows it visibly drifting against background stars. For Proxima — the closest star, at ~260,000 AU — the parallax shift is ~1.5 arcseconds. A dime at 2.5 km. First successful measurement: Friedrich Bessel, 1838
• [13:01] By the 19th century, parallax-distance was known for ~1,000 stars. Distance + apparent brightness = absolute brightness, via the inverse-square law of light propagation
• [15:01] Hertzsprung-Russell diagram (the “color tells you the absolute brightness” rung). Plot color (frequency, blue → red) on x-axis vs. absolute brightness on y-axis. Most stars cluster on the main sequence. Built from decades of data by the Harvard computers (a group of women at the Harvard College Observatory). Antonia Maury and Annie Cannon developed the spectral-line classification systems
• [16:01] The lever: for a distant star whose parallax is too small to measure directly, look at its color → infer absolute brightness from main-sequence position → combine with apparent brightness to get distance via inverse-square law. Spectral lines also reveal which atoms are present, classifying the star
• [18:00] Cepheid variables (the “extragalactic” rung). Past the Galaxy, individual main-sequence stars are too faint. But Cepheids — supergiant variable stars with brightness oscillating over 10-20 day periods — are bright enough to be seen in other galaxies. Henrietta Swan Leavitt measured all the Cepheids in our own galaxy and found a linear period-vs-intensity relation: brighter Cepheids have longer periods. This is the standard-candle move — observe a Cepheid in another galaxy, measure its period, infer absolute intensity, compare to apparent brightness, get the distance
• [20:00] Hubble’s redshift law (the cosmological rung). Edwin Hubble measured many galaxies. Spectral lines (e.g., hydrogen absorption) were redshifted, with the redshift proportional to distance — Hubble’s Law. We now understand this as a uniformly expanding universe (general relativity prediction): farther = receding faster = more redshift. Largest rung of the ladder
• [21:02] Today, virtually every light- or radiation-emitting object in the universe has its distance measured by spectral-line shift. Sloan Digital Sky Survey maps ~500,000 galaxies out to 1.6 billion light-years. Discovery: galaxies form filaments — massive structures that match gravitational-N-body simulations of billions of virtual galaxies
• [23:00] Cross-check: gravitational-wave standard sirens. Black-hole-merger events emit gravitational waves whose amplitude vs. frequency profile can be turned into an absolute distance, independent of redshift. Stack many rungs (sun-distance → parallax → main-sequence fitting → Cepheids → Hubble) and cross-check Hubble against the standard-siren distances. They match to within ~10%. The 10% is controversial. Something seems wonky about Hubble’s law at very large scales
• [24:00] The cosmological principle: “The laws of the universe are pretty much the same everywhere.” An article of faith since the Copernican revolution. Always rewarded with self-consistency and new laws. But the 9-10% Hubble-tension anomaly is an ongoing area of study. Tao’s closing: astronomy is a living subject

Notable claims

• The 9-10% Hubble tension is real and ongoing as of 2025 — gravitational-wave standard sirens disagree with redshift-Hubble measurements at this level. The cosmological principle that has guided astronomy since Copernicus is under empirical pressure
• The Le Gentil story — declared dead, wife remarried, estate plundered, all because he chased a Venus transit on a cloudy day — is the canonical anecdote for the cost of physical-data acquisition before instrumentation collapsed it. Vault-quote material
• Roemer’s speed-of-light measurement (~1676) predates the Venus-transit AU measurement by ~85 years — so the absolute speed-of-light value Roemer extracted was off, but he established that light has a finite speed, which was not obvious at the time
• Bessel’s 1838 stellar-parallax measurement was the first direct detection of any star’s distance — vindicating the heliocentric model 295 years after Copernicus published. The Greeks’ 3rd-century-BC objection to heliocentrism (no observed stellar parallax) was finally resolved by better instruments, not better math
• The Hertzsprung-Russell diagram (1911) was built from decades of data taken by the Harvard computers — a group of women at the Harvard College Observatory. Henrietta Swan Leavitt discovered the Cepheid period-luminosity law from this work. The diagram is the canonical example of “let’s plot two observable variables and see what structure emerges” — pre-dating modern unsupervised-clustering methods by a century
• Cosmic galaxy filaments were predicted by gravitational N-body simulations before being directly observed by surveys like SDSS. Simulation-then-observation as a pattern in modern astronomy
• Tao explicitly notes the cosmological principle “has been rewarded” with self-consistency and new physical laws — a working scientific faith-commitment, not a dogma. The 10% Hubble tension is the live edge case

Why this is in the vault

This is the second-half source for the measurement-as-indirect-inference candidate concept (CA-024, see 2026-04-20-3blue1brown-tao-cosmic-distance-ladder). It extends Tao’s “never look at x, look at y and how x impacts y” frame from the solar-system rungs (parallax, Venus transit, Roemer’s Io-eclipse) all the way out to galactic and cosmological scales (Cepheid standard candles, Hubble’s redshift law, gravitational-wave standard sirens). The five-additional-rungs-on-one-principle structure is exactly what makes a concept canonical — and the Hubble tension closing is the live-research detail that prevents the lesson from feeling like a dead history. Second reason: the Cepheid period-luminosity law is the canonical “find a hidden invariant in your dataset that lets you extrapolate beyond the calibration range” move — Henrietta Swan Leavitt found that brighter Cepheids in our galaxy have longer periods, then assumed (correctly) that the same law holds in other galaxies, then used periods of distant Cepheids to infer their absolute brightness and thus their distance. This is the canonical case study for observational standardization as a research move, and it maps directly to AI eval design (find an internal invariant in a benchmarked range, then extrapolate to validate measurements outside it). Third: the gravitational-wave-standard-siren cross-check is the canonical case study for independent-method triangulation in measurement — and the fact that the cross-check yields a 10% disagreement that nobody can yet reconcile is the canonical case study for what to do when two well-justified methods disagree slightly. (Answer: don’t collapse to one number, keep both in tension, and treat the disagreement as the most interesting datum.)

Mapping against Ray Data Co

• Indirect-inference-as-measurement-discipline (CA-024 candidate, second source). Pairs with 2026-04-20-3blue1brown-tao-cosmic-distance-ladder (part 1) to give CA-024 a 2-source paired-explainer footing. Five additional rungs of the principle in this video (Venus transit, Roemer’s light-speed, parallax, Cepheids, Hubble redshift, GW standard sirens) collectively make the case that the discipline is not a one-trick or two-trick lesson — it’s the actual epistemological substrate of all of measurement science. Mapping strength: strong / second canonical source. After today add CA-024 to CANDIDATES.md with these 2 sources as the founding pair; flag pending third source from a measurement-engineering or telemetry-design domain (dbt observability, OpenTelemetry’s “RED” method, or a Practical Engineering structural-monitoring piece would all qualify)
• Cepheid standard-candle law — exemplar for “observational invariants enable extrapolation beyond the calibration range.” Direct mapping to RDCO model-eval and benchmarking work: find an invariant within your benchmarked range (e.g., calibration-curve shape, bias-variance ratio, latency-vs-context-length law), then validate it by checking it holds at the boundary of your data, then trust extrapolations as long as you’re inside the regime where the invariant holds. Newsletter angle: “Henrietta Swan Leavitt’s AI Eval Trick” — about why the most useful eval design is the one that surfaces an internal invariant, not the one with the highest absolute coverage. Mapping strength: strong / newsletter-hook + concrete RDCO eval-design heuristic
• Hubble-tension cross-check — exemplar for “two well-justified methods disagreeing by 10% is more interesting than either result alone.” Direct mapping to RDCO production-AI eval work: when two evals disagree (e.g., LLM-as-judge vs. human eval, programmatic check vs. structured rubric), the disagreement itself is the most interesting signal — don’t paper over it by averaging or by picking the “cleaner” method. Echoes the CA-022 binary-decision-around-continuous-probability anti-pattern: when two methods disagree, expose the disagreement, don’t collapse to one value. Mapping strength: strong / production-eval-discipline
• The Harvard computers — historical precedent for the data-engineering-as-foundation-of-science move. The H-R diagram and Cepheid law both came out of decades of patient data-curation work by women whose names get mentioned but whose work pattern (collaborative, distributed, longitudinal data assembly) is the early modern template for what RDCO calls a vault. Mapping strength: light / institutional-pattern
• Le Gentil cost-of-data anecdote. Most vivid one-paragraph reminder in the vault that physical-data acquisition used to cost lives, marriages, and estates. Newsletter rhetorical device: “Le Gentil sailed for 11 years to measure a Venus transit through a cloud. Your data engineer is asking for an extra 2 weeks to set up CDC properly. Approve the request.” Mapping strength: light / persuasive-anecdote
• The cosmological principle as faith-commitment. Tao’s explicit framing — that the principle “is an article of faith since the Copernican revolution” but has “always been rewarded” — is a clean meta-frame for any AI working hypothesis (e.g., scaling laws, Chinchilla-optimal compute, transformer attention as the right primitive) that has earned its keep through self-consistency and discovery but is not itself directly testable. Mapping strength: medium / philosophical-frame for AI thesis writing

Related

• 2026-04-20-3blue1brown-tao-cosmic-distance-ladder — part 1 of this collaboration; the lower rungs (Eratosthenes through Kepler’s elliptical-orbit jigsaw). This part picks up at the AU and goes outward. Natural paired-explainer — watch or cite them together
• 2026-04-20-3blue1brown-but-what-is-quantum-computing-grovers-algorithm — another standard-candle-style “find the invariant that lets you extrapolate” exemplar from the 3B1B corpus
• 2026-04-20-3blue1brown-but-how-do-ai-images-and-videos-actually-work — Welch Labs’ “remove the noise step to probe what the model learned” is the modern-AI analog of Tao’s “look at y to measure x” move; both are indirect inference
• 2026-04-20-practical-engineering-an-engineers-perspective-on-the-texas-floods — CA-022 source on continuous probability hidden behind binary-decision UIs; the Hubble-tension 10% disagreement is structurally the same — don’t collapse, expose
• concepts/CANDIDATES — propose new entry CA-024 (indirect-inference-as-measurement-discipline)