Michael Nielsen – How science actually progresses

Dwarkesh frames the whole conversation around a question that turns out to matter for AI: how do we recognise scientific progress? The immediate motivation is that people are trying to close the RL verification loop on scientific discovery, and it is worth asking what closing that loop would even mean. His reading in preparation changed his view — progress in the history of science is a stranger and more elusive force than he had assumed.

The test case is Michelson–Morley. The standard telling is that an experiment in the 1880s showed the ether did not exist, created a crisis in physics, and Einstein resolved it with special relativity.

Nielsen's correction starts with what Michelson and Morley thought they were doing. The ether goes back to Robert Boyle in the 1600s: sound is vibration in air, so is light vibration in something? Boyle showed light propagates through a vacuum and sound does not, and the ether was introduced to be that something. For two centuries people argued about its nature.

The experiment was designed to discriminate among theories of the ether, and specifically to detect an ether wind: if the Earth moves through the ether, a light beam sent along the wind is accelerated and one sent against it is slowed, which should show up in interference. What they found was no ether wind. That ruled out some ether theories and not others, and Michelson kept believing in the ether for the rest of his life.

Nielsen dates the conviction precisely. The first experiment was 1881, not 1887; Michelson was still running ether experiments in the 1920s and still believed a year or two before his death around 1929. Miller continued at Mount Wilson, reasoning that at sufficient altitude the ether wind would not be dragged by the Earth, and reported a measurement. Einstein's response is the famous line: subtle is the Lord, but malicious He is not.

Dwarkesh draws the methodological point. Real history differs from the textbook picture of the scientific method because falsification is not as clean as it looks: it is unclear what has been falsified. One version of the ether, certainly. And you cannot induce special relativity from the disconfirmation of one ether theory.

Nielsen agrees that naive falsificationism is not so much refuted as shown to be far more complicated in practice. Michelson's 1881 apparatus had problems — Rayleigh among others pointed them out — so it was redone in 1887, after which most leading physicists accepted that there was no ether wind. But the question was what to do next. You had falsified some ether theories; others were untouched, and people set to work developing those.

Even the definite article is wrong, he notes: there was not "the" ether but a great many theories, with a couple of leading contenders. The response of the leading physicists was not "there is no ether" but "this tells us a great deal about what the ether must be".

Dwarkesh puts the sharper case. Lorentz, before Einstein, derived the transformations between reference frames that special relativity is built on — but interpreted them as conversions from the ether frame to non-privileged frames moving relative to it. Length contraction and time dilation were, for him, dynamical effects of moving through the ether, a pressure warping clocks and rulers. And experimentally, on his account, his interpretation and Einstein's were indistinguishable.

Nielsen softens the claim. Lorentz introduced local time without trying to give it a physical interpretation — Einstein later simply recognised it as time in another inertial frame — and Poincaré came much closer to seeing it as the time registered by clocks.

The discrimination came forty years later with cosmic-ray muons. Muons produced in the upper atmosphere decay far too slowly to reach the ground under classical assumptions, and the measured decay rates in 1940 matched special relativity exactly. Had Lorentz lived — he had been dead about a decade — he would probably have patched his theory again, but it would have been a massive setback: the mathematical convenience turns out to be what time actually is, at least for muons.

Dwarkesh revises his own claim accordingly. It is not that the two could never have been distinguished experimentally; it is that the community adopted what we now regard as the correct interpretation before any experiment preferred it. Something in human science does discriminate between theories.

Nielsen interrupts on the word itself. Process carries a connotation of something set in advance, and in practice it is far more complicated. Lorentz, whom Einstein utterly admired; Poincaré, one of the greatest scientists who ever lived; Michelson, another outstanding one — none of them ever reconciled themselves to the new picture. Great scientists can remain wrong for a very long time after the community has broadly changed its mind, and there is no centralised authority or centralised method doing the changing.

Dwarkesh's summary is that progress happens even though it is hard to articulate the process or the heuristics by which it happens.

Dwarkesh raises a recurring pattern: someone asks the right question and then fails to close it. Poincaré looks like the opposite failure to Lorentz — Lorentz had the mathematics and the wrong interpretation, Poincaré seems to have understood that simultaneity is circular to define, since velocity is defined in terms of time.

Nielsen's answer is that these cases have to be taken one at a time, because they are not all making the same mistake. Poincaré is remarkable: he appears to have grasped the principle of relativity — that the laws of physics are the same in all inertial frames — and that the speed of light is the same in all of them, which are essentially the premises Einstein deduces special relativity from.

But Poincaré also held that length contraction is dynamical, that particles are being pushed together by some external force, rather than a purely kinematic consequence of space and time being different from what we thought. In a 1909 paper he still had the dynamical picture. From the modern point of view this is simply a mistake.

Why cling to it? Nielsen says plainly that he does not know — he never met the man — and offers a guess: expertise got in the way. Poincaré knew and understood so much that he could not let go. The revealing contrast is that Einstein as a teenager, in the 1890s, also believed in the ether and knew the same material, but was not as attached to it. Perhaps the older generation were prisoners of their own expertise; he notes that some historians of science would certainly disagree.

Dwarkesh closes the loop by noting that Einstein himself is later said to have missed the correct interpretations of quantum mechanics and cosmology for the same reason — his own attachments.

Dwarkesh generalises the muon point: scientific communities seem to make progress faster than the long verification loops imply. His cleanest case is Aristarchus in the second century BC, who had heliocentrism and was dismissed on observational grounds — if the Earth moves, the stars should shift — with the escape hatch being that the stars are simply too far away for the shift to be seen.

Stellar parallax was not measured until 1838. So heliocentrism did not have to wait for its verification, and neither did the judgement that Copernicus was in some way better. And the usual retrospective justifications fail: the Ptolemaic model was more accurate, because it had centuries of epicycles added to it — and, less often appreciated, it was also in some sense simpler, since Copernicus's insistence on uniform circular motion forced him to add more epicycles than Ptolemy had. Not more accurate, not simpler: how could anyone have known ex ante?

Nielsen offers a partial answer that he finds compelling with centuries of hindsight: unification. Newton's gravitation explained the motions of the planets from Kepler's laws, and then, out of the same theory, explained terrestrial motion — why objects move in parabolas — and the tides, as the moon's and sun's gravitational effect on water. Three apparently disconnected phenomena, one set of ideas, and that convergence is what feels compelling.