The Man Who Disproved Ether Believed in It Until His Dying Day

Source: Dwarkesh Patel | Published: 2026-04-07T16:33:36Z

Michelson conducted the famous experiment that falsified the ether hypothesis, yet he publicly maintained his belief in ether until his death in 1931 — and Einstein even admitted he may never have read the paper.

The Real Story of the Ether Experiment Is Far More Complicated Than the Textbook Version

Physicist Michael Nielsen recently recounted the true history of the Michelson-Morley experiment on Dwarkesh Patel's podcast. The standard narrative goes like this: the 1887 experiment proved the ether didn't exist, triggered a crisis in physics, and Einstein proposed special relativity to resolve it.

What actually happened was entirely different. Michelson and Morley weren't testing "whether the ether exists" — they were testing the differences between competing versions of ether theory, specifically whether a so-called "ether wind" existed. They found no ether wind, which ruled out certain ether theories but fell far short of ruling out all of them. Michelson himself believed in the ether until the day he died. Another physicist, Miller, traveled to Mount Wilson in California in the 1920s to run experiments, reasoning that at sufficient altitude the Earth wouldn't drag the ether wind along, making it detectable.

Einstein later said he wasn't even sure he'd read the paper at the time. While there's evidence he most likely had, the experiment simply wasn't decisive for his thinking. What drove special relativity was something else entirely.

Lorentz Got the Math Right but the Interpretation Wrong

In the late 19th century, Lorentz derived the mathematical formulas for converting between reference frames — the Lorentz transformations, which form the mathematical foundation of special relativity. But his interpretation went like this: you're converting from the ether reference frame to other frames, and length contraction and time dilation are physical stress effects on objects moving through the ether.

He introduced a quantity called "local time" but didn't assign it much physical significance. Poincaré later came closer to recognizing it as the time actually recorded by clocks. But it took Einstein to say: space and time themselves are not what we thought they were.

About forty years later, scientists observed the decay rate of muons produced when cosmic rays strike the top of the atmosphere — they decayed far more slowly than classical theory predicted. If time really does slow down at high velocities, everything checks out. The 1940 measurements matched special relativity's predictions exactly.

Nielsen points out that the scientific community adopted the theory we now consider correct before it was experimentally verified. Some judgment process was clearly at work, but it's remarkably hard to articulate what it was.

Poincaré: Knowing Too Much Became an Obstacle

Poincaré appears to have understood the principle of relativity — that the laws of physics are the same in all inertial reference frames — and that the speed of light is invariant across all inertial frames. These are essentially the premises Einstein used to derive special relativity.

But Poincaré simultaneously believed length contraction was a dynamical effect — some external force squeezing particles together. He didn't grasp that it was purely a kinematic issue: space and time themselves are different from what we assumed. As late as a 1909 paper, he was still clinging to this dynamical picture.

Nielsen's conjecture: his expertise got in the way. He knew too much, understood too deeply, and couldn't let go. Einstein in the 1890s was a teenager who also believed in the ether, but his attachment to these ideas wasn't nearly as deep as that of the older generation. There's an enormous difference between encountering an idea when you're young and having built your entire career on it — it fundamentally changes your ability to abandon it.

Copernicus Was Neither More Accurate Nor Simpler

The standard narrative of heliocentrism doesn't hold up either. The Greek astronomer Aristarchus proposed a heliocentric model in the third century BC. It was rejected on the grounds that if the Earth orbits the Sun, we should observe stellar parallax. The only way out was that stars must be extraordinarily far away — and stellar parallax wasn't actually measured until 1838.

When Copernicus published his theory, the Ptolemaic model was actually more accurate, thanks to centuries of fine-tuning its epicycles. Even fewer people realize that Copernicus's model wasn't simpler in any meaningful sense either — his obsession with having the Earth move in perfect uniform circular motion meant he actually needed more epicycles than Ptolemy.

Neither more accurate nor simpler — so how could anyone have known in advance that Copernicus was right?

Nielsen offers a partial answer: Newton later used a single theory of gravity to simultaneously explain planetary motion, the parabolic trajectories of objects on Earth, and the tides. Three seemingly unrelated phenomena unified by one set of ideas — that kind of unification starts to become extremely compelling.

Darwin's Genius Wasn't Coming Up with Natural Selection

The basic idea behind natural selection — artificial selection — was something animal breeders had probably known for ages. Darwin's genius lay in understanding just how central this principle was. He didn't just propose an idea; he wrote the entire Origin of Species to demonstrate natural selection's connection to virtually every phenomenon in the biosphere — geology, biogeography, the fossil record.

Nielsen notes that Lucretius had a seemingly similar idea in the first century BC, but closer inspection reveals the two are vastly different. Lucretius imagined a one-time generative period that produced all species, followed by a single filter that kept the ones suited to their environment. He had absolutely no concept of a "continuous gradual process" or a "tree of life" connecting all living things.

The fact that Darwin and Wallace independently arrived at natural selection almost simultaneously tells us certain preconditions had to be in place first. Lyell's introduction of deep time in the 1830s — the idea that Earth is millions or even billions of years old — seems to have been the critical unlock. If you only have five to ten thousand years, evolution should be happening at a pace visible within a human lifetime, and we clearly don't observe that.

Falsificationism Is Far Harder to Apply Than You Think

Uranus's orbit was slightly off, so people predicted Neptune's existence — Le Verrier said in 1846 to point a telescope in a certain direction, and there it was. A towering triumph for Newtonian gravity.

Mercury's orbit was also slightly off, with its elliptical path precessing an extra 43 arc-seconds per century. People applied the same logic and predicted a planet inside Mercury's orbit, called it "Vulcan," pointed telescopes at it — and found nothing. But if you were a committed Newtonian, you'd say maybe cosmic dust was blocking the view, maybe the planet was too small to see, maybe we needed a more powerful telescope.

This pattern recurs constantly in the history of science. In the 1990s, someone discovered that the Pioneer spacecraft weren't where they were expected to be, and some people excitedly suggested general relativity might need to be overturned. The accepted explanation today: the spacecraft had slight asymmetries, with thermal radiation slightly stronger on one side than the other, producing a tiny acceleration toward the Sun.

99.9% of the time, so-called "anomalies" have these kinds of mundane explanations. But occasionally they don't. You can't tell in advance which situation you're in. This is the fundamental difficulty with falsificationism as a straightforward methodology.

The Verification Loop Can Work Against the Correct Theory for Decades

In 1815, the chemist Prout hypothesized that all atomic weights are whole numbers — that atoms are essentially all built from hydrogen. Most measured elemental weights were indeed close to integers, but chlorine's weight was 35.5. The Prout school proposed various ad hoc hypotheses: maybe chemical impurities, maybe whole-number fractions. But more precise measurements of chlorine yielded 35.46, pushing it even further from the nearest fraction.

It took 85 years before anyone discovered the concept of isotopes — chlorine's different isotopes can't be distinguished chemically, only physically. For those 85 years, the verification loop was actively hostile to the correct theory.

This has direct implications for the discussion about AI accelerating science. Some argue AI will make disproportionate progress in science because science, like programming, has tight verification loops — you can run experiments. But any single experiment is compatible with infinitely many theories. Why we converge on one theory over others over time is itself a process that's hard to articulate.

AlphaFold's Success Is Mostly a Data Collection Story

AlphaFold, the flagship success story of AI-accelerated science, has been misread, according to Nielsen. Most of AlphaFold's success is owed to the Protein Data Bank — roughly 180,000 protein structures acquired over decades through X-ray crystallography, NMR, and cryo-EM, at a cost of billions of dollars.

At its core, this is a story of humans spending decades painstakingly observing the world through experiments, then fitting a beautiful model at the end. The AI component is impressive but represents only a small fraction of the total investment.

The more interesting philosophical question: does AlphaFold count as a scientific theory? General relativity can predict phenomena it was never designed to explain — like the precession of Mercury's perihelion. AlphaFold has no such explanatory reach.

Nielsen offers three ways to look at it: first, it's not a scientific explanation in the classical sense, just a useful model; second, it contains many small explanations internally that could be extracted through interpretability research; third, it's an entirely new type of cognitive object, and we haven't yet developed the vocabulary to describe it. We can merge models, distill them — this is a major opportunity for the philosophy of science.

What If You Had Trained a Model Before Copernicus

Imagine an alternate history where we had deep learning before cosmology. You train a model to fit observational data of celestial motion, then run interpretability analysis. You'd most likely just keep discovering "oh, there's another epicycle we missed" — parameters X through Y encode this epicycle, the next set encodes another.

You can keep stacking epicycles on the Ptolemaic model forever, but it takes a human mind to integrate all the information and say: there's a more coherent way to understand this as a whole. From a gradient descent perspective, this kind of leap — to a theory that's locally worse but globally better — is very unlikely to happen on its own.

Nielsen concedes that perhaps some form of regularization or distillation could extract simpler theories from enormously complex models. But the critical shift — like from Newtonian gravity to general relativity — was driven by the recognition of a contradiction. Almost immediately after proposing special relativity, Einstein realized: in special relativity, influences can't propagate faster than light, but in Newtonian gravity, gravitational force acts instantaneously at a distance. That contradiction was the force that demanded a new theory.

The Tech Tree Is Far Larger Than We Realize

Nielsen makes a striking claim: if we ever encounter an alien civilization, they'd almost certainly have a completely different tech stack.

When computer science got started in the 1930s, Turing and Church laid down a "theory of everything" — the fundamental principles of computation. But we've spent over ninety years exploring the consequences and are nowhere close to finished. Public-key cryptography, a profoundly deep and non-obvious idea, was hiding inside computational theory since the 1930s.

Phases of matter — he learned three or four in school, but as a physicist discovered the list keeps growing: superconductors, superfluids, Bose-Einstein condensates, quantum Hall systems, fractional quantum Hall systems... We will even design new phases of matter. It looks like we're still near the bottom of the tech tree.

Knuth mentions in the preface to The Art of Computer Programming that when he started writing in the 1960s, a mathematician dismissively said computer science wasn't really a field yet — "come back when you have a thousand deep theorems." Decades later, Knuth wrote: there are obviously a thousand deep theorems now.

The Dessert Table Rebuttal to Diminishing Returns

On whether science faces diminishing returns, Nielsen uses a dessert table analogy: at a wedding with thirty desserts laid out, the best ones naturally get taken first — that's the intuition behind diminishing returns.

But what if someone behind the table keeps setting out new desserts? Scientific progress has exactly this character. New fields keep emerging, suddenly offering a fresh abundance of low-hanging fruit. Computer science is a perfect example — it was born unexpectedly from some rather esoteric questions in mathematical philosophy and logic, instantly opening up an entirely new domain. Young people flooded in because you could make major breakthroughs at 21 without spending 25 years mastering everything that came before.

Nielsen argues that if external conditions stay the same, diminishing returns will indeed set in. But diminishing returns aren't intrinsic — perhaps what's needed is another shift in external conditions. AI is clearly what many people see as the next driving force.

Why Quantum Computing Didn't Emerge in the 1950s

Von Neumann was both a pioneer of computation and the author of foundational work on quantum mechanics. He could plausibly have invented quantum computing but didn't. Feynman's 1982 paper and Deutsch's 1985 paper are widely recognized as the field's founding works.

Nielsen believes two things matured simultaneously around 1980: first, personal computers — the Apple II, the Commodore 64 — made computation vivid and exciting for far more people; second, Paul traps and the ability to manipulate individual quantum states became practical. Feynman bought one of the earliest personal computers around 1980 and reportedly injured himself carrying it home because he was so excited. A person who deeply understood quantum mechanics becoming simultaneously thrilled by these new machines — this kind of historical contingency isn't surprising at all.

Nielsen himself entered the field in 1992 after his advisor Gerard Milburn handed him a stack of papers. Reading Deutsch and Feynman, he realized these papers were asking profoundly fundamental questions, and a single person could make a contribution. Almost nobody was working in this area at the time. He benefited, in a sense, from his advisor's taste.

Open Science Changed the Political Economy of Research

Three hundred years ago, whether scientists should publicly disclose their findings was itself controversial. Galileo and Kepler sometimes published discoveries as anagrams — scrambling the letters of a sentence so that if someone else later made the same discovery, they could unscramble it and claim priority. It took over a century to get from that starting point to the modern reputation economy of peer review and publication.

Nielsen illustrates the arbitrariness of this attribution economy with a striking comparison: biologists told him that biology is so competitive you have to guard your priority and can't post to preprint servers. Physicists told him that physics is so competitive you have to post to preprint servers immediately to establish priority. The same "intensely competitive" rationale led to diametrically opposite institutional norms.

One of the open science movement's most important achievements has been turning these questions into live issues that people have opinions about and actively debate. "Publicly funded science should be open" — this short slogan condenses an entire set of political economy questions worth taking seriously.

There's a Reason LHC Papers Have Over a Thousand Authors

Nielsen once snuck into an accelerator physics conference. He discovered the attendees were world-class experts in numerical methods for inverse problems — particles get accelerated, collide, produce cascading showers of secondary particles that eventually reach the detectors. Their job is to work backward from the final data to determine what produced it.

A person could spend an entire career learning how to solve these inverse problems while knowing very little about quantum field theory, detector physics, vacuum engineering, or data processing — yet all of these are indispensable for understanding the Higgs boson. Many people understand these topics at a high level, but nobody understands all of them at the depth at which they're actually applied. Detector physics, vacuum engineering, and solving inverse problems are drastically different disciplines, and deeply understanding any one of them is serious work.

Learning One Thing Deeply vs. Knowing Many Things Shallowly

Nielsen has a sharp observation about learning: when different people say they "deeply understand" a topic, they mean completely different things. Some mean they've read a few blog posts, some mean they've read a book, and some mean they've written a book. The standard you set for yourself determines your ability to synthesize knowledge.

He's found that creative work involves two types of tasks: routine work, where you just need to avoid procrastinating, get it done quickly, or outsource it; and high-variance work, where you need to be willing to spend enormous amounts of time, go to different places, talk to different people, most of which won't have any direct payoff at any given moment. Most people are good at one type but underinvest in the other.

He also flags a trap in AI-assisted learning: talking to an LLM has some value, and that's precisely what makes it seductive — it's not completely useless, but it can substitute for the truly difficult thing you should be doing. The hardest tasks come with a strong sense of aversion; you'll grab any excuse to avoid them, and an LLM provides an ever-present next question to ask.

Computer science pioneer Alan Kay, when asked about Linux, said it has little to do with computer science — it's just a big ball of mud. There are a few ideas worth understanding in it, but most of the time what you're learning is just stuff about Linux, not any transferable knowledge. For a certain type of mind, the confusion between learning a system and understanding a discipline holds a particular allure.