Michael Nielsen – How science actually progresses

What It Covers

Michael Nielsen, quantum computing pioneer and author of the standard quantum information textbook, examines how scientific progress actually occurs — using case studies from Michelson-Morley, special relativity, Darwinism, and AlphaFold to reveal why falsification is messier than textbooks suggest, why verification loops can span decades, and what this means for AI-accelerated discovery.

Key Questions Answered

• •Falsification is non-linear: The 1887 Michelson-Morley experiment did not disprove the ether — it falsified specific ether theories while leaving others intact. Michelson himself continued believing in the ether until his death in 1931, still conducting experiments in the 1920s. Einstein may not have even considered the experiment decisive. Scientific communities adopt new frameworks before experimental verification closes, meaning "crucial experiments" are rarely as decisive as retrospective accounts suggest.
• •Expertise as cognitive trap: Poincaré understood the principle of relativity and constant light speed before Einstein, yet clung to a dynamical explanation for length contraction rather than accepting that space and time themselves are fundamentally different. Einstein, as a younger physicist less invested in prior frameworks, could subtract unnecessary assumptions. When building research teams or evaluating ideas, domain expertise can actively block paradigm shifts — fresh perspectives carry structural advantages at inflection points.
• •Verification loops can span centuries: Aristarchus proposed heliocentrism in the second century BC, but stellar parallax — the direct experimental confirmation — was not measured until 1838. Copernicus's heliocentric model was initially less accurate than the Ptolemaic system, which had centuries of epicycle refinements. Scientific communities adopted the correct framework long before experimental closure. This pattern recurs with muon decay experiments in 1940-41 confirming time dilation decades after special relativity's acceptance.
• •AlphaFold's scientific status is ambiguous: AlphaFold's success rests primarily on the Protein Data Bank — roughly 180,000 structures acquired over decades through X-ray diffraction, NMR, and cryo-EM at a cost of several billion dollars. The AI component is a small fraction of total investment. Unlike general relativity, which predicted Mercury's precession without being designed to, AlphaFold lacks generative explanatory reach. Its value may lie in archaeological extraction — using interpretability methods to surface hidden biological principles from its parameters.
• •The tech tree is vastly underexplored: Church and Turing's 1930s computability theory already contained public-key cryptography and distributed ledger systems as latent consequences, discovered decades later. Phases of matter — superconductors, Bose-Einstein condensates, fractional quantum Hall systems — keep multiplying beyond the three taught in school. Different civilizations exploring the same foundational physics would likely develop entirely different technological stacks, implying massive gains from trade between advanced civilizations persist indefinitely, making cooperation structurally rewarding.
• •Parallel research programs are structurally necessary: Neptune's discovery confirmed Newtonian gravity when Uranus's orbit deviated; Mercury's precession required general relativity when the predicted planet Vulcan did not exist. Both anomalies looked identical at the outset — no ex-ante heuristic distinguishes which exceptions overturn theories versus which reflect measurement error or missing variables. The Pioneer spacecraft anomaly, initially flagged as a potential general relativity failure, turned out to be asymmetric thermal radiation. Maintaining diverse, simultaneous research programs is the only structural solution.
• •Scientific bottlenecks shift, not disappear: Bloom et al. found that maintaining Moore's Law — 40% annual transistor density growth — required 9% annual growth in semiconductor researchers. Nielsen argues this reflects narrow measurement rather than intrinsic diminishing returns. New fields like deep learning periodically reset the difficulty curve, allowing 21-year-olds to make breakthroughs without decades of prior mastery. AI shifts programmer bottlenecks from code production to design judgment — a bottleneck with no verification loop, making it structurally harder to automate.