Science Progresses Beyond Verification Loops
Scientific progress outpaces slow experimental verification through theoretical unification, explanatory power, and community judgment, not naive falsification—as seen in relativity, heliocentrism, and more.
Verification Loops Are Longer and Weirder Than Textbooks Suggest
Michael Nielsen dismantles the simplified narrative of scientific progress, emphasizing that experiments like Michelson-Morley didn't directly "prove" no ether or inspire Einstein's special relativity. Einstein later claimed he might not have even known the 1887 paper, though evidence suggests he did—it wasn't pivotal. Michelson and Morley aimed to detect an "ether wind" from Earth's motion through the luminiferous ether, a medium for light waves proposed since Robert Boyle in the 1600s. Their null result ruled out some ether models but not all; Michelson believed in ether until his 1931 death, running experiments into the 1920s from Mount Wilson.
Dwarkesh Patel highlights Dayton Miller's 1920s claims of detecting ether effects at altitude, prompting Einstein's quip: "Subtle is the Lord, but malicious He is not." Nielsen notes leading physicists like Rayleigh accepted no ether wind by 1887 but adapted theories rather than abandoning ether. Hendrik Lorentz derived transformations matching relativity's math but interpreted length contraction and time dilation as ether-induced stresses, with "local time" as a mathematical trick. Henri Poincaré grasped relativity principles and light speed invariance but clung to dynamical length contraction via external forces, even in a 1909 paper.
"Things are often much more complicated than you think," Nielsen says. Falsification isn't straightforward: null results falsify specific ether variants, not "the ether." Lorentz's view was experimentally indistinguishable from relativity until 1940s muon decay experiments, where cosmic muons survived atmospheric transit due to time dilation—matching special relativity precisely. Yet the community shifted to Einstein's view decades earlier, without decisive proof.
Patel points to Einstein's own later resistance to quantum mechanics interpretations, showing even geniuses cling to priors. Nielsen speculates Poincaré's "expertise... got in the way," unlike teenage Einstein, less attached to ether dogma.
Explanatory Unification Drives Adoption Over Prediction or Simplicity
Progress accelerates beyond verification via theories unifying disparate phenomena. Patel cites Aristarchus's 3rd-century BC heliocentrism, dismissed by Athenians expecting stellar shifts (absent due to vast distances; parallax measured only in 1838). Copernicus's model wasn't simpler or more predictive than Ptolemy's epicycle-laden geocentric system—it added epicycles for circular orbits.
Nielsen ties this to Newton: his gravity unified planetary laws (Kepler), terrestrial parabolas, and tides via moon/sun pull. "You have... three very different disconnected phenomena all being explained by this one set of ideas. That starts to feel very compelling."
Patel probes natural selection's late arrival (Darwin 1859), despite farmers observing it. Nielsen hints at deeper cognitive or cultural barriers. On gradient descent discovering general relativity: unlikely, as it optimizes narrow losses, not broad unification.
Aliens likely follow divergent "tech stacks," Nielsen argues, challenging linear progress trees. No universal path; civilizations trade ideas, not converge on identical science.
Community Judgment and Credit Assignment Shape Progress
No centralized "process" exists—"Process carries connotations of something set in advance. It’s much more complicated," Nielsen warns. Great scientists like Lorentz, Poincaré, and Michelson stayed "wrong" long after consensus shifted. Progress relies on decentralized judgment: elegance, unification, fruitfulness.
Nielsen, early quantum computing pioneer and open science advocate, questions credit mechanisms. Prolificness vs. depth: internalizing knowledge demands deliberate practice. Timestamps hint at infinite deep principles remaining, quantum draw, science's magician era (Newton as last).
Patel frames for AI: closing RL loops for discovery needs recognizing progress sans fast verification. Nielsen's hypotheses imply AI must emulate human heuristics—unification, not just prediction.
"Great scientists can remain wrong for a very long time after the scientific community has broadly changed its opinion. But there’s no centralized authority or centralized method."
"Einstein subtracts from Poincaré's vision and says, ‘No. Space and time are just different than what we thought, and here’s the correct picture.’"
"Newton's unification starts to feel very satisfying once they eventually realize it."
Key Takeaways
- Study case-by-case histories to grasp why theories succeed: unification trumps isolated predictions or simplicity.
- Avoid naive falsification; null results constrain, but don't dictate—interpretations evolve via community debate.
- Expertise can blind: younger outsiders like Einstein shed priors faster than entrenched giants.
- For AI scientific discovery, emulate human progress signals: explanatory breadth, not just empirical fit.
- Aliens' divergent stacks mean no universal tech tree—focus on tradeable principles over convergence.
- Long loops (e.g., 2000 years for parallax) show theory leads experiment; reward bold unification early.
- Rethink credit: science needs mechanisms valuing depth over prolific output.
- Internalize via practice: reading alone insufficient—rebuild ideas from scratch.
- Progress isn't linear or procedural; it's emergent from decentralized genius clashes.