Michael Nielsen – Why aliens will have a different tech stack than us

The Michelson-Morley experiment (1887) did not prove the ether's non-existence but rather falsified specific "ether wind" theories, leading to decades of continued belief in the ether by its discoverer, Albert Michelson, who died believing in it.

• • Michelson continued conducting experiments into the 1920s, and physicist Edward Miller reported detecting an ether wind at high altitudes, which Einstein dismissed with the quote, "Subtle is the Lord, but malicious He is not."
• • The experiment's true historical impact was not the immediate adoption of Special Relativity, but the development of the Lorentz transformations by Hendrik Lorentz to save ether theories through physical length contraction and time dilation.
• • Einstein's Special Relativity and Lorentz's ether-based interpretation were empirically indistinguishable for decades until muon decay experiments in the 1940s confirmed that time dilation was a fundamental kinematic property, not a physical pressure effect.
• • Henri Poincaré understood the relativity of simultaneity but clung to a "dynamical" interpretation of length contraction, suggesting that expertise can sometimes prevent scientists from recognizing a more fundamental kinematic truth.

Scientific progress often relies on heuristics and "taste" rather than strict falsification, as multiple competing theories can fit the same data while one is eventually chosen based on explanatory power or aesthetic simplicity.

• • Aristarchus proposed heliocentrism in the 2nd century BC, but it was rejected for lack of observable stellar parallax, which was not measured until 1838; the theory was accepted long before empirical verification due to its theoretical elegance.
• • Newton's Principia (1687) unified disparate phenomena (planetary motion, terrestrial projectile motion, tides) under one theory, a feat that provided strong non-empirical motivation for its acceptance despite the Ptolemaic system being more empirically accurate at the time due to its complex epicycles.
• • Charles Lyell's work on "deep time" (1830s) was a necessary precursor to Darwin's Origin of Species (1859), providing the geological timescale required for evolution to appear plausible.
• • The "verification loop" in science is often hostile or extremely long, as seen in the 85-year delay in accepting the existence of isotopes to explain non-integer atomic weights (e.g., Chlorine at 35.5), where the data actively argued against the correct theory.

AI's role in science differs from classical discovery in that it often prioritizes high-accuracy fitting of complex models over parsimonious explanatory principles.

• • AlphaFold's success is largely attributed to the existing database of 180,000 protein structures from decades of experimental work, with the AI serving as a small fraction of the total investment; it predicts folding but lacks the broad explanatory reach of a theory like General Relativity.
• • Deep learning models may function as a new type of scientific object that requires "archaeology" to extract interpretable principles, similar to how 19th-century physicists had to manually simplify 100-page equations before the advent of tools like Mathematica.
• • Current AI research faces a bottleneck in generating novel "design ideas" rather than code generation, mirroring the shift where the verification loop for coding (unit tests) is solved, but the loop for conceptual novelty remains open.

The "tech tree" of scientific knowledge is vast, path-dependent, and likely never fully explored by any single civilization.

• • Fields like computer science emerged from esoteric logic questions in the 1930s, demonstrating that new domains can open up "low-hanging fruit" even when established fields face diminishing returns.
• • Different civilizations might develop entirely different technological stacks (e.g., oral vs. visual biases in reasoning), leading to massive gains from trade if they can share their unique discoveries in the "tech tree."
• • Comparative advantage in trade may be limited by transaction costs and the difficulty of transferring tacit "process knowledge" (e.g., manufacturing skills in China) compared to pure information.

Scientific breakthroughs often require a confluence of historical contingencies and the maturation of external conditions.

• • Quantum computing did not emerge in the 1950s because the necessary conditions—salient personal computing power and the ability to trap single ions—did not mature until the 1980s (Feynman, 1982; Deutsch, 1985).
• • Michael Nielsen entered the field of quantum information in 1992 not because he discovered it was a low-hanging fruit, but because he recognized the deep, open questions in Feynman and Deutsch's work as a high-variance, high-reward area.
• • The "market for follow-ups" in science is driven by individuals identifying specific, provocative problems where they can contribute, rather than a centralized method of discovery.

The "Open Science" movement aims to reform the political economy of science by shifting from a priority-protection model to a collective attribution economy.

• • Historical practices like publishing results as anagrams (Galileo, Kepler) were attempts to secure priority in a pre-digital economy; modern preprint culture (physicists) and journal dominance (biologists) reflect different strategies for the same goal.
• • Complex projects like the LHC require "collective science" where no single individual understands all specialized components (accelerator physics, detector physics, inverse methods), necessitating a network of experts rather than a lone genius.

Deep learning requires "forcing functions" and high-stakes creative artifacts to achieve true understanding, rather than passive consumption of information.

• • Podcasting and AI-assisted learning can create an illusion of understanding by glossing over difficult intermediate steps; true compounding requires implementing problems (e.g., coding the Transformer) or writing deep expository essays.
• • The "Equal Odds" rule suggests that high-volume output increases the probability of producing a seminal work, but avoiding the "great project" trap requires a balance between routine execution and high-variance creative work.
• • Learning is most effective when the "stakes" are raised (e.g., the fear of public judgment or the need to explain to an expert) or when the curriculum includes specific practice problems that force integration of knowledge.