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

Summary

Dwarkesh Patel interviews Michael Nielsen (quantum computing pioneer, author, Astera Institute fellow) about the nature of scientific progress — how discoveries actually happen versus how we mythologize them, what this means for AI-driven science, and why the “tech tree” of possible knowledge is far larger than we assume. The conversation uses rich historical case studies (Michelson-Morley, Einstein, Darwin, Newton) to argue that science cannot be reduced to a simple verification loop, and that different civilizations might explore entirely different branches of knowledge.

Key Takeaways

• Falsification is far messier than textbooks suggest. The Michelson-Morley experiment didn’t “prove the ether doesn’t exist” — it ruled out some ether theories. Michelson himself believed in the ether until he died in ~1929. There’s no clean moment where a theory is falsified and everyone moves on.

• Scientific communities adopt correct theories before experimental proof. Special relativity was widely accepted before the 1940 muon decay experiments that could distinguish it from Lorentz’s ether-based interpretation. Some non-experimental “taste” or aesthetic guides progress faster than verification loops.

• Expertise can be a prison. Poincare had almost all the pieces of special relativity but couldn’t let go of a dynamical interpretation of length contraction. Lorentz had the math right but the interpretation wrong. Einstein, younger and less attached, made the leap.

• AI and the verification loop problem. People assume AI will accelerate science because it accelerates domains with tight verification loops (like coding with unit tests). But science has infinitely many theories compatible with any experiment. The hard part — choosing which theory is actually correct — isn’t easily automated.

• AlphaFold is mostly a data story. The billions of dollars spent acquiring 180,000+ protein structures via X-ray diffraction, NMR, and cryo-EM are the main achievement. The AI model, while impressive, is a small part of the total investment.

• AI models as a new kind of scientific object. Three interpretations: (1) they’re not real explanations, just useful models; (2) they contain many small explanations that interpretability can extract; (3) they’re a genuinely new type of object we can do operations on (merge, distill, compress) — analogous to how Mathematica lets physicists work with 100-page equations that were previously intractable.

• The tech tree is vastly larger than we think. Computer science started with a “theory of everything” (Turing/Church in the 1930s) and we’ve spent 90 years discovering deep ideas hidden within it (public-key cryptography, etc.). We’re near the bottom of the tree, not the top.

• Alien civilizations would likely have different technology. Different sensory biases, different historical contingencies, different paths through the tech tree. This implies massive gains from trade between civilizations — making friendliness more rewarding than domination.

• Diminishing returns are an illusion of static snapshots. The “dessert buffet” argument (best ideas get picked first) breaks down because new fields keep appearing — like someone restocking the buffet with better desserts. Computer science, quantum computing, deep learning — each opened an explosion of low-hanging fruit.

• The bottleneck keeps moving. Programmers are no longer bottlenecked on writing code (AI handles that), but now they’re bottlenecked on having interesting design ideas — something with no verification loop.

Detailed Notes

The Michelson-Morley Myth vs. Reality

The standard story: Michelson-Morley proved the ether doesn’t exist, creating a crisis that Einstein solved. The real story: the experiment tested competing theories of the ether (particularly whether an “ether wind” existed). It ruled out some theories but not all. Lorentz developed the correct mathematical transformations (Lorentz transformations) but interpreted them as effects of moving through the ether. His theory was experimentally indistinguishable from special relativity until muon decay experiments in ~1940. Michelson continued believing in the ether until his death. The “process” by which science converges on truth is not a method — great scientists can remain wrong long after the community shifts.

Why Did Darwinism Take So Long?

Natural selection seems conceptually simpler than gravity, yet Newton’s Principia (1687) preceded Darwin’s Origin (1859) by 172 years. Key blockers: (1) Deep time — you need Lyell’s geology (1830s) establishing millions/billions of years for evolution to be plausible; (2) Biogeography from colonial voyages; (3) Paleontology revealing intermediate species. Darwin’s genius wasn’t the idea of selection (known to animal breeders) but building the comprehensive case that it explains the entire biosphere. The simultaneous independent discovery by Alfred Wallace suggests the building blocks had to be in place first.

The Problem with Automating Science

The Prout hypothesis example: In 1815, chemist Prout hypothesized all atomic weights are whole numbers (made of hydrogen). Chlorine at 35.5 was a problem. For 85 years, the verification loop was actively hostile to the correct theory — the anomaly was due to isotopes, which couldn’t be chemically distinguished. You needed the concept of isotopes (discovered ~1900) to resolve it. Similarly, the Pioneer spacecraft anomaly looked like it might falsify general relativity but turned out to be asymmetric thermal radiation. 99.9% of apparent exceptions to theories are mundane effects. There’s no ex ante heuristic to tell which case you’re in.

Quantum Computing’s Origin Story

Von Neumann could have invented quantum computing in the 1950s — he was pioneering both computation and quantum mechanics. Two historically contingent developments around 1980 made it possible: (1) personal computers made computation salient to physicists (Feynman famously got excited buying an early PC); (2) Paul traps enabled manipulation of single quantum states. Nielsen himself entered the field in 1992 via Gerard Milburn, who handed him a stack of foundational papers (Feynman 1982, Deutsch 1985) at a time when almost nobody was working on it.

On Learning and Depth

Nielsen distinguishes two types of work: routine work (minimize procrastination, outsource, do fast) and high-variance work (requires time stuck, exploration, talking to different people). Being stuck is “maybe even the most important part.” Essays written in 2 days teach less than those that took 3 months. The difference in “going deep” between people is enormous — for some it means reading blog posts, for others it means writing the definitive book. LLMs are seductive because they provide an easy escape from the aversive feeling of being stuck, which is precisely the feeling that produces real understanding.

Open Science

The attribution economy of science is socially constructed. Example: physicists upload preprints to establish priority quickly because “physics is so competitive”; biologists avoid preprints to protect priority because “biology is so competitive.” Same reasoning, opposite conclusions. The modern political economy of science (papers, attribution, reputation) was constructed over a century after Galileo/Kepler’s era of publishing discoveries as anagrams. Open science is about updating this economy for the internet age — making code, data, and in-progress ideas shareable with proper credit.

Quotes / Notable Moments

“Newton was not the first of the age of reason. He was the last of the magicians, the last great mind which looked out on the visible and intellectual world with the same eyes as those who began to build our intellectual inheritance rather less than ten thousand years ago.” — Keynes on Newton

“If you’re attempting to reduce science to a process, you’re attempting to reduce it to something where there is just a method which you can apply, and you turn the crank and out pops insight. You can do a certain amount of that, but you’re going to get bottlenecked at the places where your existing method doesn’t apply.”

“I wish there was a very large number of biographies of people who are fantastically talented who just missed.” — Nielsen on what we could learn from brilliant people who never achieved breakthroughs

On LLMs and learning: “It’s entertaining but not necessarily anything else… it makes it easier because instead of doing some intermediate thinking, there’s always a next question you can ask a chatbot.”