All podcasts / Dwarkesh Podcast / Summary

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

2026-04-07 - 123 min - source - Read full transcript
Dwarkesh Patel (host)Michael Nielsen

Key insights

The popular Michelson-Morley story is largely myth: it didn't disprove the ether or directly cause special relativity.
Michelson and Morley were testing specific ether-wind theories, not the concept of ether itself, and Michelson personally believed in some form of ether until he died in the late 1920s, still running experiments on it. Einstein later said he wasn't even sure he'd read the paper, and the result was not dispositive for his thinking.
scientific-progress
Long, actively misleading verification loops are the norm in science, not a rare edge case.
Prout's 1815 hypothesis that all atomic weights are whole-number multiples of hydrogen looked increasingly wrong as chlorine's measured weight (35.46) drifted further from a clean fraction over an 85-year period, until isotopes were discovered and vindicated something close to the original idea. For most of that period the evidence pointed the wrong way.
scientific-progress
AlphaFold's success is mostly a story about data acquisition, not AI algorithms.
The Protein Data Bank represents decades of X-ray diffraction, NMR, and cryo-EM work and several billion dollars spent obtaining roughly 180,000 protein structures. The AI model-fitting step that made headlines is a comparatively small fraction of the total investment behind the result.
ai-and-verification-loops
AI may not accelerate science as uniformly as it accelerates coding, because science lacks coding's tight verification loop.
Coding benefits from unit tests that cheaply falsify wrong answers. Any finite set of experiments is compatible with infinitely many theories, so which explanation the scientific community eventually adopts depends on hard-to-formalize taste, follow-up work, and historical contingency rather than a clean verification signal AI could optimize against.
ai-and-verification-loops
Diminishing returns in research are a property of static fields, not an intrinsic law, because entirely new fields keep opening and resetting the curve.
Using the analogy of a dessert buffet that keeps getting restocked, Nielsen argues that measured diminishing returns (e.g. needing ~9% more semiconductor researchers per year to sustain Moore's Law) hold within an existing field, but computer science itself arose almost by accident from abstruse 1930s questions in mathematical logic and reset the entire game, letting young researchers make major progress again.
ai-and-verification-loops
An alien civilization would likely have discovered a substantially different technology stack than humanity's, not converged on the same science.
Nielsen's core thesis: the tech tree implied by any deep theory of everything, such as computation theory from Turing and Church, is far larger than realized, and different starting biases (visual versus aural cognition, for example) push different intelligences down different branches, so exploration order and outcome are highly path-dependent rather than convergent.
alien-tech-trees
Divergent tech trees imply large permanent gains from trade between civilizations, which should make cooperation more valuable than conquest.
If different civilizations end up with different but complementary capabilities because they explored different parts of the tree, comparative advantage generates ongoing mutual benefit from trade rather than a single winner-take-all outcome, an observation Dwarkesh flags as potentially important for how civilizations, or humans and future AI, choose to coordinate.
alien-tech-trees
Biology can be treated as a naturally occurring alien tech dump that humans have barely begun to decode.
Proteins like kinesin, hemoglobin, insulin, and the ribosome represent roughly four billion years of evolutionary R&D. There are tens of thousands of papers on hemoglobin alone and it is still not fully understood, illustrating how much can remain hidden inside a system humans have had extensive data on for a century.
alien-tech-trees
Durable learning requires a high-stakes creative artifact that forces internalization, not passive exposure.
Nielsen reports that essays he wrote in a couple of days are barely remembered, while ones that took three months he can still recall in detail fifteen years later. The binding constraint isn't time spent but being stuck and forced to work through difficulty, which he calls the most important part of the process even though it used to feel like pure annoyance.
deep-learning-as-a-skill
Interview-based learning risks producing broad but shallow, quickly-depreciating knowledge because there is no forcing function to clamp it.
Dwarkesh contrasts guests with a legible curriculum, such as implementing a transformer after talking to Ilya Sutskever, against guests like Ada Palmer where the conversation was excellent but left no exercise that locked in real understanding, meaning the knowledge fades the way any unclamped black-box mapping of inputs to outputs does.
deep-learning-as-a-skill
Raw prolificness, not selective depth, predicts a creator's most important work.
Citing Dean Keith Simonton's equal odds rule, any given piece of work by a person has roughly the same odds of being their most important, so periods of high output (Darwin's enormous correspondence, Einstein's 1905, Shakespeare's pace) explain career-defining breakthroughs better than waiting for the single great insight, with rare exceptions like Gödel who published very little.
deep-learning-as-a-skill
Scientific attribution norms are a social construction, not a fixed consequence of how competitive a field is.
Nielsen recounts asking physicists and biologists why one field used preprints and the other didn't, and getting mirror-image answers: physicists said their field was too competitive not to post preprints immediately, while biologists said their field was too competitive to risk posting before publication. The same underlying pressure produced opposite conventions, showing the reputation economy is constructed by agreement rather than derived from first principles.
open-science-economics

Books referenced

Media referenced

Companies

Techniques and frameworks

Summary

Dwarkesh Patel and physicist-writer Michael Nielsen spend the conversation dismantling the tidy textbook version of how scientific progress happens, using it to stress-test the idea that AI can simply "close the verification loop" on science the way it has for coding. They open with Michelson-Morley: contrary to the popular story, the experiment didn't disprove the ether or directly hand Einstein special relativity. Michelson kept believing in some form of ether until he died decades later, and the real history involves competing interpretations (Lorentz's mathematically identical but physically different account), decades-long gaps between having the right pieces and recognizing what they mean, and case-by-case idiosyncrasy in why some brilliant scientists (Poincare, Michelson) never made the leap while others (a young, less expert Einstein) did.

That pattern repeats across their examples: the muon decay experiments that took forty years to confirm time dilation, the Mercury-versus-Uranus case where an identical anomaly-plus-hidden-planet hypothesis was right once (Neptune) and wrong once (the never-found "Vulcan"), and the 85-year saga of Prout's whole-number atomic weight hypothesis, which looked more falsified over time until isotopes vindicated a version of it. The throughline is that naive falsificationism cannot explain real scientific change, because any dataset is compatible with many theories, and there's no ex ante heuristic for knowing which anomaly is the one that matters. This directly undercuts the assumption that AI will accelerate science the way it accelerates coding: code has cheap, fast verification via unit tests, while science's verification loops can be long, ambiguous, or actively pointing the wrong direction for a human lifetime.

The conversation's most distinctive idea, and its YouTube title, is Nielsen's claim that an alien civilization would likely have arrived at a substantially different technology stack than humanity's rather than converging on the same science. His argument is that deep theories like computation theory (Turing, Church) imply enormous, still-mostly-unexplored tech trees, evidenced by public-key cryptography and cryptocurrency both hiding inside 1930s computation theory for decades before anyone found them. Different starting cognitive biases would push different intelligences down different branches. Dwarkesh extends this into a striking implication: divergent tech trees imply large permanent gains from trade between civilizations, which could make cooperative relationships more valuable than domination, a point Nielsen says he hadn't considered before. They also reframe biology itself as a naturally occurring alien tech dump, with proteins like the ribosome and kinesin representing four billion years of R&D that humans have barely begun to reverse-engineer despite a century of hemoglobin and insulin research.

They apply the same skepticism to AlphaFold, arguing its headline success is mostly a story about the Protein Data Bank's decades of expensive experimental data acquisition rather than a story about AI algorithms, and debate whether AlphaFold-style models constitute "explanations" at all or are a genuinely new kind of object that interpretability work might excavate into legible principles later, the way Mathematica turned previously unworkable 100-page equations into objects people could keep manipulating. On the economics of research, Nielsen reframes measured diminishing returns (like the finding that sustaining Moore's Law requires roughly 9% more semiconductor researchers per year) as a property of already-mature fields rather than a law of nature, since entirely new fields (computer science emerging from 1930s mathematical logic) periodically reset the curve and let young researchers make major progress again.

The final third turns personal and practical, as Dwarkesh probes Nielsen on how to actually learn deeply from interviews rather than accumulating shallow, fast-depreciating knowledge, a professional hazard he worries about as a podcaster. Nielsen's answer centers on "clamping": durable understanding requires a demanding creative artifact (an essay, a class, an implementation) that forces you to get stuck, and being stuck is now, in his view, the most valuable part of the process rather than something to avoid. He also invokes Simonton's equal odds rule to argue that raw output volume, not waiting for a single insight, predicts a creator's most important work, while cautioning that different people are wired toward either routine efficiency or high-variance risk-taking and rarely both. They close on open science, where Nielsen argues that scientific attribution norms (preprints, credit, priority) are entirely socially constructed rather than derived from how competitive a field is, illustrated by physicists and biologists giving opposite justifications for opposite preprint conventions.

Notable Quotes

"Newton was not the first of the age of reason. He was the last of the magicians." - Michael Nielsen, quoting John Maynard Keynes

"It's funny too, the way we tell the history of science, it sounds so simple. You just focus on the right exception and you realize that you need to throw out the old theory and lo and behold, your Nobel Prize awaits. But in fact, these exceptions are all over the place." - Michael Nielsen

"There's an infinite number of theories that are compatible with any given experiment. Over time, why we latch onto the one we think is more correct in retrospect is, as we're discussing, hard to articulate." - Dwarkesh Patel

"Most parts of the tech tree are never going to be explored. There are just too many interesting ways of combining things. There are too many deep ideas waiting to be discovered, and not only we, but nobody ever is going to discover most of them." - Michael Nielsen

"Spending time stuck is incredibly important. That used to just be annoying. Now it seems like it's maybe even the most important part of the whole process." - Michael Nielsen