Dwarkesh PodcastMichael Nielsen on Dwarkesh Patel: Why Ether Died Slowly
Lorentz fit Einstein equations while keeping the ether ontology; Michelson-Morley only ruled out ether wind, so a single result cannot force a paradigm shift.
CHAPTERS
Michelson–Morley: what the experiment did (and didn’t) falsify
Nielsen re-tells the Michelson–Morley story as scientists at the time understood it: not a clean disproof of “the ether,” but evidence against specific ether models (like an ether wind). The chapter highlights how later textbook narratives compress a messy research program into a simple crisis-and-resolution arc.
- •Michelson–Morley aimed to distinguish among competing ether theories, not “test the ether” in the abstract
- •The null result ruled out an ether wind but still left many ether variants alive
- •Michelson kept working on ether experiments for decades and never fully abandoned it
- •Why naive falsificationism struggles: experiments typically underdetermine theory choice
Lorentz vs. Einstein: same equations, different meanings
They discuss how Lorentz developed transformations that mathematically match special relativity but interpreted them through an ether frame. The conversation emphasizes that physics often advances by changing interpretation and ontology—not just by fitting equations.
- •Lorentz transformations emerged before Einstein, but with an ether-centric interpretation
- •“Local time” began as a mathematical device before being recognized as physical time in another frame
- •Poincaré approached key principles but retained a dynamical picture of length contraction
- •Early adoption of Einstein’s interpretation preceded some later decisive experiments
When verification arrives decades later: muons and time dilation
Nielsen uses cosmic-ray muon lifetime measurements (mid-20th century) as an example of a long verification loop that eventually makes the ‘time really changes’ interpretation feel unavoidable. This illustrates how communities sometimes commit to frameworks before the most decisive empirical clinchers appear.
- •Muon decay in the atmosphere provides strong confirmation of relativistic time dilation
- •This kind of evidence came ~40 years after special relativity’s formulation
- •Lorentz-style patching might have been possible in principle but would become increasingly strained
- •Scientific consensus can shift on grounds beyond immediate decisive experiments
Copernicus vs. Ptolemy: progress without being simpler or more accurate (yet)
Dwarkesh presses on why heliocentrism was adoptable despite weaker predictive accuracy and even added epicycles in early Copernican models. Nielsen points to later unification—especially Newtonian synthesis—as a key reason some frameworks feel more compelling than their early scorecards suggest.
- •Ancient parallax objection delayed heliocentrism’s verification until the 1800s
- •Early Copernicus wasn’t clearly simpler or more accurate than refined Ptolemy
- •The later Newtonian unification (terrestrial + celestial + tides) made the overall picture compelling
- •Scientific progress can be ‘global’ (unificatory) rather than locally error-minimizing
Why natural selection wasn’t obvious earlier: prerequisites and “making the case”
They explore why Darwin’s idea took so long to land despite breeders understanding pieces of selection. The key barrier wasn’t just the core mechanism but assembling deep time, geology, biogeography, and a persuasive evidential web—plus grappling with missing mechanisms like heredity.
- •Artificial selection was known, but Darwin’s leap was its explanatory centrality across biology
- •Deep time (Lyell) and fossil evidence were critical enabling context
- •Darwin lacked genetics; missing heredity mechanisms complicated acceptance and development
- •Independent discovery (Wallace) suggests ‘building blocks’ had become ripe in the 1850s
Automation limits: AlphaFold, data accumulation, and what counts as explanation
AlphaFold is treated as both a signature AI success and a reminder that decades of experimental infrastructure (protein databanks, imaging techniques, funding) were the main driver. They debate whether large neural models are ‘scientific explanations’ like GR, or a new kind of object from which explanations might be extracted.
- •AlphaFold depends heavily on massive prior experimental datasets and institutions
- •Classic explanation ideals: simple principles, few parameters, broad counterfactual reach
- •Possible middle ground: models aren’t explanations, but contain extractable ‘local explanations’
- •Analogy to tools like Mathematica: formerly intractable objects become workable intermediates
Could gradient descent find general relativity? Big theory shifts and forcing functions
Dwarkesh worries that optimization over observational fit might just produce “more epicycles,” missing global theory swaps. Nielsen frames Einstein’s path to GR as driven by incompatibilities (finite signal speed vs. action at a distance), plus long exploration through ugly intermediates before landing on a simple final form.
- •Pure curve-fitting risks entrenching complex patches rather than triggering paradigm flips
- •Einstein’s forcing function: SR forbids instantaneous influences, so Newtonian gravity must change
- •Major advances often pass through messy, wrong intermediate theories
- •Progress may require multiple parallel research programs, not one monolithic optimizer
Why aliens will have a different tech stack: the tech tree is vast and path-dependent
Nielsen argues that “science finished by a theory of everything” is a category error: even after fundamentals, there’s immense unexplored combinatorial space (as in computer science). Differences in perception, embodiment, and early choices could steer civilizations into distinct regions of the tech tree—yielding genuinely different technological stacks.
- •Computer science had its ‘TOE’ (Turing/Church) early, yet decades of deep discoveries followed
- •Many deep primitives (e.g., public-key cryptography, ledgers) were latent in the foundations
- •Physics analogy: proliferating ‘phases of matter’ suggests vast remaining discovery space
- •Path dependence: perceptual/cognitive biases and contingent history can steer exploration directions
Diminishing returns vs. “new desserts”: why new fields keep appearing
They examine the common ‘low-hanging fruit is gone’ story and Nielsen’s counter: progress is not a fixed buffet—new desserts get restocked when new fields, tools, and representations open up. Attention, fashion, and institutional dynamics determine which frontiers get resourced and which remain invisible.
- •Diminishing returns assumes a static, visible menu and similar preferences across researchers
- •New fields arise from unexpected roots (e.g., logic/philosophy of math → computer science)
- •Centralized attention makes it feel like only a few ‘big’ areas exist at once
- •Resource intensity (e.g., deep learning) may reflect rapid scaling once a seam is found
Are there infinitely many deep principles? Measuring progress and the Bloom “ideas get harder” result
Dwarkesh asks whether there are endlessly many Noether/Church–Turing-level principles left, given empirical evidence that maintaining progress requires more researchers. Nielsen is skeptical that narrow productivity metrics capture the arrival of new fields and spillovers, and suggests institutional and tool changes can reset the difficulty curve.
- •Speculation: repeated discovery of new “fundamental primitives” suggests depth remains
- •Bloom-style studies track narrow metrics; they can miss new paradigms and externalities
- •Institutional and tooling shifts (security for researchers, instruments, automation) change slopes
- •AI and advanced instrumentation already act as ‘robots’ extending scientific reach
Quantum computing’s origin story: why the field ignited in the 1980s and why Nielsen joined early
Nielsen explains why quantum computing could have been invented earlier (von Neumann era) but wasn’t: computation became salient and single-quantum-state control emerged around the same time. He recounts how a mentor handed him foundational papers in 1992, and why Deutsch/Feynman-style questions felt both deep and tractable.
- •Key contingencies: personal computing salience + experimental control (ion traps, etc.)
- •Foundational papers (Feynman 1982, Deutsch 1985) reframed computation as physical simulation
- •Nielsen’s entry: exposure via Gerard Milburn’s paper stack in 1992
- •Heuristic: find underexplored areas with fundamental questions and reachable contributions
Open science and the credit economy: what ‘success’ looks like and what’s still missing
They treat open science as a redesign of the political economy of knowledge: papers, code, data, and preprints all sit in different credit regimes. Nielsen emphasizes that attribution norms are socially constructed (physics vs. biology preprints) and that better incentive systems can unlock more collective progress.
- •Open science wins: making open access, open code, and open data salient and more standard
- •Historical parallel: shifting from anagrams/priority games to journal-based disclosure norms
- •Preprint culture shows credit norms can invert across fields depending on incentives
- •Core challenge: assign reputation/credit for intermediate artifacts (datasets, code, partial ideas)
Prolificness vs. depth—and how to actually internalize what you learn
The conversation turns personal: how creators balance routine output with high-variance exploration, and how “deep learning” (in the human sense) often requires demanding stakes and time spent stuck. They discuss why podcasts (and now LLMs) can create an illusion of understanding unless paired with forcing functions like exercises, writing, or building artifacts.
- •Two modes: routine execution (avoid procrastination) vs. high-variance exploration (embrace uncertainty)
- •Simonton-style ‘publish more’ vs. counterexamples like Gödel; fear of judgment can block output
- •Internalization often requires a demanding context—projects with real stakes, not just conversation
- •LLMs can accelerate shallow progress while enabling avoidance of the hardest cognitive work
