Why Science Doesn’t Work the Way You Think — And Why the Future Will Be Even Stranger
Insights from a conversation between Michael Nielsen and Dwarkesh Patel
The Clean Story of Science Is a Fiction
Science is often taught as a linear process: hypothesis, experiment, falsification, progress. That story is false.
Consider the Michelson–Morley experiment. It is widely framed as the moment the ether theory was disproven, paving the way for relativity. Yet Albert A. Michelson himself continued to believe in the ether for decades after the experiment.
The implication is structural:
Evidence does not eliminate theories cleanly. Instead, it produces competing interpretations, patches, and ambiguity. Scientific progress emerges from navigating this ambiguity, not resolving it.
Evidence does not eliminate theories cleanly. Instead, it produces competing interpretations, patches, and ambiguity. Scientific progress emerges from navigating this ambiguity, not resolving it.
Science Often “Knows” Before It Can Prove
A deeper pattern emerges in high-impact breakthroughs: the scientific community frequently converges on correct ideas before definitive experimental validation.
Albert Einstein’s special relativity gained acceptance before direct experimental confirmation such as muon decay experiments in the 1940s. Meanwhile, Hendrik Lorentz had mathematically equivalent equations but a different interpretation.
The choice of Einstein over Lorentz was not driven by data. It was driven by conceptual coherence and aesthetic judgment.
Conclusion:
Science is not purely empirical. It is partly guided by taste, intuition, and theoretical elegance.
Science is not purely empirical. It is partly guided by taste, intuition, and theoretical elegance.
The Darwin Problem: Why Obvious Ideas Arrive Late
On the Origin of Species appeared in 1859—nearly two centuries after Principia Mathematica.
Yet natural selection appears simpler than Newtonian mechanics.
Why the delay?
Because Darwin’s theory required a missing prerequisite: deep time. Without geological insights from figures like Charles Lyell, evolution was not just unknown—it was impossible.
This reveals a non-obvious constraint on innovation:
Ideas do not emerge when they are simple. They emerge when the conceptual infrastructure exists.
Ideas do not emerge when they are simple. They emerge when the conceptual infrastructure exists.
AI Will Not “Solve” Science
Modern narratives suggest AI will automate scientific discovery. The reality is narrower.
Take AlphaFold. It is framed as an AI breakthrough. In practice, it rests on decades of experimental data from the Protein Data Bank—billions of dollars of accumulated scientific effort.
The core limitation is deeper:
- Any dataset supports infinitely many theories
- Scientific progress involves selecting among them
- Selection depends on insight, not just optimization
AI excels in tight feedback loops (e.g., coding). Science often operates in loose, delayed, or ambiguous feedback environments.
Result:
AI will accelerate local bottlenecks but will not resolve unknown unknowns.
AI will accelerate local bottlenecks but will not resolve unknown unknowns.
The Most Important Idea: The Tech Tree Is Path-Dependent
The dominant assumption: advanced civilizations converge on the same science and technology.
This assumption is likely false.
Examples:
- Computer science emerged from abstract mathematical logic—not engineering demand
- Quantum computing emerged only in the 1980s despite earlier theoretical readiness
- Phases of matter continue to expand beyond classical categories
Technology is not inevitable. It is contingent.
Different starting conditions, sensory systems, or environments would produce entirely different technological trajectories.
Implication:
An alien civilization would not just be more advanced—it would be different in ways that are difficult to anticipate.
An alien civilization would not just be more advanced—it would be different in ways that are difficult to anticipate.
Knowledge Is Not Universal—It Is Local
This leads to a powerful economic insight:
If two civilizations develop along different paths, their knowledge sets will be complementary, not redundant.
The gains from exchange would be enormous—not because one is superior, but because each has explored different branches of the possibility space.
This reframes progress:
Innovation is not just depth within a domain. It is breadth across unexplored domains.
Innovation is not just depth within a domain. It is breadth across unexplored domains.
Science Is a Social System, Not Just an Intellectual One
Scientific output is shaped by institutions, incentives, and norms.
In the 17th century, scientists like Galileo Galilei concealed discoveries using encoded messages to preserve priority. Today, open science and preprints dominate some fields.
Yet norms differ:
- Physics emphasizes rapid disclosure
- Biology historically emphasized secrecy
These are not natural laws. They are coordination mechanisms.
Conclusion:
The structure of credit systems determines the structure of knowledge production.
The structure of credit systems determines the structure of knowledge production.
Learning Is Not Consumption—It Is Struggle
A final observation cuts through modern knowledge culture:
Depth correlates with effort, not exposure.
Work that takes months embeds itself. Work that takes days evaporates.
The critical variable is not access to information, but the presence of a forcing function:
- A problem that cannot be solved passively
- An artifact that must be created
- A constraint that demands resolution
Without this, one accumulates familiarity—not understanding.
Synthesis
Three structural truths emerge:
- Science is not linear, empirical, or clean—it is interpretive and messy
- Progress depends on hidden prerequisites and long feedback loops
- The space of possible knowledge is far larger than what humanity has explored
The consequence is uncomfortable:
The current scientific worldview is not the inevitable endpoint of rational inquiry. It is one path through a vast, largely unexplored landscape.
Understanding this shifts the objective:
From optimizing within known frameworks → to identifying entirely new branches of the tree