Raw Content
“Seeing Science Like a Language Model” - Article Summary
Core Argument
Dan Shipper argues that language models are revealing fundamental limitations in how centuries of scientific thinking have approached truth. The traditional reductionist scientific method—breaking complex systems into simple, predictable parts—has reached its limits in understanding phenomena like human behavior and complex systems.
Key Points
Historical Context: Shipper opens with the observation that our current scientific worldview (Copernican, Newtonian) represents a cultural shift from intuitive perception. Just as people once found heliocentrism preposterous, we now accept ideas contradicting direct experience through trust in “the science.”
The Replication Crisis: Psychology exemplifies these limitations. Research built on linear regression assumes straightforward relationships between variables, but human behavior is contextual and nonlinear. Small changes yield dramatic shifts; major interventions produce nothing.
The Real Problem: According to psychologist Tan Yarkoni, the issue isn’t merely replication failure but generalizability. Scientists overstate how universal their findings are. Better replication would require duplicating the entire original context—researchers, environment, subjects, interventions.
“Hard-to-Vary” Explanations: Physicist David Deutsch’s concept defines scientific progress: explanations functioning as precisely engineered machines where every component serves essential purposes.
The Core Insight: Shipper suggests language models demonstrate that some truths resist the reduction and explicit categorization that enabled scientific progress. This presents a fundamental challenge to reductionist thinking itself.