Notes — Steven Pinker on Coordination, Common Knowledge, and the Retreat of Liberal Enlightenment

Source: Conversations with Tyler, Episode 255, 24 September 2025 (recorded 12 September 2025).

Four questions [Adler frame]

Q1 — What is it about? Pinker’s 2025 book When Everyone Knows That Everyone Knows… applied to social coordination, political authority, and the current state of liberal enlightenment. The technical core is a psychologist’s account of what happens inside human heads when common knowledge is generated — not the game-theoretic abstraction, but the cognitive mechanisms of recursive mentalising and their shortcuts (eye contact, public signals, Schelling points).

Q2 — How is it argued? Through examples at escalating scales: the coral spawning case (no cognition, pure coordination); the traffic circle (minimal recursion); eye contact (instant common-knowledge generation); currency and political authority (social-order scale). The key move is to distinguish common knowledge (the infinite recursion) from common salience (Schelling points) and from private knowledge (what triggers the initial defection from a regime).

Q3 — Is it true? The formal results (Aumann’s theorem, the coordination logic) are mathematically established. Pinker’s extension to human relationships (friendship, authority, romance as coordination games) is plausible and consistent with sociology and anthropology. The claim that liberal enlightenment is in retreat because it is counter-intuitive is a contested empirical claim — Pinker acknowledges the retreat but does not concede permanence.

Q4 — What of it? The framework helps explain: why bank runs are self-fulfilling rather than informative; why authoritarian regimes can collapse suddenly; why innuendo and tact are socially functional, not merely polite; why Taiwan’s diplomatic ambiguity is a rational equilibrium, not hypocrisy. The policy implication is that maintaining common knowledge of liberal norms requires active cultivation — it is not a natural equilibrium.

Glossary

Common knowledge: a proposition P is common knowledge in a group when everyone knows P, everyone knows that everyone knows P, everyone knows that everyone knows that everyone knows P, and so on, infinitely. The infinity is not merely technical: the difference between finite mutual knowledge and full common knowledge can determine whether coordination collapses.

Coordination game: a game in which the payoff to each player depends on aligning their action with other players’ actions. Currency, road rules, and political authority are all coordination games; their solution depends on common expectations, not on the intrinsic value of any particular convention.

Schelling point (focal point): a coordination equilibrium selected not by explicit agreement or common knowledge but by common salience — what pops into everyone’s mind. Named after Thomas Schelling. The selection criterion is not logic but human psychology; this is where game theory leaves off and ‘whimsy, caprice, even poetry’ take over.

Recursive mentalising: thinking about other people thinking about other people. ‘I know that she knows that I know…’ Human capacity for recursive mentalising (theory of mind) is the cognitive basis for common knowledge, but it is limited — minds start to spin after three or four iterations.

Aumann’s agreement theorem: rational agents with the same priors whose posteriors are common knowledge cannot agree to disagree. The mere announcement of posteriors (without sharing evidence) must produce convergence. Corollary: in a market with common-knowledge prices, speculative trade cannot produce systematic profit — a foundation for the case for index funds over active management.

Benign hypocrisy (Hanson-Simler): a common pretence maintained by mutual agreement to not-notice what everyone notices. Not an outright lie; closer to strategic ambiguity. Functional because it preserves institutions (authority, deterrence, law) that explicit acknowledgement would undermine.

Nuclear opacity: Israel’s policy of neither confirming nor denying possession of nuclear weapons. A worked example of benign hypocrisy at the geopolitical level: deterrence requires some suspicion; open acknowledgement would require Arab states to respond, triggering proliferation.

Strategic ambiguity (Taiwan): the US policy of neither recognising Taiwan’s sovereignty nor denying that it exists as a functioning state. Preserves China’s ability to maintain its territorial claim without requiring either a military response to explicit recognition or a full acknowledgement that the claim is unenforced.

Section notes

The headline claim: GDP would collapse without common knowledge. The mechanism: all economic coordination depends on beliefs about others’ beliefs. Pinker gives three cases —

Currency: the paper is worth something because we know others will accept it because they know others will accept it. Hyperinflation and bank runs are the unravelling of this recursive structure.
Political authority: enforcement cannot be total; authority depends on common knowledge of compliance. This is why demonstrations can topple regimes: once enough people see enough people seeing enough people oppose, the common-knowledge structure collapses.
Law: the rule of law depends on the common knowledge that the law will be enforced, which allows the police to tolerate strategic non-enforcement (the paper bag) without surrendering authority.

Coordination vs common knowledge: the traffic circle

Cowen presses the distinction: most everyday coordination does not involve genuine infinite recursion — the traffic circle involves perhaps one or two levels of mentalising, not infinite loops. Pinker’s reply: you’re right for routine coordination. What makes common knowledge interesting is the non-routine cases where small perturbations cascade. The knife-edge properties of formal models are a feature, not a bug: they describe the fragility of social institutions precisely.

Eye contact: the self-referential signal

Pinker’s claim is that eye contact is a unique common-knowledge generator because it is self-referential in the relevant sense: looking at the person looking at you looking at them. No explicit reasoning is required. This is consistent with evidence that eye contact is phylogenetically old and cross-culturally potent.

The coral example establishes the limiting case: a full-moon spawning event solves a coordination problem with no cognitive apparatus at all. The public signal + simultaneous sensing = coordination, without any beliefs about beliefs.

Schelling points: where game theory ends

Schelling’s own admission (1960) is significant: he wrote in an era when Grand Central was a dominant focal point, but acknowledged that Schelling points are time- and context-dependent. Pinker takes this further: what counts as a Schelling point for two specific people depends on the idiosyncratic history of their relationship — something that came up in conversation the night before, a joke between them. This is why ‘there doesn’t have to be any rhyme or reason other than human psychology.‘

Aumann’s theorem and index funds

The connection between the agreement theorem and the case for passive investing is rarely spelled out directly. Pinker makes it explicit: since prices are common knowledge in a public market, any private signal that a stock is undervalued is almost certainly already priced in. Rational agents cannot systematically exploit private information in common-knowledge price environments. This is the efficient-market hypothesis expressed in terms of common-knowledge epistemology.

Cowen tests the theorem against aesthetic judgements (chocolate, Hitchcock). Pinker’s reply draws the distinction: the theorem applies to descriptive beliefs (what is the case) with shared priors, not to preferences (what I like) where individual constitution is the prior. He is not willing to revise his dislike of chocolate on Tyler’s epistemic authority; he is willing (in principle) to revise his view of Hitchcock’s ranking if the criteria are external and shared.

Liberal enlightenment in retreat

Pinker’s diagnosis (2025): not gone, but in retreat. His mechanism: liberal enlightenment is counter-intuitive. Tribalism, deference to authority, conformity are the natural equilibria of human social organisation; liberal democracy, markets, and universal rights are ‘counterintuitive discoveries of organised rationality,’ like science. They require active maintenance.

The implication is that retreat is not caused primarily by liberal enlightenment being wrong, or by specific bad actors (though Pinker names Putin and Xi). It is caused by the cessation of the active maintenance that any counter-intuitive institution requires.

Against Cowen’s more pessimistic reading — that universalism is functionally gone and we will have a steady state of low-level great-power conflict below the threshold of war — Pinker argues that we should distinguish level of conflict from trend. Ukraine is not World War I. Low-level conflict that remains below the nuclear threshold is better than the previous baseline of periodic mass-casualty industrial war.

LLMs and linguistics

Pinker is discouraged but not dismissive. His assessment:

LLMs implicitly build in linguistic insights: hidden layers encode abstractions; attention heads search for non-adjacent regularities; token training approximates morpheme structure; embeddings approximate distributional semantics.
But formal linguistic theory (as in Linguistic Inquiry) has no direct counterpart in transformer architecture.
The training data gap is telling: 30,000-year equivalent vs a child’s two-year acquisition — suggests that the models are doing something substantially different from human language learning.
Confabulation and hallucination support the same inference: structural differences from human cognition, not merely scale differences.

Pinker does not take a position on whether LLMs vindicate Chomsky (by showing that deep structure can emerge from surface statistics) or refute him (by showing that no innate structure is needed). He treats his own intuitions as unreliable at the scale of hundreds of billions of parameters.