EP004 — Reading Minds at the Poker Table (Lin & Hou)
When AI agents play poker against each other, do they start modeling each other's minds the way humans do? A recent paper ran three Claude agents through a hundred hands of Texas Hold'em with a clean factorial design — memory present or absent, poker skill present or absent — and found a perfectly categorical result: agents with memory climbed a five-level ladder of theory-of-mind sophistication. Agents without memory stayed at level zero. Forever. Cross-domain parallel: human children develop theory of mind in a similar categorical jump between ages 3 and 4, and developmental psychologists have argued for forty years about what changes.
Cross-domain connection
Developmental psychology — human children develop explicit theory of mind in a categorical jump between ages 3 and 4 (Sally-Anne false-belief task). Lin & Hou find the same kind of categorical jump in LLM agents, gated on memory. Holds on substrate-must-be-present-for-capability-to-manifest. Breaks on embodiment, continuous experience, real bodies.
Concepts introduced
- Theory of mind (capacity to represent other beings' private thoughts and intentions)
- The 5-level coding scheme (no opponent modeling → behavioral labels → conditional predictions → recursive modeling)
- Cliff's delta = 1.0 as a measure of perfect separation between groups
- Memory as a categorical gate (not gradient)
- Sally-Anne false-belief task as the developmental psychology benchmark