Ethan Mollick's D&D Encounter Test: watching AI models fight mind flayers

eldonredwards · x · 2026-09-08

Ethan Mollick proposes the "Encounter Test," a nerdy AI benchmark: have a model simulate a D&D fight between two creatures and see how long it takes to break the rules. GPT-4o did best; Gemini was cute with similar outcomes.

Two years later he re-ran it with a stronger model ("GPT-6"): its mind flayer even got a "101-bone skeletal rig." He built Underdark Duel, a reproducible web page logging every roll, with adjustable encounters and downloadable records. Under 5E 2014 rules he found no major errors, though the mind flayer's tactics could be more creative.

Related event: Ethan Mollick Proposes 'Encounter Test' Benchmark Using D&D Battles(2 posts)→

Original post →

More from Fun

Fun channel →