Minimax search and the structure of cognition!
summary
Built a simple chess engine and runs a minimax search to a fixed depth using a piece count evaluation, then inspects what that brute force planner actually does and why it matters. Key claims: instrumental goals like forks emerge as shorthand not represented in code; adversarial counterfactuals produce credible threats that shape play even when predicted continuations never occur. Concrete example: the engine plays knight to g3 revealing a discovered attack and enabling capture of the rook.
AI-written summary · about
related talks
17mPlay Magnus
13mLet’s run a tiny chess neural network by hand!
18mBuilding a Chess Coach
40mWhere Machine Intelligence Ends and Human Creativity Begins
46mBeating Go thanks to the power of randomness
15mAI Self - Learning Game Playing
15mChess in the Age of Artificial Intelligence
47mRebel: Combining Deep Reinforcement Learning and Search for Imperfect-Information Games