Korva is a scroll of the internet’s best ideas.
Beaten by 3-year-olds, he trained AI
Most chess engines, AlphaGo and AlphaZero among them, are hybrids: part neural network, part traditional programming and search. An engine built entirely from a neural network is still an unsolved problem in the research literature.
Amjad Masad, the CEO of Replit, had never been a researcher and didn't know how to construct a neural network from scratch. He had always been into AI theory, though, and AI had gotten really good at software. So he set himself a challenge: could it do other things, like train a fully neural-network chess engine?
He trained one, and it has an Elo score of 1300 on Lichess, and 1700 when it plays against humans. Working with AI feels to him like having a genius intern who is sloppy and bad at ideas. It has to be directed, and with direction he can do anything.
How does a traditional chess engine pick a move?
Traditional engines such as Stockfish search a tree of possible moves, using alpha-beta pruning to skip branches that cannot beat a move already found. At the end of each branch, an evaluation function scores the position. Stockfish added a small neural network to that evaluation step in version 12, released in 2020, called NNUE. The search itself remained ordinary programmed code.
How did AlphaZero learn chess without human games?
AlphaZero was given only the rules of chess and then played against itself, adjusting its network after each game based on who won. DeepMind's 2017 paper reported that this self-play training ran for about nine hours before the match against Stockfish. Unlike a pure network, it still paired its neural network with Monte Carlo tree search, which explores promising lines of play. The same method was applied to Go and shogi in that paper.
Why do Lichess ratings differ from other rating pools?
Lichess calculates ratings with Glicko-2, a refinement of the Elo system devised by the statistician Mark Glickman. A rating has no absolute meaning: it only describes how a player performs against the other players in the same pool. A rating earned among bots, or among mostly beginners, therefore cannot be compared directly with one earned in another pool. The Elo system itself was created by Arpad Elo and adopted by the international chess federation FIDE in 1970.
More from Casey Newton
