project · released
Reflex, a transformer that plays 2048 without search
reflex looks at a 2048 board once and picks a move. no tree search, no rollouts. i designed and trained it. i'm Akshat Kushwaha, an ai systems engineer in india.
how it works
the board goes in as 21 tokens: 16 tiles, 4 move tokens built from what each move would leave behind, and a value token. eight pre-norm attention blocks score the moves, and it plays with one forward pass per move.
it has 14.5M parameters and learned by imitating an expectimax teacher across 460M of its moves.
how well it plays
over 10,000 fresh games it averages 164,395 points and builds the 2048 tile 96.6% of the time, at about 9 ms a move on a laptop cpu.
what failed
adding DAgger dropped play from 56k to 4k while every training metric still looked healthy. that was the most useful thing i learned on this project, and it is written up in the paper along with the rest.