Definition
Plain language
A small AI trained to play the board game Othello that became a famous test case for figuring out what a model represents inside itself.
As stated in the literature
A model trained to predict legal Othello moves, used in interpretability research as evidence that a network builds an internal world model of the board; follow-up work found it tracked 'current player' versus 'opponent' rather than fixed colors, a cautionary tale about assuming the form of a learned representation.
Also called: Othello
Why it matters: It's a vivid reminder that a model may build a real internal model of its world, but in a form different from what we expect, so we must check rather than assume.
For example, researchers probing it found it wasn't tracking fixed black-and-white pieces but rather 'my pieces' versus 'your pieces,' overturning their first assumption about what it had learned.
Heard on the show
“There's a lovely cautionary tale they nod to, from a model trained to play the board game Othello.”Episode 098 — Finding Millions of Readable Concepts Inside a Real, Deployed AI Model