Fork Around and Find Out Part 2: One Head does the Summing

Per LessWrong, researchers studying how a chess-playing AI represents knight forks have found a crucial piece of the puzzle: the network's ability to recognize knight forks—a tactic where a single knight move attacks two pieces—emerges at a specific computational layer. Using a technique called logit lens, they found the fork commitment snaps into focus after the model's fifth transformer block. Control experiments ruled out simpler explanations: the network isn't just identifying the best moves, but something more specific about the tactic itself. The team's evidence points toward a compositional representation—the model assembles forks from simpler components like checks and piece attacks rather than recognizing them as a single holistic pattern, though more investigation is needed to pin down the exact mechanism.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton