Learning Go with Eric Jang

TL;DR

• 1

  Go is about territory and captures, not just surviving stones: Eric Jang explains that stones die when all four orthogonal neighbors are surrounded, and that territory includes enclosed empty intersections too.

• 2

  Threats in Go work a bit like checks in chess: In the demo, one black move forces an immediate white response because ignoring it would let black capture a stone or even an entire group.

• 3

  The beauty of Go is 'lose the battle, win the war': Jang emphasizes that sacrificing a small group can be correct if it helps you gain more territory or win elsewhere on the board.

• 4

  Tromp-Taylor rules are what Go AIs train on because they are fully unambiguous: Unlike human play, where a suicidal move is usually illegal, Tromp-Taylor allows the move and then resolves it as dead, producing the same outcome without ambiguity.

• 5

  Human scoring depends on agreement, computer scoring does not: Humans effectively stop when both players agree the game is over and which stones are dead, while Tromp-Taylor counts stones and empty intersections algorithmically.

• 6

  A recent code bug came from an edge case in scoring contested territory: Jang uses a position where enclosed white stones look dead to humans, but under Tromp-Taylor some shared intersections score for neither side and others can still count for white.