Dwarkesh Patel Learns Go
TL;DR
Go is about territory, not just captures — the explainer starts with the core mechanic: black moves first, stones live on intersections, and you win by controlling territory plus surviving stones, not by greedily taking every piece.
Capturing in Go works by cutting off all four orthogonal neighbors — Dwarkesh is walked through the idea that stones die when they lose their “oxygen,” with the teacher explicitly correcting him that diagonals do not matter, only the cross-shaped connections do.
A local loss can still be the right move if it wins you the board — the clip lands on Go’s signature idea that you can “lose the battle but win the war,” as small groups get sacrificed for larger strategic advantages elsewhere.
Rule sets matter more than casual players realize — the speaker contrasts Chinese and Japanese rules with Tromp-Taylor rules, noting that Tromp-Taylor is fully unambiguous and is what Go AIs train and score against.
Human scoring and AI scoring diverge in edge cases — humans settle dead groups by mutual agreement (“I think the game is done”), while Tromp-Taylor counts stones and empty intersections algorithmically, even in positions a human would treat as already lost.
This wasn’t just a game lesson — it was setup for talking about neural nets — the speaker says explicitly that how humans resolve ambiguous endgames “will actually map later to how we think about the deep neural network,” framing the Go lesson as groundwork for AI discussion.
Summary
Learning the board: stones, territory, and “oxygen”
The clip opens with the most stripped-down explanation of Go: black goes first, players place black and white stones on intersections, and the goal is to control as much territory as possible. The teacher uses a vivid metaphor — a surrounded stone is “cut off from oxygen” — to make captures intuitive instead of abstract.
Dwarkesh gets the first key correction
As they play a few sample moves, Dwarkesh tries to infer the logic and guesses diagonals might determine whether a stone is trapped. He gets a clean correction: not diagonals, but the cross-section — the four orthogonal neighbors — which makes the whole capture mechanic click.
Checks, forced moves, and reading ahead
The teacher compares a threatened stone to “a check in chess,” because sometimes you must respond immediately or lose it on the next move. That turns into a short tactical sequence where Dwarkesh realizes a little block of stones is basically doomed, and the teacher confirms it: yes, you should just assume that group is gone.
The beautiful part of Go: lose the battle, win the war
From there, the lesson zooms out from tactics to strategy. The teacher says this is what makes Go beautiful: you can let your opponent capture some stones if it gives you better position elsewhere, a classic “lose the battle but win the war” dynamic that gets even richer on larger boards.
A scoring edge case that broke the code
Then the clip pivots to a concrete board formation that recently exposed a bug in the speaker’s code. Looking at a surrounded white group, Dwarkesh naturally assumes white controls the interior, but the teacher explains that black actually owns that area because white is enclosed and has no realistic path to life.
How humans end a Go game
This leads into a surprisingly human description of scoring: players more or less agree that the game is done, then agree which stones are alive or dead, and if they disagree, they keep playing. The speaker even frames that agreement as both humans’ “value function” reaching consensus, which is a very AI-flavored way to describe an old board-game ritual.
Tromp-Taylor rules: what computers need from Go
Finally, the speaker explains why AIs use Tromp-Taylor rules: they’re completely unambiguous and can be resolved algorithmically. Under Tromp-Taylor, you count controlled stones and empty intersections not touched by the opponent, which can produce results that feel odd to humans — for example, white may still get points in an area a person would say is obviously doomed — and the game ends only by resignation or two consecutive passes.
Was This Useful?
Share
Keep Reading
Make Alcreon Yours
Tune your feedFive quick questions, and the feed ranks what matters to you first.Or just get notified
The weekly Echo. Signal worth keeping in your inbox.
Every new piece, announced on X.
Read Next
See all
Playbook
The Retirement Email Isn't a Warning
Model retirements now arrive every few weeks; the config-eval-rehearsal loop turns each deprecation email from a fire drill into an afternoon swap.
Playbook
The Cheapest Model That Passes
OpenRouter lists 400 models behind one API. The fix for choosing isn't a better leaderboard, it's a four-step protocol that ends in a real eval.
Playbook
Cheap Models, Hard Tasks
Most agent workflows route every step to the frontier model by default. The bill scales with how chatty the agent gets, even when most steps don't need that brain.