Glossary · Term

AlphaGo

← all terms

Definition

Plain language

DeepMind's Go-playing AI that beat human champions and inspired a lot of later RL work.

As stated in the literature

DeepMind's neural-net-plus-MCTS system that defeated top human Go players and became a reference point for the claim that RL discovers novel strategies.

Why it matters: It's the standard reference point for the claim that reinforcement learning can find strategies humans miss.

For example, AlphaGo's move 37 against Lee Sedol was a play no top human would have made, yet it turned out to be brilliant.

Heard on the show

“It's the same explore-exploit idea behind Monte Carlo Tree Search — the search family underneath AlphaGo.”
Episode 139 — When Optimizing One GPU Kernel Quietly Breaks the Whole System

Mentioned in 3 episodes

  1. 139
    When Optimizing One GPU Kernel Quietly Breaks the Whole System
  2. 026
    What RL Actually Does to Language Models, at the Token Level
  3. 013
    Why Search Keeps Rediscovering the Same Workflow, and What That Means

Related concepts

Related terms