DeepMind’s algorithm that learned to play chess, Go, and Atari games using one general system was MuZero. Announced in 2020, MuZero combined model-based planning with reinforcement learning while learning only the aspects of the environment needed for effective decision-making.
Earlier DeepMind systems had important but narrower achievements. AlphaGo specialized in Go, while AlphaZero learned chess, shogi, and Go from the rules of those games. MuZero differed by not being given the complete rules or an explicit simulator for the environments it mastered.
The system learned an internal model of rewards, actions, and relevant future observations, then used tree search to choose actions. “One general system” does not mean MuZero was a human-level intelligence in every setting; its demonstrated domains were selected games. Its significance was the combination of planning and learned models across very different tasks.