David Silver, a principal research scientist at DeepMind, on AlphaGo, AlphaZero, and MuZero, applying reinforcement learning to real world problems, and more
David Silver of DeepMind, who helped create the program that defeated a Go champion, thinks rewards are central to how machines—and humans—acquire knowledge.
Context & Ripple Effects
This interview lands at the end of a four-year arc that Silver himself built. The lineage runs from the self-taught Go agent that beat its own predecessor 90% of the time through AlphaZero mastering chess, shogi, and Go without human games, to [[a:961419|MuZero, which plans without even knowing the rules and is now working on YouTube video compression]].
What makes the piece matter is the framing: Silver argues rewards — not imitation of human data — are the core mechanism of learning for machines and humans alike. That claim converts the game-playing results from curiosities into an argument that reinforcement learning is a general method, with MuZero's compression work as the first concrete evidence.
First-order effects
- Silver's reward-centric thesis gives DeepMind a public rationale for pushing beyond board games, and MuZero's YouTube compression project is the named proof point that the lab's research line is already attached to a production system.
Second-order effects
- If learning-from-rewards displaces learning-from-human-examples as the default recipe, other labs face pressure to build self-play and rule-free planning into their own pipelines rather than relying on curated datasets — a capability race set by DeepMind's published sequence of results.
Third-order effects
- A field organized around rewards rather than human data shifts the bottleneck from collecting examples to specifying objectives — meaning whoever defines the reward function increasingly determines what the resulting systems optimize for, a structural question for governance as these methods move into real-world deployment.
The trend: Reinforcement learning is migrating from closed games with known rules toward open real-world tasks, with DeepMind's AlphaGo-to-MuZero progression marking each step of the retreat from human-provided knowledge.