1. A New Direction: From Human Data to the Era of Experience

David Silver proposes that AI must move beyond the "Era of Human Data" to the "Era of Experience" — where AI generates its own data through world interaction.

2. AlphaGo and AlphaZero: Learning Beyond Human Data

AlphaGo initially used human professional Go data but improved through self-play. AlphaZero used no human data at all — achieving superhuman play in Go, chess, and shogi. Move 37 against Lee Sedol was a creative move humans wouldn't conceive even once in 10,000 attempts.

3. The Power of Reinforcement Learning

AI learns from rewards (+1 for wins, -1 for losses). AlphaZero's simple algorithm — random moves, game results, policy/value updates, repeated — produced the world's best player.

4. AI's Potential Beyond Human Data

"AI dependent on human data can only reach human level. True innovation happens when AI learns independently and discovers new things."

5. AlphaProof: New Frontiers in Mathematics

AlphaProof proved mathematical theorems without human proofs, achieving silver-medal level at the International Mathematical Olympiad.

6. Challenges and Future

Applying reinforcement learning to the real world (without clear success criteria) remains challenging. Safety, ethics, and alignment with human values require careful design.

7. Conclusion

"If we truly want superhuman intelligence, it's time to move beyond human data."