Interview
David Silver: AlphaGo, AlphaZero, and Deep Reinforcement Learning | Lex Fridman Podcast #86
- Cryptocurrency development is projected to potentially redefine the nature of money.
- Reinforcement learning is anticipated to serve as a core component of future human-level intelligence systems and remains a necessary element for progressing beyond current human performance levels in game playing.
- The fundamental learning capability of neural networks is characterized as a true, universal property expected to improve without bound, with the expectation that future systems will view current complex reinforcement learning approaches as overly intricate compared to simpler ideas that will endure.
- Predictions regarding specific matches include a 4-1 scoreline where delusions were expected to occur in approximately one out of five games, and a conviction that deep learning systems without search capabilities would inevitably reach professional and world champion levels in Go once initial milestones were achieved.
- If AlphaZero were operated with greater computational resources, it is predicted to defeat previous systems by a margin of 100 games to zero, a process expected to continue indefinitely within a human lifetime, though the game of Go itself is expected to set a performance ceiling unreachable by any computational device constructed from the 10 to 80 atoms available in the universe.
- Future applications of reinforcement learning are expected to extend into diverse real-world domains such as chemical synthesis and quantum computation, with algorithms likely to be applied by others in unexpected ways to solve significant societal problems.
- To achieve clear answers regarding intelligence, well-defined problems are required, with the next level of advancement envisioned as systems able to solve goals more effectively than humans can, potentially accessing abilities previously thought exclusive to the human mind such as intuition and creativity.