Is Human Data Enough? With David Silver

This episode explores the concept of the "era of experience" in AI, advocating for a shift from relying on human data to AI systems learning through their own interactions and self-generated experiences. David Silver argues that while large language models have made significant strides by assimilating vast amounts of human-generated text, true progress requires AI to surpass human knowledge through independent discovery. He contrasts this approach with AlphaZero, which mastered Go without any human data, learning solely through self-play and reinforcement learning. Against the backdrop of AlphaGo's initial reliance on human data, Silver highlights the "bitter lesson" that human knowledge can sometimes limit AI's potential by hindering its ability to learn independently, referencing AlphaGo's move 37 as an example of AI creativity exceeding human understanding. More significantly, the discussion pivots to the challenges of applying reinforcement learning in less structured environments, where clear metrics of success are absent, and the potential for AI to define its own goals based on human input, while acknowledging the risks of unintended consequences and the need for careful alignment. The episode concludes with insights on the potential for AI mathematicians to transform the field and the broader implications of untethering AI from human data to achieve superhuman intelligence.