Is human data enough

A conversation with David Silver, introducing the ideas in Welcome to the Era of Experience.

The point is that AI performs better as human involvement is reduced—data obtained from humans, human judgments such as RLHF, and so on. I found it interesting that algorithms obtained through meta-RL (using RL to create the RL algorithm itself) outperformed algorithms designed by humans. In other words, “excluding humans twice” made things better.