The TLDR in the scaling RL to 10^26 FLOPs post is a proposal to find a way to do next-token prediction on the Internet using RL. How does Morris' arguments compare to those of the Lu's around finding the dual of RL?
The TLDR in the scaling RL to 10^26 FLOPs post is a proposal to find a way to do next-token prediction on the Internet using RL. How does Morris' arguments compare to those of the Lu's around finding the dual of RL?