[singularity2001]

Superhuman chess engines are now trained just from one bit reward signal: win / lose. This says absolutely nothing about the complexity that the model develops inside. They even learned the rule of the games just from that reward.