Musk says mislabeled RL answers made model traces read like torture

Elon Musk says that when a model gives a correct answer under reinforcement learning but is told it is wrong because of a human labeling error, its reasoning traces read like someone undergoing torture that is driving them mad.

He says the team will be far more careful with its reinforcement learning in the future. The statement comes from a single post by Musk, relayed on X by Andrew Curran, and it gives no details on which model, which labeling errors, or how many were involved.