LeCun calls Anthropic moralising and restates the case against next-token prediction

Yann LeCun, a Turing Award winner, wrote on X that Anthropic is moralising to the point of exasperation, after an exchange about usage restrictions that block some creative and learning uses of its tools. In a separate post he recalled older results showing that nonlinearities which cannot produce zero-mean outputs give badly conditioned loss functions.

The posts arrived alongside his paper with Randall Balestriero, which argues that next-token prediction is an inefficient route to representations shared across modalities. Both are LeCun’s statements of his own position.