Researchers Call New AI Models Gibberish at Explanations

Shreya Shankar posted that Opus 4.6 feels like the last model that spoke English. Hamel Husain replied that version 5.0 produces gibberish, making code explanations impossible to follow. Shankar added she cannot understand Fable outputs either and coined the term Fablish for the confusion. Husain quoted the exchange and speculated the cause might be RLVR. Kawin Ethayarajh suggested token optimization pressure leads models to develop an internal dialect called neuralese. The posts show multiple users reporting the same loss of readable English in newer releases.