“Geoff Hinton's argument for language model understanding is that the ability to accurately predict the next word in a sequence requires the model to have developed a sophisticated internal model of semantics.”