"Model designed to guess next tokens from a sequence, based on its training data, fails to guess next tokens from sequences not similar to its training data."

What's the news?