AI researchers are using techniques inspired by neuroscience to study how language models work — and to reveal how perplexing they can be.