Researchers at the A.I. company Anthropic claim to have found clues about the inner workings of large language models, possibly helping to prevent their...