Anthropic can now track the bizarre inner workings of a large language model