"... lets you trace the outputs of language models back to their full, multi-trillion-token training data in real time."