I thought that was about LLMs being trained on compressed data. But I might be thinking about a different paper.