LLMS CAN INVENT THEIR OWN COMPRESSION

Summary from the TLDR AI newsletter

Models can discover their own compression tricks to pack more information per token.