The fact that the best compression algorithms are, in the LLM setting, still catastrophically bad at prediction suggests that the overlap between compression and prediction is weak at best.
replyreply
This shows that prediction algorithms (AI models) are also very good at compression, but compression algorithms (like the ones used in gzip) are not likewise very good at prediction. Which is evidence that compression is necessary but not sufficient for prediction.
reply