How Is Compression Prediction?
Summary
The article surveys the equivalence between compression and prediction from information theory. It explains that the ideal payload length equals the cumulative log-loss under a probabilistic model, and discusses the role of the Kraft–McMillan inequality, Kolmogorov complexity, and Shannon entropy. It emphasizes that compression and prediction are tied only after the coding problem (object family, serialization, decoder contract) is fixed, and that the slogan 'Compression is Prediction' is true only at the level of code lengths, not in choosing representations or models.