Over recent weeks on Hacker News, the claim that compression is prediction has been debated. Luca Lombardo analyzes this through 3Blue1Brown videos on entropy and an ngrok article on arithmetic coding. The core insight: probabilistic models assign conditional probabilities, entropy coders convert these to bits.

For sequence x₁:n with model Q, ideal payload length is −log₂Q(x₁:n) = Σ−log₂Q(xᵢ|x<ᵢ), which equals cumulative logarithmic loss. Lombardo confirms the mathematical equivalence, noting its roots in Shannon's information theory. However, compression involves more than model application—it requires agreement on representation, alternatives, model availability, and decoder operations.

The real question is what must be fixed before equivalence applies and what costs remain unmeasured.