Pedro Domingos: Calling LLM Training "Compression" Is a Fallacy

UW professor Pedro Domingos argues in a dispute with Christian Szegedy that minimizing cross-entropy is not equivalent to maximizing compression, and that calling LLM's maximum likelihood training "compression" is a harmful misconception.

2026-09-02 ~ 2026-09-02 · 2 related posts

• Pedro Domingos: Minimizing Cross-Entropy Is Not Maximizing Compression — pmddomingos · 2026-09-02
• Pedro Domingos: Calling likelihood maximization 'compression' is a travesty — pmddomingos · 2026-09-02