Pedro Domingos: Calling likelihood maximization 'compression' is a travesty

pmddomingos · x · 2026-09-02

Pedro Domingos argued in a reply that likelihood is maximized by memorizing data, which is largely what LLMs do. He stated that calling this process 'compression' is a travesty, challenging the notion that any maximum-likelihood model constitutes a compressor.

Related event: Pedro Domingos: Calling LLM Training "Compression" Is a Fallacy(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →