Dwarkesh interviews OpenAI researcher: backprop squeezing now, sample efficiency breakthrough next

PTrubey · x · 2026-09-30

PTrubey highlights Dwarkesh Patel's interview with an OpenAI researcher, articulating a key 1-2 year thesis: OpenAI and Anthropic are squeezing maximum efficiency from the existing backpropagation paradigm, while others tackle core ML problems — sample efficiency and continual learning — that could step-change the entire industry. Full video recommended.

Related event: OpenAI Researcher: Sample Efficiency and Continual Learning May Upend Backprop Paradigm(2 posts)→

Original post →

More from AGI Musings

AGI Musings channel →