AI researchers debate how close we are to recursive self-improvement – Dwarkesh Podcast

AI researchers debate how close we are to recursive self-improvement – Dwarkesh Podcast — Transcript

Market views

Bullish

AILABS— Distillation and continual learning prevent consolidation in AI model providers, allowing smaller labs to keep up with frontier labs by distilling capabilities daily and accessing the same training data.
ALTDATA— Chinese companies are using router services to collect valuable prompt distribution data from US frontier models, which is extremely useful for distillation and represents a strategic data advantage.
INFRA— Infrastructure and compute remain fundamental bottlenecks for AI progress, with labs being compute-constrained for the foreseeable future and hardware improvements (like GB200s) enabling larger model serving.

Not investment advice.