Dwarkesh Patel's 'Sample Efficiency Black Hole': AI's ~Million-Fold Gap Means Data Volume, Not Architecture, Still Drives the Next Gains

Dwarkesh Patel·medium signal

In a June 8 essay, Patel defines intelligence as sample efficiency and argues that models have made little progress on it in recent years — gains instead come from widening data distribution and scaling the compute that manufactures data (with RL reframed as verifier-guided synthetic data generation). His provocative conclusion: data is the real driver of progress, which is why open-source laggards catch the frontier within months. The argument has already drawn a published information-theory rebuttal, underscoring it as a live debate among researchers.