Analysis: from 2019 to 2025, gains in pretraining compute efficiency came mostly from data improvements rather than model improvements (Dwarkesh Podcast)

OpenAI claims an internal model, described as “significantly more capable than GPT-6 Astra,” solved the Navier-Stokes Millennium Prize problem using 10,000 concurrent agents running for 88 hours. Mathematician Tristan Buckmaster alleges the system leveraged his and Anthropic researcher Levent Alpöge’s unpublished work, though OpenAI says it cannot rule out that de-identified data from their usage may have contributed to model improvements.