Program and guest

A Dwarkesh Podcast video essay, featuring a monologue by host Dwarkesh Patel, running 44 minutes and 26 seconds. Original English title: The avalanche that hit math is coming for AI development.

Key points

  • The bottom line: AI may reach the level of the very best human AI researchers in just a few years, or even months; by 2030, possibly as soon as next year, AI could reach the level of the top human experts in every field. 0:21
  • Recursive acceleration is already happening: after spending three months inside Anthropic, independent evaluator METR concluded that Anthropic’s AI progress on clearly defined tasks is 2 times faster than last year’s, and that AI use in R&D may account for most of the acceleration. 1:07
  • OpenAI’s September blog: the 10% of researchers who use internal coding agents most heavily have seen their average daily usage rise from about $ 1000 at the start of summer to more than $ 7000 by the end of August; Anthropic said around the same time that Claude already leads 26% of its AI R&D work, while the figure was still below 1% in February. 1:31 1:48
  • The avalanche in mathematics: in 2023, models could not reliably handle even four-digit addition; in 2024, they could barely solve high-school math competition problems; in 2025, they won gold at the International Mathematical Olympiad; and a month ago, OpenAI solved a Millennium Prize problem in 88 hours. 3:17
  • Median year: 2028. This assumes AI labs push ahead at full speed, do not slow themselves down, and are not shut down by governments. He also uses simulated economies to illustrate the generalization capabilities of an intelligence explosion: AI could be placed in a simulated economy made up of millions of AIs that start companies together and experience hundreds of years of corporate competition to discover the best organizational structures. 8:27 19:33

Background and notes

This episode is a video essay by Dwarkesh Patel, running 44 minutes and 26 seconds, with 574 timestamped subtitle segments covering 0:00–44:23. The views and timestamps in the article are based on the original subtitles; each timestamp marks the start of the corresponding spoken segment. This article covers the opening argument (0:00–9:00) and the intelligence explosion section (19:33–21:30); the technical details discussed in the latter part of the program have not been organized point by point. The release time was approximately 04:11 on 2026-11-14 Hong Kong time, converted from the relative time shown on YouTube when retrieved, with hour-level precision.

Source links

Further reading

Related topics