Dwarkesh Patel: AI Researchers by 2028, Superintelligence 2030

"It fucking solved a Millennium Prize problem in 88 hours." That line, pulled from Dwarkesh Patel's new Substack essay, is the argument in miniature. Patel is making the case that the trajectory math has just walked, from dunce to competition winner to open-problem solver in three years, is the preview reel for AI research itself.

He runs the sequence flat. 2023: cannot reliably add four-digit numbers. 2024: barely any high school competition problems. 2025: gold at the International Mathematical Olympiad. Earlier this year, open Erdős problems. A month ago, the Millennium Prize problem in 88 hours.

Patel's real move is that AI research will go faster than math, not slower. Pure math, he argues, demands exactly the capacity current models struggle with. He grants the worry by quoting Toby Ord: "While current AI systems are starting to automate mathematical proof, there is currently very little evidence that they can ask powerful questions." AI research does not carry the same bar. Labs reward grinding through random conjectures that make training more efficient, and Patel leans on Chris Olah's observation that "the transformer architecture was a natural combination of many ideas…taken to quite aggressive lengths."

The receipts are lab-sourced. Patel cites METR evaluators saying "Anthropic's overall rate of progress on well-scoped tasks is 2x what it was a year ago, and that a 'majority of this acceleration today is likely attributable to Anthropic's extensive use of AI to do R&D.'" He writes that "the heaviest 10% of AI users among OpenAI's researchers went from using around $1,000 a day of internal coding agents at the beginning of the summer to over $7,000 a day by the end of August." And he states that "Claude 'leads' more than 26% of Anthropic's AI R&D work," up from under 1% as recently as February.

The numbers carry the prediction. Patel puts his median guess for AI matching the best human researchers at 2028, superintelligence "by 2030, but plausibly as early as next year," and a "25% chance that it doesn't happen until after 2032." He quotes Ajeya Cotra for the aggressive bound: at face value, if AI is "nearly doubling the rate of AI capabilities progress" and the acceleration is smooth, "we would reach extreme superintelligence within a year." This is one writer's forecast on figures sourced to the labs themselves, and the essay never names which Millennium Prize problem it refers to. Our Anthropic feed this same week carried Amodei calling compute Anthropic's "deepest business dilemma".