Dwarkesh Patel Says He Now Expects Superintelligence by 2030, as AI Starts Automating Its Own Research
The avalanche that hit math is coming for AI development.
Context
One of AI's most followed interviewers now argues that AI could match the best human AI researchers within a couple of years, a view that shapes how investors, labs and policymakers read the pace of progress.
Dwarkesh Patel hosts the Dwarkesh Podcast, known for long interviews with lab leaders and researchers. On November 13 he published an essay titled The avalanche that hit math is coming for AI development, with a 44 minute narrated video. The subtitle reads: Why I now expect superintelligence by 2030.
Deep dive
Patel writes that earlier this year he would have said superintelligence was at least the better part of a decade away, but that he has updated his views a great deal over the last three months. He now believes AI systems as good as the best humans at AI research are only a couple of years away, and possibly only a couple of months away. He argues that automating AI research will compress so much progress into a single year that, by 2030 and plausibly as early as next year, AI will be at least as good as the best human experts in every field. He calls this a wild claim that he himself did not believe a couple of months ago.
His case leans on evidence that AI already speeds up frontier AI work. He cites METR, which he says concluded after three months embedded inside Anthropic that Anthropic's rate of progress on well scoped tasks is twice what it was a year ago, with most of the acceleration likely due to its use of AI for research. He also cites an OpenAI post saying its heaviest 10% of internal users went from about $1,000 a day of coding agent use to over $7,000 a day by late August, and an Anthropic post saying Claude leads more than 26% of its AI research work, up from under 1% in February.
The essay's title points to mathematics, where Patel says AI is now resolving long standing problems at a fast pace, and he argues AI research is the next verifiable domain to follow. He estimates that automated research could deliver three to six years of AI progress in a single year. These are Patel's forecasts and arguments, not measured results.
What to expect next
Watch how researchers and lab leaders respond to the essay, since Patel says he has been debating this question on and off his podcast. The evidence he cites, from METR's work inside Anthropic to OpenAI's and Anthropic's own reports on AI use in research, will be tested by the next model releases. If AI research automation speeds up as he expects, it should show up in shorter gaps between frontier models over the coming year.