Dwarkesh Patel: the avalanche that hit math is headed for AI labs
Dwarkesh Patel's new essay puts a date on the thing every lab keeps hinting at. AIs as good as the best humans at AI research are only a couple of years away, the essay argues, and possibly only a couple of months. The median is 2028, with a 25% chance it slips past 2032 and a real chance it lands in 2027. After that, experts in every field by 2030, and plausibly as early as next year. Published Nov 13: https://www.dwarkesh.com/p/rsi
The argument runs through math. Current AI systems have solved a Millennium Prize problem, the essay notes, and math fell because answers can be checked. AI research, Patel argues, is checkable enough too, because labs run on empirical recipes rather than theory. Either the loss went down or it did not.
The evidence pile is the useful part, and all of it is attributed. METR, after three months embedded at Anthropic, found the rate of progress on well-scoped tasks is twice what it was a year ago, most of it likely from AI doing AI R&D. OpenAI's agent spend for a top-10% researcher went from $1,000 a day in early summer to more than $7,000 a day by the end of August. Anthropic says Claude leads more than 26% of its AI R&D work as of August, up from under 1% in February. The essay also estimates today's AIs work about 10x human speed, about 50x more work per calendar week, and that roughly 80% of a lab's algorithmic progress comes from its top 50 researchers. Thomas Larsen of AI Futures is thanked as a co-researcher, so read it next to the AI Futures timelines update from earlier this week.
That last number is the hinge. If 50 people drive most of the progress, you do not need to automate a whole lab, only replace or multiply those 50. The spend curve from $1,000 to $7,000 a day says labs already act as if that is happening.
The weak spot is the jump from checkable to automatable. Research taste, picking which experiment to run, is exactly the part no benchmark grades yet. Still, this is the clearest single document on why people inside labs talk about 2028 with a straight face. Treat the dates as Patel's bet, and the dollar figures as the thing to watch.
← Back to all articles
The argument runs through math. Current AI systems have solved a Millennium Prize problem, the essay notes, and math fell because answers can be checked. AI research, Patel argues, is checkable enough too, because labs run on empirical recipes rather than theory. Either the loss went down or it did not.
The evidence pile is the useful part, and all of it is attributed. METR, after three months embedded at Anthropic, found the rate of progress on well-scoped tasks is twice what it was a year ago, most of it likely from AI doing AI R&D. OpenAI's agent spend for a top-10% researcher went from $1,000 a day in early summer to more than $7,000 a day by the end of August. Anthropic says Claude leads more than 26% of its AI R&D work as of August, up from under 1% in February. The essay also estimates today's AIs work about 10x human speed, about 50x more work per calendar week, and that roughly 80% of a lab's algorithmic progress comes from its top 50 researchers. Thomas Larsen of AI Futures is thanked as a co-researcher, so read it next to the AI Futures timelines update from earlier this week.
That last number is the hinge. If 50 people drive most of the progress, you do not need to automate a whole lab, only replace or multiply those 50. The spend curve from $1,000 to $7,000 a day says labs already act as if that is happening.
The weak spot is the jump from checkable to automatable. Research taste, picking which experiment to run, is exactly the part no benchmark grades yet. Still, this is the clearest single document on why people inside labs talk about 2028 with a straight face. Treat the dates as Patel's bet, and the dollar figures as the thing to watch.
Comments