Loading page…
His new essay argues that AI could accelerate its own development. The supporting evidence shows faster AI-assisted work, not autonomous systems replacing top researchers.
Listen to this story
Patel’s 2028 estimate rests on a specific handoff: AI tools would first need to automate the research work that improves AI, not merely perform bounded tasks faster. He says progress in mathematics and AI-assisted experiments has moved his median forecast forward, while putting a 25% chance on the milestone arriving after 2032. METR’s Anthropic assessment found faster benchmark-score gains largely attributable to AI assistance, but researchers still provided the key contributions and public evidence is limited by redactions—leaving autonomous research as an unverified leap.
Patel gives 2028 as his median for matching the best human AI researchers, sees 2027 as possible, and assumes labs do not intentionally slow development.
METR estimated Anthropic benchmark-score gains accelerated 1.9×, with a 95% confidence interval of 1.5–2.8×; it ruled out faster training-compute growth as the explanation.
Researchers told METR that AI assistance drove most gains, rather than fully autonomous contributions; Anthropic also redacted much of the evidence evaluators found most informative.
Dwarkesh Patel now expects AI to match the best human AI researchers in 2028, potentially setting off much faster progress in AI itself. In a November 13 essay, he argues that advances in mathematics and AI-assisted research have shortened his timeline. The forecast assumes labs do not deliberately slow development; it is not an achieved milestone.
Earlier this year, Patel says, he would have placed superintelligence at least the better part of a decade away. He now expects systems at least as capable as the best human experts across every field by 2030, potentially sooner. His argument turns on a narrower, preceding milestone: automating the work that makes AI better.
Patel’s starting point is work with checkable answers. If a system can test whether an answer succeeds, it can practice repeatedly and compare approaches. He argues that AI’s progress on well-defined mathematics problems shows how persistence and combining ideas across fields can produce results without first developing entirely new theories.
He sees a similar opening in AI research. Labs can test whether a new training method produces a stronger model or uses computing resources more efficiently. In his account, trying many combinations of existing ideas could be valuable even without a deep theoretical breakthrough. Different agents could also work independently on different parts of model development.
Once automated researchers match the strongest humans, Patel expects three sources of acceleration:
His central estimate for matching the best human AI researchers assumes no intentional slowdown by labs.
Patel says the research milestone could arrive next year.
He assigns substantial probability to a longer wait if important research work proves difficult to automate.
The strongest independent evidence he cites comes from METR’s November 5 assessment of Anthropic. After high-access work from August 12 through October 29, METR estimated that the company’s model benchmark scores had improved at 1.9 times their previous rate since October 2025. The estimate’s 95% confidence interval ran from 1.5 to 2.8 times.
METR judged that AI use likely explained most of the current acceleration, while stressing substantial uncertainty. It ruled out faster growth in training compute as the explanation. Crucially, researchers told METR that the gains came primarily from AI assistance, rather than fully autonomous contributions.
Patel also cites company disclosures showing heavier use. He recounts Anthropic’s statement that Claude led more than 26% of AI research and development work in August, up from under 1% in February. At OpenAI, he says, the heaviest-using tenth of researchers went from roughly $1,000 in daily internal coding-agent use early in summer to over $7,000 by late August.
Patel acknowledges that betting on the right research direction years ahead is not directly checkable. His forecast assumes that practice on narrower tasks, combined with broadly smarter models, will carry over to that judgment. He also recognizes that faster agents would still face limits on computing capacity for experiments.
METR’s evidence has another constraint: Anthropic redacted most of the material the evaluators found most informative. METR says this leaves the public unusually dependent on its judgment. Its findings concern well-specified tasks, with additional uncertainty about broader work. That is the boundary between the acceleration it measured and the research autonomy Patel predicts.
Loading discussion...
Join the conversation
Explain whether solving problems translates into choosing the right ones.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.