Video essay "Recursive self-improvement has started," published November 13, 2026

Dwarkesh Patel, the host of the Dwarkesh Podcast, has sharply shortened his forecast for superintelligence. In a video essay uploaded November 13, he said that earlier this year he would have put the arrival "at least the better part of a decade" away. After three months of debate and of watching AI progress in verifiable domains, he now believes the world is "only a couple of years away, and possibly only a couple of months away" from AIs as good as the best humans at AI research. He expects AIs at least as good as the best human experts in every field by 2030, and plausibly as early as next year. He called this "a wild claim," and one he did not believe himself a few months ago.

Evidence that AI is already accelerating AI progress

Patel anchors the argument in data from the labs. METR, the independent evaluation organization, spent three months embedded inside Anthropic. It concluded that Anthropic's rate of AI 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."

OpenAI published a September blog post showing that its heaviest 10% of AI users among researchers went from about $1,000 a day of internal coding-agent usage at the start of summer to over $7,000 a day by the end of August. Anthropic's own September post reported that Claude now leads 26% of the company's AI R&D work, up from under 1% as recently as February.

Mathematics as the template for AI research

The biggest driver of Patel's update has been the speed of improvement in AI mathematics. He noted that in 2023 models could not reliably add four-digit numbers, in 2024 they could barely solve high school competition problems, and in 2025 they won gold at the International Math Olympiad. Earlier this year they began solving niche open problems, and a month ago OpenAI solved a Millennium Prize problem in 88 hours. Patel said hundreds of results that top mathematicians spent decades pursuing are now being produced "as side quests to hype up future model releases."

He acknowledged the counterpoint from Toby Ord, who wrote that there is "very little evidence that they can ask powerful questions or find profound new fields of mathematical enquiry." Patel's retort is that the same sentence also means AIs solved a Millennium Prize problem without those abilities. He argues the skew toward verifiable work matters more for machine learning than for mathematics. Mathematics ultimately prizes new theory, whereas AI labs care about what empirically makes training more efficient or models smarter, and the last decade and a half of deep learning progress has not required much new theory.

He illustrated the point with the transformer. Citing conversations with Sholto Douglas and a quote from Chris Olah, he said none of its core pieces surprised active researchers. The architecture only proved clearly superior once half a dozen tweaks were combined correctly, which took research taste plus a large number of small experiments. That, in his view, is the persistent combinatorial search at which AIs have excelled in mathematics, and "the kind of thing that Claude 7 could have invented" if asked to make attention work better.

Timeline: 2028 median, with a fat tail on either side

Patel's median for automated researchers matching the very best human AI researchers is 2028, assuming labs go full steam ahead and are not slowed by themselves or by government. He gave at least a 25% chance that it does not happen until after 2032, because important parts of AI research may prove hard to automate. At the other end, he said some of the best-calibrated researchers he knows expect the milestone as early as next year.

He noted that 2027 is also the straightforward implication of METR's finding fed into a takeoff model. He quoted Ajeya Cotra: "If we take it at face value that AI is now nearly doubling the rate of AI capabilities progress and assume that acceleration from AI is smooth, this naively implies that we would reach extreme superintelligence within a year." Patel's median is not 2027 only because he is uncertain how accurate such modeling is.

Why automated researchers would be faster, better and more numerous

Patel gave three reasons progress accelerates once the milestone is reached. First, AIs already work at about 10 times human speed and would work around the clock, so each automated researcher would do roughly 50 times more work per calendar week than a human. He cautioned that compute for experiments would not speed up, so returns would diminish steeply, but he still expects algorithmic progress to run many times faster.

Second, quality. He estimates around 80% of algorithmic progress at a frontier lab comes from its 50 top researchers, which he said explains why Zuckerberg was willing to pay hundreds of millions of dollars for top talent. In an automated lab every researcher would be at that level.

Third, quantity. Patel argued AI research resembles formalizing a theorem more than writing a novel. He pointed to an animation of many Claudes jointly formalizing Fermat's Last Theorem across thousands of modular subtasks, each locally checkable. AI labs already split work across teams on data filtering, synthetic RL environments and attention kernels. He acknowledged steep diminishing returns, noting that "having 100,000 collaborators doesn't mean that you'll go 100,000 times faster, probably not even 100 times faster," but said it is implausible that orders of magnitude more labor would not yield at least a couple times faster progress.

Spark or software-only singularity: a coin toss

Patel framed the key question as whether the self-improvement loop outruns the growing difficulty of finding new ideas. He noted that AI companies already have to quadruple R&D compute and double their workforce every year to maintain a constant rate of progress. If AI improves faster than ideas get harder to find, the result is a "software-only singularity," in which, without added compute, the swarm reaches within months a superintelligence that "scrapes against the physical limits of our current chips."

If ideas get harder faster, the result is what he calls a "spark": roughly three to six years of progress at the current rate compressed into one calendar year, after which things "simmer down" until more compute is built. He gave each scenario 50%.

Even the spark case, he said, would feel extreme. He compared a four-year jump to going from GPT-3.5, which could write a passable limerick but was "mostly a useless toy," to a hypothetical GPT-6 "Astra" that builds whole video games from scratch and produces Fields Medal-worthy contributions at a rate that desensitizes observers. Starting in 2027 or 2028, he said, such a leap would yield a mind vastly superhuman in even slightly verifiable domains. He listed designing the next generation of AI accelerators better than Nvidia, earning more than the current human versions of Citadel and Jane Street given equal market access, and finding a room-temperature superconductor if given remote access to a materials lab.

How superintelligence gets good at hard-to-verify real-world work

Patel addressed the obvious objection that a datacenter-trained system cannot match experts whose edge comes from years of real-world experience, in areas such as building businesses, winning court cases, trading or running campaigns. He does not claim the answers can be baked into model weights. His claim is that scaled-up RL will give AIs the general skills to ramp up quickly in new contexts, such as seeking information, planning, persisting through obstacles, tracking what they have learned and collaborating.

He expects automated engineers to build tens of thousands of times more RL environments than exist today, far larger and more open-ended, "each to today's video games what a Bugatti is to a horse and buggy." Examples include dropping an AI into a simulated economy of millions of other AIs to co-found a company, or running historical backtests. One such task would ask whether an AI trained on a corpus ending in 1700 can derive general relativity, and another whether it can predict the Cuban Missile Crisis from newspapers up to 1962. Automated researchers would also improve the learning algorithms themselves, particularly online learning and sample efficiency....