Podcaster Dwarkesh Patel now puts his median year for AIs that match the very best human AI researchers at 2028. In an essay published on 13 November he wrote: “If you asked me earlier this year how far away we are from superintelligence, I’d probably have said at least the better part of a decade”.
Update, 14 November 2026, 20:00 UTC: a Meta researcher proposes a different test for what a model should keep learning. David Lopez-Paz of FAIR at Meta published an essay on arXiv titled “Against All Likelihood” (2611.09385, dated 6 November). It argues that maximum likelihood, “the dominant training principle for large language models”, is “neither necessary nor sufficient for truth”, and that a statement’s truth should be judged “by its invariance across relevant transformations of the environment”, drawing on Nozick, Wigner, Poincaré and Cassirer. It sketches a test-time procedure in which a model generates different perspectives or framings of a question, answers under each, and extracts what stays the same across the answers. On continual learning, the area Patel names as the main risk to his timeline, the essay suggests “reframing continual learning as learning invariances across a different kind of environment, such as unrelated groups of users”, and says that “if continual learning is the problem of deciding what new data is worth learning from, statistical invariance may serve as the guiding principle.” The essay is a position paper: it reports no experiments or benchmark results.
He still gives “a 25% chance that it doesn’t happen until after 2032”, which leaves 75% on the milestone arriving by 2032, by our arithmetic. He wrote that he had “updated my views a ton” over the last three months.
Business Pill · The Model and the Harness
A short Business Engineer explainer on the difference between an AI model and the harness around it. It teaches the general idea only and does not discuss Patel or his essay.
The key insight: As we read the essay, Patel’s update rests on others’ measurements rather than new data of his: independent evaluator METR’s finding of a 2x progress rate at Anthropic, and the labs’ own disclosures of OpenAI’s jump to over $7,000 a day of agent use for its heaviest researchers, and Claude leading more than 26% of Anthropic’s R&D work, up from under 1% in February.
Patel wrote that “we’re only a couple of years away, and possibly only a couple of months away, from AIs that are as good as the best humans at AI research”. He called it “a wild claim, and it’s one that I myself didn’t believe even a couple of months ago”.
His median comes with a condition: “assuming the labs don’t intentionally slow down”. He added that “we could hit this milestone in 2027”.
He also wrote that automating AI research would bring, “by 2030, but plausibly as early as next year”, AIs “at least as good as the best human experts in every field”.
The probabilities Dwarkesh Patel states in his 13 November 2026 essay: 25% that AIs matching the best human AI researchers arrive only after 2032 (75% by 2032 is our arithmetic), and even odds between a “spark” and a “software-only singularity” once AI research is automated.
The Evidence He Cites
Patel cites METR’s finding, after three months embedded at Anthropic, that the lab’s rate of progress on well-scoped tasks is “2x what it was a year ago”; our piece on the METR report covered it.
He cites OpenAI’s figure that its heaviest 10% of AI users among researchers went from “around $1,000 a day of internal coding agents” to “over $7,000 a day by the end of August”, about 7 times in one summer by our arithmetic.
And he cites Anthropic’s report that Claude “leads” more than 26% of its AI R&D work, “up from under 1% as recently as February”. Patel notes that the fully autonomous share, by Anthropic’s definition, “is still 0%”.
Two Scenarios, Even Odds
Patel splits what follows automation into two outcomes and, “If pressed”, gives each a 50% chance: a “spark” and a “software-only singularity”, in which he wrote that the AI swarm improves itself “within a matter of months”.
In the spark case he expects “something like three to six years of AI progress in a single year”. His case for compute: “OpenAI’s end-of-2027 fleet should be able to support 2.8 million GPT-3-training-run-sized experiments a year, or about 7,660 per day”.
He wrote that “around 80% of the algorithmic progress at a given frontier lab is driven by its top 50 researchers”, which is why automating that group matters most in his model. Our AGI timelines tracker records how other forecasts have moved.
The Structural Read
As we read it, the median of 2028 carries his stated condition that the labs do not intentionally slow down, and his 25% tail sits on continual learning staying unsolved.
As we read the scenarios, both of his outcomes are fast: even the “spark” case means three to six years of progress in a single year, in his words.
Dwarkesh Patel, essay, 13 November 2026
“I know this is a wild claim, and it’s one that I myself didn’t believe even a couple of months ago.”
Three Implications
FOR FORECASTERS Patel puts 25% on after 2032, which leaves 75% on the milestone by 2032, by our arithmetic.
FOR LABS AND INVESTORS The evidence he cites is a 2x progress rate at Anthropic from independent evaluator METR, plus lab disclosures such as about a 7x rise in OpenAI researchers’ agent spend over one summer, by our arithmetic.
FOR READERS OF THE TRACKER Our AGI timelines tracker records how other forecasts have moved.
The framework records: “Model capability is exponential; product capability is approximately linear.”
As we read it, Patel’s essay is an argument about the first curve; the framework does not discuss Patel or his forecast.
The milestone years in Patel’s 13 November 2026 essay, shown as years after 2026 (our arithmetic). All years are his forecasts.
What Is Not Established
These are Patel’s forecasts, not measurements. The essay does not give the exact earlier year he would have named, only “at least the better part of a decade”. The figures he cites from METR, OpenAI and Anthropic are theirs; we did not re-derive them for this piece.
Patel puts the main risk to his timeline on continual learning, giving “a 25% chance that we won’t be able to automate Ilya-level researchers until after 2032”. The post that flagged the essay called November “the month of the Big Update”.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Dwarkesh Patel moved from a decade-plus view earlier this year to a 2028 median for AIs that match the best human AI researchers, with 25% on after 2032 and even odds between a fast “software-only singularity” and a slower “spark”. The evidence he leans on is a 2x progress rate at Anthropic from independent evaluator METR, plus the labs’ own disclosures: about a 7x jump in agent spend at OpenAI, by our arithmetic, and Claude leading over a quarter of Anthropic’s research work.
A note on sourcing. Patel’s words and forecasts come from his essay published on dwarkesh.com on 13 November 2026, read in full on 14 November. The METR, OpenAI and Anthropic figures are as Patel cites them. The 75% and the 7 times are our arithmetic. Patel was not contacted for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
