Dwarkesh Patel: Why He Now Expects Superintelligence by 2030

Dwarkesh Patel now expects superintelligence by 2030, driven by AI that automates AI research, after saying earlier this year that it was most of a decade away. He makes the case in a 44-minute video essay, I thought superintelligence was a decade away. Not anymore., and in a longer written version with footnotes, The avalanche that hit math is coming for AI development. This post covers both, because the footnotes hold several of the numbers and caveats the narration leaves out.

TLDW

Patel argues that AI research is mostly a verifiable domain, like the well-scoped math problems that AI models went from failing to dominating in about three years, so AI will soon match the best human AI researchers. His median year for that is 2028, with a real chance of 2027 and a 25% chance it slips past 2032. Once research is automated, a swarm of faster, smarter and more numerous AI researchers triggers recursive self-improvement (RSI), which he splits evenly between a software-only singularity and a one-time “spark” of three to six years of progress in one year. The second half of the essay takes on the strongest objections: continual learning, research taste, data and experiment compute.

Thoughts

Experiment compute is the objection that decides whether any of this happens, and it is the one place where the video and the essay give different numbers. On camera, Patel says that by the end of 2027 OpenAI could “almost” give 10,000 agents enough compute for a GPT-3-sized training run every single day. The essay says “thousands of agents”, and its last footnote shows the working: about 7,660 such runs a day if the entire projected fleet did nothing else, or one run every four days for 10,000 agents if a third of the fleet goes to experiments. That is a four-fold gap between the spoken claim and the written arithmetic, and it rests on OpenAI quadrupling compute two more years in a row. The argument still stands on the weaker number, but only if small experiments keep predicting large ones, which is a hope backed by two examples (scaling laws and muP), not a law.

The 50/50 split between a singularity and a spark is the most honest number in the piece and also the least supported. Everything depends on one ratio: how fast automated researchers improve versus how fast ideas get harder to find. Patel gives the second half of that ratio (labs already quadruple R&D compute and double headcount each year to hold the same pace) and no measurement of the first. A coin toss is what a forecaster says when the key quantity is unknown. The essay’s footnote makes the spark more demanding than it sounds: three to six years of progress with flat compute means four to nine years of raw algorithmic progress in twelve months, which he himself calls a pretty aggressive estimate. So the cautious branch of the forecast is already an extraordinary claim.

The reversal on continual learning matters more than the dates, because Patel spent a year and a half as the best-known proponent of the view that models cannot do real jobs until they learn on the job. He has not claimed the problem is solved. He has claimed it may not need solving first: longer effective context could stand in for weight updates, and online learning is itself a research problem with clean metrics, so amnesiac models might crack it. That is a clever move, but it is circular in one respect. It assumes the models can do frontier research without the capability in order to invent the capability. He keeps a 25% chance that this fails and top researchers are not automated until after 2032, which is a large tail to carry alongside a headline of “by 2030”.

The present-day evidence is thinner in print than in the narration. The video says Claude now leads 26% of Anthropic’s AI R&D work, up from under 1% in February. The essay’s first footnote adds that “leads” means completing most of a task from a high-level prompt while a human supervises, and that the share of fully autonomous work is still 0%. METR’s finding of a doubled rate of progress is for well-scoped tasks. Another footnote supplies the right frame, Amdahl’s law: if AI works 50 times faster on half of a researcher’s job, the researcher gets only about 2 times faster, about 8.5 times at 90% and more than 30 times at 99%. The whole forecast lives in the gap between 90% and 100%, which is exactly where nobody has data.

The step from superhuman AI researcher to superhuman everything is the weakest link. It leans on two analogies: evolution built a general learner from a crude survival signal, and Napoleon and Rockefeller needed only a couple of years of experience. Patel concedes the measurement problem directly, since progress on unverifiable tasks is hard to measure by definition. Evolution also took an enormous number of generations in the real world, not in a simulated one, so the analogy cuts both ways. Even the ramp-up estimate moves between versions: years of deployed context in “a few days” in the video, “weeks” in the essay. Readers should treat the 2028 research milestone and the 2030 all-fields milestone as two separate bets with very different evidence behind them.

Key Takeaways

Patel’s median year for AI that matches the very best human AI researchers is 2028, assuming labs do not slow down on purpose and governments do not slow them.
He gives at least a 25% chance the milestone does not arrive until after 2032, and says 2027 would not surprise him.
He expects AI at least as good as the best human experts in every field by 2030, and plausibly as early as next year.
METR, after three months embedded at Anthropic, concluded that Anthropic’s rate of progress on well-scoped tasks is 2x what it was a year ago, with most of the acceleration likely due to AI doing R&D.
Patel cites an OpenAI post showing the heaviest 10% of its researchers went from about $1,000 a day of internal coding agents to over $7,000 a day across the summer.
Anthropic reported that Claude leads more than 26% of its AI R&D work as of August, up from under 1% in February. Fully autonomous work is still 0%.
Math is his template: models could not reliably add four-digit numbers in 2023, won International Mathematical Olympiad gold in 2025, and, by his account, solved a Millennium Prize problem (Navier-Stokes) in 88 hours about a month ago.
AI labs care about empirical results, not theory, so the grind-through-conjectures skill that is only partly useful in math is enormously valuable in machine learning.
An automated researcher works about 10x human speed around the clock, roughly 50x more work per calendar week, though experiment compute caps the gain.
Patel’s sense is that about 80% of algorithmic progress at a frontier lab comes from its top 50 researchers, so a workforce where everyone is at that level is a second speedup.
Two outcomes at 50% each: a software-only singularity within months, or a “spark” of three to six years of progress in one year followed by a compute-limited slowdown.
Even the spark equals a jump the size of GPT-3.5 to GPT-6 in 12 months.
Pretraining, post-training and inference efficiency each have measurable proxies (held-out loss, sample efficiency, latency), which is why he calls AI R&D largely verifiable.
Research by Patel and Jerry Han found that over six years, more than three times as much pretraining progress came from data improvements as from model improvements.

Chapters

00:00 AI Is Already Speeding Up AI Progress

Patel says he held decade-long timelines until about three months ago and changed his mind after debates on and off the podcast. Before the speculative part, he lists three current data points. METR’s assessment found Anthropic’s progress on well-scoped tasks has doubled in a year. OpenAI’s heaviest internal users increased their coding agent spend sevenfold over the summer. Anthropic says Claude leads about a quarter of its R&D work.

02:03 Math Is a Preview of What Happens to AI Research

Models went from failing arithmetic to solving a Millennium Prize problem in roughly three years, a collapse Patel marks with a line from East of Eden about how gods fall. He grants Toby Ord’s point that AI proves things without yet asking powerful questions, then turns it around: it solved a Millennium Prize problem without that ability. Machine learning rewards exactly this kind of persistent recombination. His example, from conversations with researcher Sholto Douglas, is the transformer, whose parts (attention, residual connections, layer norm, Adam) were all in the literature before 2017 and needed many small experiments to assemble.

08:15 The 2028 Median for Automated AI Researchers

In every domain with checkable answers and thousands of practice rounds, AI has gone from below average to better than the best humans within a few years. Patel expects the same for AI research, including the parts that cannot be checked directly, because models generalize beyond their training tasks. He puts the median at 2028, a 25% chance after 2032, and notes that plugging METR’s doubling into a takeoff model points to 2027. He does not fully trust that straightforward modeling, which is why his median is later.

11:10 Faster, Smarter and More Researchers

Three things compound once AI matches top researchers. Each one works at about 10x speed, nights and weekends included. Each one is as good as the top 50 people who drive most of a lab’s progress. There are far more of them, and AI research parallelizes well because teams already work on separate pieces with local ways to check results. He compares it to a swarm of Claudes formalizing Fermat’s Last Theorem, and admits returns diminish steeply in every case.

14:24 A Singularity or a Spark

Whether the acceleration sustains itself depends on a race between smarter researchers and harder ideas. If the researchers win, a software-only singularity produces a superintelligence at the physical limits of current chips within months. If ideas get harder faster, the result is a spark: several years of progress in one, then a wait for more compute. Patel says a post-spark system would out-design Nvidia, out-trade Citadel and Jane Street given equal market access, and lead the search for a room-temperature superconductor.

17:21 Superintelligence as a Meta-Learner

Most valuable work, such as building a company or winning a court case, cannot be graded inside a datacenter. Patel’s answer is that training will produce general skills for ramping up fast, not baked-in answers. Automated engineers would build tens of thousands of times more training environments, including simulated economies and backtests such as deriving general relativity from data that stops at 1700. A deployed system could read a company’s entire history on its first morning, and he expects top human performance in almost all fields within months, which lands in 2028 or 2029 in his median case.

25:59 Appendix: Why AI R&D Is Largely Verifiable

The appendix breaks AI R&D into pretraining, post-training and inference efficiency, and gives a measurable target for each. A model can be trained to minimize the compute needed to reach a fixed loss, in the style of the public nanoGPT speedrun, or to predict whether a small-scale improvement holds at larger scale. He cites the lab Magic matching leading open-source base models for half the compute of the original GPT-3 (roughly $500,000 today), ByteDance’s EdgeBench for on-the-fly learning, and an internal Anthropic benchmark of around 500 real problems its staff solved this year.

33:59 Objection: Continual Learning

This was Patel’s own objection as recently as last December. He now gives two reasons it may not block automation by 2028. Scaling in-context learning to sessions that last effectively many months may be enough for a single instance to learn on the job. And continual learning is a research problem with clean proxies, so models without it might solve it. He still calls the risk serious and ties his 25% tail to it.

36:44 Objections: Research Taste and Data

Critics say the proxies assume today’s paradigm and that choosing which hill to climb is the real work. Patel replies that the transformer and reasoning models both showed up on benchmarks that predated them, and that paradigm-agnostic tests are easy to write. On data, he argues the AIs will compile it themselves. He quotes a DeepSeek paper saying improvements in synthesized data and environments account for essentially all of its observed post-training gains, and that the pipelines building those environments are increasingly automated.

39:58 Objection: Experiment Compute

Unlike math, AI research needs experiments, and each agent in a large swarm gets a smaller share of compute. Patel answers that the share is still large: the final model in the 2017 transformer paper would cost about $50 to train today, around one ten-thousandth of GPT-3’s training compute. Labs currently waste compute to save scarce researcher time, and a swarm would do the opposite, reading every rollout and running many small ablations. It could also develop a better science of learning from small experiments, extending results like scaling laws and muP.

Notable Quotes

“Current AIs are slicing through well-scoped problems in math like a hot knife through butter.”
Dwarkesh Patel, in the video, on why math changed his mind

“If we take it at face value that AI is now nearly doubling the rate of AI capabilities progress and assume acceleration from AI is smooth, this naively implies that we would reach extreme superintelligence within a year.”
Ajeya Cotra, writing about the METR report, as quoted in Patel’s essay

“In the last four years, we’ve gone from a preschooler to a savant intern.”
Dwarkesh Patel, in the video, on the size of the jump a spark would repeat

“Rockefeller and Napoleon needed experience, yes, but only a couple years’ worth, not decades.”
Dwarkesh Patel, in the video, on how fast AI could ramp up in the real world

“It’s like watching somebody do cartwheels down the track and still keep pace with Usain Bolt.”
Dwarkesh Patel, in the video, on models solving hard problems without long-term memory

“Human thinking speed is a biological accident. It would be a strange coincidence if our speed happened to be exactly the point beyond which thinking faster was no longer helpful.”
Dwarkesh Patel, in a footnote to the written essay, on why faster researchers matter

Watch the full video essay here, or read the written version with footnotes and links at dwarkesh.com.