Is AI about to automate its own research? What the data says

Is AI about to automate its own research? What the data says

Some forecasters now expect AI to match the best AI researchers by 2028. We checked that case against METR, Anthropic and OpenAI data and show where it is weakest.

ZharfAI ResearchNovember 17, 202619 minutesSource: Dwarkesh Patel

Will AI soon do the work of the best AI researchers, and then improve itself? The most specific recent case for yes is Dwarkesh Patel's essay "The avalanche that hit math is coming for AI development," published on November 13, 2026. It forecasts that AI systems will be as good as the very best human AI researchers around 2028, and the resulting burst of automated research will produce AI at least as capable as top human experts in nearly every field by 2030. Patel says he held a much longer view only three months earlier. What changed his mind is a mix of new measurements from inside the frontier labs and the speed with which AI moved through well-defined problems in mathematics.

This article follows the essay section by section, including the technical appendix that carries most of its weight. Where the essay summarizes another organization's measurement, we went to that organization's own publication and report what it actually says. Where the essay forecasts, we label the forecast as Patel's. The last two sections are ZharfAI's assessment of where the argument is strongest and weakest, and what a team building with AI today should do with it.

Evidence status: the timelines and probabilities below are Patel's personal forecasts, not measurements. Acceleration figures were checked against METR, Anthropic and Epoch AI's publications; claims we could not open at the source are attributed to the essay. The charts are ZharfAI's own, drawn from the published data; Patel's original figures are in his essay. The cover is Philip James de Loutherbourg's An Avalanche in the Alps (1803, Tate Britain), the public-domain painting the essay uses as its own image.

The forecast, stated plainly

The essay's central object is the automated AI researcher: a system that does the work of a frontier lab's best researchers, at machine speed and in large numbers. Recursive self-improvement, abbreviated RSI, is what happens when such systems improve the next generation of themselves. Patel's forecasts, with the confidence he attaches to each, are these:

MilestonePatel's estimate
AI as good as the best human AI researchersMedian 2028; 2027 plausible; about 25% chance it slips past 2032
Shape of the first year after that point50% "spark" (roughly three to six years of progress compressed into one), 50% software-only singularity
AI at least as good as top human experts in almost every fieldBy 2030, within months of deployment, in his median scenario
Assumption behind all of the aboveLabs do not deliberately slow down

Two points matter for reading the rest. First, his "best researcher" bar is deliberately high: he names researchers of the calibre of Ilya Sutskever and Noam Shazeer. Second, the 25% tail is not a footnote to him; most of the appendix argues about which bottlenecks could produce it.

The acceleration evidence, checked against its sources

Patel opens with three data points meant to show that AI is already speeding up AI development. All three come from the labs themselves or from an evaluator working inside one, so the details of each measurement matter.

What the essay saysWhat the primary source says
METR found Anthropic's rate of progress on well-scoped tasks is twice what it was a year ago, mostly because of AIMETR's assessment (November 5, 2026) says progress on well-specified tasks is about twice the March 2024 to October 2025 rate. Its benchmark index grew 1.9x faster (95% interval 1.5x to 2.8x), and the authors would not be surprised by 1.3x or 3x. They guess "a majority (perhaps ≈80%)" of the acceleration is due to AI.
OpenAI's heaviest 10% of AI users went from about $1,000 a day to over $7,000 a dayOpenAI's September 2026 post, as tabulated by Epoch AI: the 90th-percentile researcher spent $972 a day in the week ending May 31 and $7,047 in the week ending August 15. The median rose from under $1 in January to $601. Values are at API list prices, not OpenAI's internal cost.
Claude "leads" more than 26% of Anthropic's AI R&D work as of August, up from under 1% in FebruaryAnthropic's measurement page shows 26% at "leads" in August, after 1%, 3%, 12%, 14% and 22% from March to July. "Leads" means Claude completes most of a task from a high-level prompt while a person supervises. No measured work is fully autonomous.

Daily coding-agent spend per OpenAI researcher

Weekly mean in US dollars at API list prices, log scale; January to August 15, 2026

Daily spend, US dollars
Week endingMedian90th percentile
2026-01-04$0$1.95
2026-01-11$0$10
2026-01-18$0.17$19
2026-01-25$0.51$29
2026-02-01$0.94$34
2026-02-08$2.24$43
2026-02-15$1.72$43
2026-02-22$1.09$44
2026-03-01$5.66$86
2026-03-08$10$112
2026-03-15$10$112
2026-03-22$13$156
2026-03-29$14$239
2026-04-12$29$345
2026-04-19$42$413
2026-04-26$45$460
2026-05-03$66$573
2026-05-10$82$645
2026-05-17$80$688
2026-05-24$101$795
2026-05-31$137$972
2026-06-07$148$971
2026-06-14$161$1,111
2026-06-21$156$1,020
2026-06-28$151$1,169
2026-07-12$163$1,642
2026-07-19$236$2,188
2026-07-26$304$4,413
2026-08-02$377$5,274
2026-08-09$434$6,100
2026-08-15$601$7,047

Data: OpenAI’s September 2026 post, tabulated by Epoch AI (CC BY). Self-reported by OpenAI; programmatic use excluded. Chart: ZharfAI.

On a log scale both lines are close to straight, which is what Epoch's fit describes: spending roughly doubled every month for the median and the 90th-percentile researcher alike.

Share of Anthropic’s model R&D work that Claude “leads”

Monthly, 2026; under 1% in February. “Leads” means Claude completes most of a task from a high-level prompt while a person supervises.

Data: Anthropic’s measurement page, level AL4 on Epoch AI’s automation scale; levels assigned by a Claude-based judge; Anthropic’s 90% intervals are not shown. Fully autonomous work (AL5): none. Chart: ZharfAI.

The caveats are part of the evidence. METR worked for eleven weeks and says Anthropic redacted a majority of the evidence METR found most informative. It also says the benchmark trend could partly reflect overfitting to the benchmarks. Anthropic's levels are assigned by a Claude-based judge over a fixed basket of about 15,000 tasks. Epoch notes that OpenAI's numbers are self-reported, exclude programmatic use, and price tokens at list rates. None of this makes the trend false. It means the strongest evidence for acceleration is about benchmarkable, well-specified work, measured mostly by the labs doing it.

Why mathematics is his leading indicator

Patel's main reason for updating is mathematics. In his telling, models went from unreliable four-digit addition in 2023 to Olympiad gold in 2025, to open Erdős problems early in 2026, to what OpenAI describes as a solution of a Millennium Prize problem about a month before the essay. We have not reviewed that last result; the essay links OpenAI's announcement.

He takes on the strongest counterpoint directly. Toby Ord argues that current AI is automating proof, not the harder work of asking deep questions or opening new fields. Patel accepts the description and reverses its implication. In mathematics, a long list of solved conjectures matters less than new theories. In machine learning, labs mostly want recipes that make training more efficient, and a decade and a half of deep learning progress has needed little new theory. Grinding through many well-scoped problems is therefore worth more in AI research than in mathematics.

His example is the transformer. The 2017 paper combined pieces that already existed. According to a forthcoming post by researcher Sholto Douglas that Patel previews, attention beat recurrent networks only together with residual connections (2015), layer normalization (2016), the Adam optimizer (2014), interleaved MLP layers, positional encodings and learning-rate warmup. The bottleneck was finding the combination, which took taste and many small experiments. Persistent, systematic recombination of known ideas is exactly what models have shown in mathematics.

Faster, smarter, more numerous

Once automated researchers reach the top human level, Patel names three reasons algorithmic progress would speed up.

Faster. He estimates current models already work about ten times faster than people, citing OpenAI's claim that GPT-5.2 finished its GDPval tasks 11 times faster than experts and generation rates of around 60 tokens per second. Add round-the-clock work and he gets roughly 50 times more work per calendar week per automated researcher. Experiments do not run faster, so he expects steep diminishing returns, but still a many-fold speedup.

Smarter. His sense is that about 80% of a frontier lab's algorithmic progress comes from its top 50 researchers. If every automated researcher works at that level, the average quality of the research effort rises sharply.

More. Parallel agents help when work splits into modules that can each be checked locally. Patel points to Anthropic's formalization of Fermat's Last Theorem by many cooperating Claude agents; our reading of the Fable 5.2 system card covers the 1,000-agent run behind that result. AI research looks more like that formalization than like writing a novel. Labs already run separate teams for data filtering, RL environments, kernels and so on, each with its own measurable target.

A footnote in the essay explains why partial automation already matters, using Amdahl's law: the part of a job that cannot be sped up sets the pace. We computed the full curve for Patel's assumption that AI does its share of the work 50 times faster and the rest stays at human speed:

Share of a researcher's work AI can doOverall speedup
50%1.96x
75%3.8x
90%8.5x
95%14.5x
99%33.6x

Partial automation: the work left to people sets the pace

A researcher’s overall speedup by the share of work AI does, if AI does that share 50 times faster (Amdahl’s law)

The 50× assumption is from Dwarkesh Patel’s essay. Calculation by ZharfAI: speedup = 1 ÷ (human share + AI share ÷ 50).

The table makes the argument's real hinge visible. Going from 90% to 99% automation quadruples the speedup, so the last few percent of human-only work matter more than everything before them. That is why the essay spends its appendix on whether the hardest parts of research can be automated.

Patel's intuition pump for why thinking speed matters comes from Ryan Greenblatt and Eli Lifland's slow-corporations essay. Picture a lab with OpenAI's compute whose researchers work one hour a week. It would progress much more slowly. Compared with a company of automated researchers, he argues, today's labs are that slow lab.

Singularity or spark

The pace after automation depends on two opposing trends. RSI keeps making the researchers faster, smarter and more numerous. Meanwhile, ideas get harder to find. Patel's evidence is that labs already roughly quadruple their R&D compute and double their workforce every year just to keep progress steady; Epoch AI measured frontier training compute growing four to five times a year from 2010 to May 2024.

Singularity or spark: Patel’s two paths after research is automated

Vertical axis: AI progress in years at today’s pace. A schematic, not a quantitative forecast.

Based on the figures in Dwarkesh Patel’s essay (November 2026): equal odds for the two paths, and three to six years of progress in the spark’s first year (center line: about four). After year one the spark is drawn at the pre-automation pace. Chart: ZharfAI.

If AI improves faster than ideas get harder, the result is a software-only singularity: without new hardware, successive models improve until they run into the limits of existing chips, within months. If ideas get harder faster, the result is a spark: three to six years of AI progress in one year, then a slowdown until more compute is built. Patel puts 50% on each. A footnote adds that the spark implies four to nine years of raw algorithmic progress in that year, because compute would grow little.

Even the spark is dramatic. Patel compares it to the jump from GPT-3.5 to GPT-6 Astra happening within twelve months, starting from 2027 or 2028. After a spark, he expects AI to accelerate compute growth itself: chip design, fab layout and teleoperated robots for construction. Even without that help, Epoch AI estimates that the world's AI computing capacity has grown about 3.3 times a year since 2022.

From the datacenter to the real world

The obvious objection is that most valuable work, such as running a company or winning a case, cannot be graded inside a datacenter. Patel does not claim the answers can be trained into the weights. His claim is that scaled reinforcement learning will produce strong general learners that adapt quickly to new settings.

His sketch of future training has three parts:

  • Far larger environments. He expects automated engineers to build tens of thousands of times more RL environments: long-running, open-ended, multi-agent simulations such as an economy of millions of AIs in which one must co-found a company.
  • Historical backtests. Models would predict real discoveries and events from data cut off before them, for example deriving general relativity from a corpus ending in 1700, or forecasting the Cuban Missile Crisis from newspapers up to 1962.
  • Better learning algorithms. Sample efficiency and online learning have measurable proxies, so automated researchers can improve the learning method itself, not only the data.

For transfer, his evidence is that today's coding agents already orient themselves in unfamiliar codebases, and that recent models dominated older ones in his social-deduction game evaluation. His deeper analogy is evolution: a crude objective, passing on genes, produced minds that write biographies and build semiconductors. AI training uses objectives far closer to today's work and actively tests transfer on held-out tasks.

He accepts that some knowledge only comes from contact with reality, and argues it can be gathered fast. Parallel copies of a model could read a company's entire record and interview every employee in a day. Only feedback that takes calendar time, such as quarterly results, stays slow. On that basis he expects top human performance in most fields within months of deployment, 2028 or 2029 in his median case.

Is AI research verifiable? The appendix, mapped

The appendix argues that nearly all algorithmic progress falls into three areas, each with proxies a machine can check. Verifiability does two jobs: a lab can train models on these targets through thousands of practice rounds, and agents can work in parallel because each can check whether its change beats the current best before merging it.

AreaProxy in the essayHow it would be used
PretrainingCompute needed to reach a fixed loss; loss on held-out textAn RL objective such as "train a GPT-2-sized model to this loss with the least compute"; the public nanoGPT speedrun is an existing version
Research tasteWhether a gain at 100M parameters survives at 10BA costlier "taste" environment that still fits a frontier lab's budget
Post-trainingCompute, data or time needed to learn a new environment; speed of learning in unseen environmentsByteDance's EdgeBench, which Patel cites, includes such tasks
Direct R&D skillAn internal benchmark of real research problemsPatel cites Anthropic's CoBench, about 500 problems its staff solved this year, from a redacted Anthropic risk report
InferenceGPU utilization, KV cache per token, FLOPs per forward pass, latencyDirectly measurable on every change

The idea is close to what we described in our guide to verifiable rewards in post-training: when a result can be checked independently, training can reward finished work rather than persuasive text, but the checker becomes part of the product and can be gamed.

Patel adds an argument about model size. Better pretraining lets a smaller model match a larger one, which makes experiments and RL cheaper, lets more automated researchers run on the same hardware, and shortens the loop itself. During RSI, he expects labs to stop long giant training runs, because a much smaller model started a week later would overtake them. He also cites a claim by the lab Magic that it matched leading open base models for half the compute of the original GPT-3. We have not verified that claim.

The four objections he answers

Continual learning. For more than a year, Patel's main reason for skepticism was that models do not learn on the job: each session starts from scratch. He now sees two ways around it. Scaling in-context learning to sessions that are effectively months long might be enough. Or models without on-the-job learning might solve continual learning anyway, since it has clean proxies. He keeps a serious chance that neither works, and this is the main source of his 25% tail beyond 2032.

Choosing what to measure. If reaching superintelligence needs new paradigms, today's proxies may point at the wrong hill. His reply: the transformer beat an existing translation benchmark, and reasoning models showed up on existing coding and math benchmarks, so old proxies have captured large jumps. Paradigm-agnostic outer tests also exist, such as solving open problems or rediscovering a known theory from data cut off before it.

Data. He argues data growth continues during RSI, with models generating and filtering it. In research he published with Jerry Han, data improvements explained more than three times as much pretraining progress over six years as model changes. He also cites DeepSeek's account that synthetic data and environments, built by increasingly automated pipelines, explain almost all of its recent post-training gains.

Experiment compute. This is the objection where the arithmetic is easiest to check, so we reproduced it. Patel starts from 1.74 million H100 equivalents for OpenAI at the end of 2025, a figure he takes from Epoch AI's chip-holdings data:

StepValue
Quadruple in 2026 and again in 20271.74M × 16 ≈ 27.8M H100e
Annual compute at 10¹⁵ FLOP/s each≈ 8.8 × 10²⁹ FLOP per year
GPT-3 final training run≈ 3.14 × 10²³ FLOP
GPT-3-sized runs per year≈ 2.8 million, about 7,660 a day
A third of the fleet shared by 10,000 agentsOne GPT-3-sized run per agent about every four days

The numbers reproduce. They also assume that every chip runs at its full nominal rate. Real training runs use only part of that peak rate; at an illustrative 40% utilization, the interval stretches to roughly ten days, which does not change Patel's qualitative point. For scale, the big model in the original transformer paper trained for 3.5 days on eight P100 GPUs.

He also argues that compute-limited researchers would use compute differently: smaller and more numerous ablations, close reading of every rollout, and a better science of extrapolation. The GPT-4 technical report predicted the model's final loss from runs using at most 10,000 times less compute, and μP lets hyperparameters tuned on a small model transfer to a large one.

Where the argument is strongest and weakest

This section is ZharfAI's assessment, not Patel's.

Strongest. The direction of the measurements is consistent across three separate publications, though each rests on data from the labs themselves, and the mechanism is concrete. AI research is unusually rich in checkable targets, and the Amdahl table shows why even partial automation compounds. The appendix converts a vague debate about "AGI" into specific, testable sub-claims.

Weakest.

  • The evidence measures well-specified work. METR says its estimate covers well-specified tasks and that translation to broader work is more uncertain. The forecast requires the opposite end: research taste and choosing directions.
  • The transfer step is an inference. Mathematics shows that training on many checkable problems generalizes to harder checkable problems. That research judgment will transfer the same way is plausible, not demonstrated.
  • Much of the data is redacted or self-reported. METR could not publish most of what persuaded it, Anthropic's levels come from a model judge, and OpenAI's spend series is priced at list rates.
  • Simple extrapolation is fragile. The 2027 date comes from feeding a doubled rate into a smooth takeoff model. Ajeya Cotra's post on the METR report, which Patel quotes, describes that implication as naive, and Patel himself keeps his median later for this reason.
  • The physical world sets its own clock. Compute growth depends on chips, power and capital; deployment depends on regulation and on feedback that arrives quarterly or yearly. The spark scenario explicitly depends on that slower clock.

None of these refutes the essay. Together they explain why Patel's own distribution is wide, and why its tails deserve as much planning attention as its median.

What to watch, and what it means for teams building with AI

The essay is useful less as a date than as a list of indicators that would confirm or weaken it. The signals worth tracking over the next year:

  1. The share of AI-led and fully autonomous work in lab disclosures. Anthropic reports 26% "leads" and none fully autonomous; movement at the autonomous level would matter more than another rise at "leads."
  2. Independent assessments with less redaction. METR's next report, and similar audits of other labs, would show whether the acceleration holds outside well-specified tasks.
  3. Gaps between model generations. If RSI is real, release intervals and capability jumps should both change visibly.
  4. Compute deployment against plan. The experiment-compute argument assumes roughly fourfold yearly growth continues through 2027.

For a company building with AI today, including teams in Iran that work with limited access to frontier infrastructure, three consequences follow even if Patel's median is too early:

  • Expect rapid capability churn. Design systems so the model can be replaced without rewriting the product: keep prompts, tools and evaluation separate from any one provider.
  • Own your verifiable proxies. Patel's argument is that progress concentrates where results can be checked. The same holds inside a company. A test set of real cases with checkable answers is what lets you adopt each new model safely and quickly.
  • Plan for scenarios, not a date. Economy-wide effects are uncertain; our analysis of Anthropic's economic scenarios shows how to turn such uncertainty into decisions about tasks, budgets and staff transitions.

Patel promises a follow-up essay on what such systems would be like and how they would change the world. The test of this one is simpler: whether the measurements he relies on keep moving in the direction he expects.

Source notes — reviewed November 17, 2026

The essay was read in full, including footnotes and appendix, on November 17, 2026. The four charts are ZharfAI's: the OpenAI series is plotted from Epoch AI's published table (CC BY), the Anthropic bars from the six values Anthropic labels, the scenario chart from the numbers in the essay, and the speedup curve from our own calculation. Patel's original figures are not reproduced here and can be seen in the essay. The cover is a 16:10 crop of the public-domain reproduction on Wikimedia Commons.

Claims attributed to the essay without a link here, including OpenAI's mathematics results, GDPval speed, Magic's pretraining result, EdgeBench, CoBench and DeepSeek's post-training account, were not independently reviewed.

TagsRecursive Self-ImprovementAI Research AutomationSuperintelligenceDwarkesh PatelMETRAI Forecasting

Related Posts

Name one process for a discovery call

If this note maps to a real system in your organisation, start with the services page or a shipped case study.