Source: raw/Recursive_self-improvement_has_started.md Reactions section also draws on Moonshots ep. 311 and All-In (paths in that section). Creator: Dwarkesh Patel · URL: https://www.youtube.com/watch?v=1NHPDzpfR6Y · Duration: not stated in the transcript (video essay plus a technical appendix) · Platform: YouTube (yt-podcast feed)

Dwarkesh Patel’s November video essay says he has moved his timeline by most of a decade in three months. His median for AI as good as the best human AI researchers is now 2028, with at least a 25% chance it slips past 2032 and a real chance it lands in 2027. The argument rests on one claim: AI research is about as verifiable as mathematics, where models went from failing high-school contests to solving open problems in about three years. These are forecasts, not measurements. The useful parts for a practitioner are the data points he collects and the list of verifiable proxies, which double as a map of where AI-driven speedups will show up first.

Key Takeaways

  • His numbers. Median 2028 for automated researchers at the level of the best humans, assuming labs are not slowed by choice or by government. At least 25% chance after 2032. 2027 is “the straightforward implication” of feeding METR’s 2x finding into a takeoff model.
  • Two scenarios, 50/50. If AI gets smarter faster than ideas get harder to find, a “software-only singularity” reaches the physical limits of current chips within months. If not, a “spark”: three to six years of progress at today’s rate compressed into one year (he gives “GPT-3.5 to GPT-6 Astra in just 12 months”), then a slowdown until more compute is built.
  • Why it would speed up: faster, smarter, more. Today’s models work at “about 10 times human speed” and do not sleep, which he puts at about 50x more work per researcher per calendar week. He estimates about 80% of a frontier lab’s algorithmic progress comes from its top 50 researchers; automation would make every agent that good. Experiment compute stays the bottleneck, so returns diminish.
  • Why ideas get harder. Labs already have to quadruple R&D compute and double headcount every year to keep the same rate of progress (his figures).
  • Verifiability is the whole case. Every bucket of AI R&D has a cheap automatic score: held-out loss for pretraining, compute or data needed to reach a score for post-training, GPU utilization and latency for inference. That lets labs RL models on the research itself and lets thousands of agents check their own work before merging.
  • He reversed his own continual-learning objection. In December 2025 he called automated AI research without on-the-job learning “super implausible”. He now thinks longer effective context may be enough, and that continual learning is itself an ML problem with verifiable proxies. This is a change from his August essay, Eight Predictions for the Era of Continual Learning.

The evidence he cites

All figures are his relays; none was checked against the original posts here unless noted.

  • METR at Anthropic: progress on well-scoped tasks is 2x a year ago, and a “majority of this acceleration today is likely attributable to Anthropic’s extensive use of AI to do R&D.” Matches the wiki’s account in METR’s October 2026 assessment.
  • OpenAI internal usage (September blog post): the heaviest 10% of AI users among OpenAI researchers went from about 7,000 a day by the end of August. New to the wiki.
  • Anthropic (September): Claude “now leads 26% of Anthropic’s AI R&D work, up from under 1% as recently as February.” The baseline month differs from other relays; see the contradiction callout in Anthropic on Recursive Self-Improvement.
  • Mathematics: “a month ago, where OpenAI solved a Millennium Prize problem in 88 hours.” The wiki records this as OpenAI’s Navier-Stokes result from an internal model (Lab Slowdown Warnings); the 88-hour figure is his.
  • Data beats architecture. With Princeton student Jerry Han, he found that over six years more than three times as much pretraining algorithmic progress came from data improvements as from model improvements. He quotes a DeepSeek post-training paper: improvements in “the scale, diversity, and verifiability of synthesized data and environments account for essentially all of the observed gains.”
  • Pretraining headroom. He says the lab Magic pretrained a model matching DeepSeek V4 Pro as a base model for half the compute of the original GPT-3, which he prices at about $500,000 today.
  • Benchmarks as RL environments. Anthropic’s internal CoBench holds “around 500 real problems that their researchers and engineers have solved this year.” ByteDance’s EdgeBench measures how fast a model learns an unseen environment.
  • Compute per agent. If frontier compute keeps more than quadrupling yearly, by end of 2027 OpenAI could give 10,000 agents enough compute each for a GPT-3-sized training run every day. For scale: the final model in the 2017 transformer paper took three days on eight P100s, about $50 to rent today.

The verifiable proxies, by bucket

BucketProxy he namesWhat an agent swarm can optimize
PretrainingLoss on a held-out setCompute needed to train a fixed-size model to a fixed loss (nanoGPT-speedrun style)
Research tasteDoes a 100M-parameter result hold at 10B?A “taste-bench” environment that scores scale predictions
Post-trainingCompute, data or time to reach a score on new environmentsFaster RL; faster in-context learning on unseen tasks
AI R&D directlyRedo recent human work (CoBench)The last three to six months of lab progress as practice
InferenceGPU utilization, KV cache per token, FLOPs per forward pass, latencyServing cost

Objections he answers

  • Continual learning: see Key Takeaways. He keeps a 25% chance this is what delays automation past 2032.
  • Finding the right proxy for a new paradigm: the transformer was built to beat existing translation benchmarks, and reasoning models moved existing coding and math benchmarks, so old proxies caught big jumps. Paradigm-agnostic outer loops also exist (“train on data before 1905; how long until it finds general relativity?”).
  • Data: producing better data is itself ML research that AIs are getting good at.
  • Experiment compute: today’s labs waste compute to save scarce researcher time; agents can trade labor for compute (more small ablations, reading every rollout, fixing broken environments) and find more scaling regularities like muP.
  • Real-world domains: he argues RL on vast simulated environments builds general skills that transfer, and that an AI shadowing a CEO could read every document on day one and interview every employee in parallel. Feedback that needs months of real time stays a bottleneck.

Same-week reactions: open-loop vs closed-loop RSI

Sources: Moonshots ep. 311 (raw/Jensen_s_Departure_NVIDIA_s_succession_plan_AI_doomer_response_and_Demis_talks_RSI_311_Live.md); All-In (raw/Jensen_OUT_as_Nvidia_CEO_-_Spy_Scandal_Export_Restrictions_Market_Impact.md). Quotes checked against the transcripts.

  • Demis Hassabis draws the line Patel’s scenarios depend on. In a clip played on Moonshots, he says today’s RSI is “early days”: AI is “making a human engineer maybe three, four, five times more productive”, with “the human’s very much still in the loop”. The next phase, “closed loop” RSI without a human reviewer, “I think carries the most risk”, because the result “will likely be quite alien” and hard to control. He adds that labs face “an intense … competitive race to improve your coding systems”. Alex Wissner-Gross, on the same show: “I think we’re pretty far into the RSI loop at this point”, with a human “pressing a button approving a pull request”.
  • Paul Christiano, opening the Hinton lectures in Toronto on November 9 (clip played on Moonshots): “I think there is a high risk of an AI takeover within the next three years.” Building AI “along the most rapid trajectory” would be “something like a coin flip”. He declines to give a number beyond that. The Moonshots host introducing the clip notes Christiano sits on OpenAI’s foundation board. Wissner-Gross and Emad Mostaque (who puts his own p(doom) at 20%, down from 50%) push back on the takeover claim. ^[inferred: speaker names from on-air address; the transcript has no labels]
  • All-In: RSI as an IPO pitch. One host, relaying people who attended Anthropic management meetings, says Anthropic presents itself as “building this recursive self-improvement engine”, with compute cost per unit of intelligence falling and a possible “winner take all scenario”. Another says Anthropic is “marketing it in their road show … ‘you should buy our stock because we’re doing recursive self-improvement’”, and calls this odd for a lab that says RSI is dangerous. ^[ambiguous: the coordinator’s note attributes the first to David Friedberg and the second to David Sacks; the transcript has no speaker labels]

Try It

  1. Use his proxy table to predict where your tools improve next. Work with an automatic score (tests pass, loss, latency, a checkable answer) is where models get better fastest. Move the parts of your own workflow that you can score into automated loops first; see Verifier-First Loops.
  2. Plan for faster release cycles, not a date. His 2027-2032 range is wide. A cheap hedge is to re-run your own task bench on every new model release rather than on a calendar.
  3. Quote the OpenAI 7,000-a-day figure as a ceiling, not a norm. It describes the heaviest 10% of researchers at one lab.

Open Questions

  • None of the cited posts was fetched. The OpenAI September usage post, Anthropic’s September uplift post, the DeepSeek paper, Magic’s announcement, CoBench and EdgeBench are all relayed by Patel. The Fable 5.2 system card says Anthropic withheld CoBench results “because we are updating the evaluation” (Fable 5.2 + Mythos 5.2); the 500-problem description is his.
  • The 26% baseline. He says “under 1% as recently as February”; other relays say 1% in March, 1% at the start of the year, or 3% in April.
  • The 88-hour figure for OpenAI’s Millennium Prize result has no source in the vault.
  • The Jerry Han data-vs-architecture study is not linked; the “3x” ratio cannot be checked here.
  • His speed and talent-concentration figures (10x human speed, 50x weekly output, 80% from the top 50 researchers) are his estimates.