Evidence On AI's Recursive Improvement—and The Rules It Could Raise
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Evidence On AI's Recursive Improvement—and The Rules It Could Raise on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get the latest gadgets delivered free with Prime

  • Fast, free delivery on millions of items
  • Prime Video, Amazon Music and more included
  • Member-only deals all year
Start your free Prime trial Free trial for eligible customers · Cancel anytime
As an affiliate, we earn on qualifying purchases.

TL;DR

Evidence cited in a viral essay shows AI systems helping researchers develop AI, but does not document a system independently building its own successor. A pair of congressional bills would pause some AI development and create new oversight, but both remain in committee.

A viral essay by podcaster Dwarkesh Patel has renewed debate over whether AI systems are beginning to improve AI, but the evidence it cites documents AI-assisted research with humans still involved, not a system independently designing and building its successor. The debate also puts two proposed congressional bills under scrutiny: both would restrict some AI development, but neither has advanced beyond committee referral.

In a November 13 essay titled “The avalanche that hit math is coming for AI development,” Patel argues that AI is speeding up the work of making AI and forecasts that progress could accelerate. The essay draws on findings from the nonprofit evaluator METR and figures published by Anthropic and OpenAI. Its forecasts are Patel’s analysis, not new experimental evidence.

Anthropic reported that, as of August 2026, Claude “leads” 26% of the company’s AI research and development work, using an automation scale in which a human supervises work completed from a high-level prompt. The company said more than 90% of the work it measured was at or above a level where AI collaborates with people, and that no measured subset of its AI research and development runs fully autonomously.

METR’s November 5 assessment examined Anthropic’s internal data from August 12 to October 29. It judged that Anthropic was improving model capabilities on well-specified tasks at about twice the rate seen from March 2024 to October 2025, while stressing high uncertainty: its estimate could be closer to 1.3 times or three times. METR said AI assistance likely accounts for most of the acceleration, but also said AI was not yet making highly significant, fully autonomous contributions to the most important parts of model training. It noted that Anthropic redacted much of the evidence it considered most informative.

At a glance
reportWhen: As of November 16, 2026; the essay appe…
The developmentA viral essay has revived debate over recursive AI self-improvement, drawing attention to evidence of AI-assisted research and to two proposed U.S. bills that would restrict some AI development.

Has it started? · explainer

AI already does much of the work of building AI. Not alone.

How far AI research and development is automated at one lab, on a scale from AL0 (no AI) to AL5 (no human in the loop), and where the proposed ban would draw its line.

Anthropic’s AI R&D work, August 2026

Automation levels from Epoch AI’s scale, as measured by Anthropic

AI collaborates or moreabove 90%
AL3 and up
AI leads, a human supervises26%
AL4
Fully autonomousno measured part
AL5 · the closed loop

Anthropic: “We are not there yet.”

The bill’s line: “the capacity to automate or greatly accelerate” AI R&DS.5493 / H.R.10538 list it as a “superintelligence precursor”. How much is “greatly” is not defined.
  1. AI collaborates on most of the workAbove 90% of the work is at or above “AI collaborates”, where AI does large chunks under close human direction.
  2. AI leads about a quarterIn 26%, Claude completes most of a task from a high-level prompt while a human supervises.
  3. No closed loopNo measured part runs fully autonomously. That is the step “recursive self-improvement” usually means.
  4. Where would a ban start?The bill names “greatly accelerate” as a warning sign, so its reach depends on a measurement nobody yet makes in a standard way.

Sources: Anthropic, “Measurements for understanding the pace of AI development inside frontier labs” and “When AI builds itself” (figures as of August 2026, Anthropic’s own measurements); Ban Artificial Superintelligence Act of 2026, S.5493 and H.R.10538, sec. 3(6)(A). Bars show shares of Anthropic’s AI R&D work.

Graphic: Powered by Thorsten Meyer AI

What AI-Assisted Research Could Change

The distinction between AI helping researchers and AI independently building a successor matters for both policy and public understanding. Assistance can speed up parts of development while people continue to set goals, judge results and decide whether to train or deploy a model. That is a meaningful shift in research workflows, but it is not the same as a self-directed improvement loop operating without human oversight.

The findings also come with limits. METR’s estimate concerns well-specified tasks at one company, and the evaluator describes its conclusion as uncertain. Anthropic supplied internal evidence, much of which was redacted. The available material therefore offers a partial account of how AI contributes to development, not a complete or independently reproducible picture of the research process.

For readers, the immediate issue is not that a fully autonomous system has been documented, but that AI assistance may be increasing the pace of some AI research. If that pattern continues, questions about who sets research priorities, how progress is measured and what oversight is appropriate could become more pressing. Those implications remain distinct from the stronger claim that AI has taken over its own development.

Products Worth Considering

Anker SOLIX C1000 Gen 2 Portable Power Station with 100W Solar Panel

Anker SOLIX C1000 Gen 2 Portable Power Station with 100W Solar Panel

Bose QuietComfort Headphones - Wireless Bluetooth Over-Ear Noise Cancelling Headphones, Petal Pink Limited Edition

Bose QuietComfort Headphones – Wireless Bluetooth Over-Ear Noise Cancelling Headphones, Petal Pink Limited Edition

As an affiliate, we earn on qualifying purchases.

The Evidence Behind the Viral Claim

Recursive self-improvement, or RSI, usually refers to a feedback loop in which an AI helps create a more capable AI, which then helps develop a successor. Anthropic describes the endpoint as an AI system able to fully autonomously design and develop its own successor. The company says that endpoint has not been reached and is not inevitable, though it could arrive sooner than institutions expect. That warning is a company assessment, not proof that autonomous self-improvement is already occurring.

OpenAI’s September figures, as reported by The Decoder after it reviewed the company’s blog post, provide another example of human-supervised automation. OpenAI said it had met its goal of an “automated research intern,” according to its own measurements, and aims for a full automated AI researcher by March 2028. The Decoder reported that tasks lasting under 15 minutes succeeded 86% of the time without intervention, while more than half of successful tasks lasting four to eight human hours required at least one human step. OpenAI said people still set priorities and decide on scaling, pauses and deployment. The figures are company-reported and, according to The Decoder, had no independent review cited.

The policy debate has a concrete legislative counterpart. Sen. Bernie Sanders introduced the Ban Artificial Superintelligence Act of 2026 as S.5493 on September 23; Rep. Greg Casar introduced its House counterpart, H.R.10538, on September 24. The bills would pause training, modification and fine-tuning of large AI systems until a new Department of Artificial Intelligence established rules. Their provisions include recursive self-improvement and systems capable of automating or greatly accelerating AI research and development. Both bills were referred to committee; the House measure had 12 Democratic cosponsors, according to the source material.

A separate proposal from Rep. Ro Khanna, called the Human Control Over AI Act, was described in a September CNBC report as a summary. The source reports that no bill text was available. The renewed attention comes after the Associated Press called both chambers of Congress for Democrats on November 3, which supporters of the proposed ban see as an opportunity in the next Congress. That political interpretation does not change the bills’ current status.

““As of October 2026, Anthropic is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025.””

— METR, in its November 5 assessment of Anthropic

Limits of the Available Evidence

The evidence does not establish that any AI system has independently designed and built its own successor. It also does not establish how much of the reported research acceleration would persist across companies, tasks or longer periods. METR’s estimate is explicitly uncertain, and its analysis relied on internal Anthropic data, much of which was redacted.

OpenAI’s reported performance figures are also company measurements relayed by The Decoder, rather than an independently reviewed evaluation in the material available here. The figures do not show that a system can set research priorities or decide on its own to scale, pause or deploy a model; OpenAI says people retain those decisions. The proposed federal rules have not been enacted, and the source material does not establish the eventual fate of the bills or the details of Khanna’s proposal.

Congressional and Research Milestones

The next immediate policy milestone is whether the Sanders and Casar bills receive committee action. As of November 16, both remained referred to committee, so they impose no new requirements on businesses or users today. Their text would pause covered activities pending rules if enacted, but the proposals have not become law.

On the research side, further independent evaluation and clearer disclosure could help show how much AI contributes to model development and whether those contributions are becoming more autonomous. OpenAI has stated a goal of a full automated AI researcher by March 2028; that is a company target, not a confirmed outcome. The central question remains whether current human-supervised assistance develops into a system capable of managing its own successor development—and what evidence would be needed to establish that change.

Key Questions

Has recursive AI self-improvement been demonstrated?

Not in the fully autonomous sense described by Anthropic. The evidence cited reports AI assisting with research and leading some work under human supervision. No system independently designing and building its own successor is documented in the source material.

What did METR find about Anthropic’s progress?

METR judged that Anthropic’s model capabilities on well-specified tasks were improving at about twice the rate seen between March 2024 and October 2025. It gave the estimate with high uncertainty, saying the true acceleration could be closer to 1.3 times or three times, and said AI assistance likely accounts for most of the increase.

Would the proposed U.S. bills ban AI research?

The Sanders and Casar proposals would pause training, modification and fine-tuning of large AI systems until a new department established rules. They also cover recursive self-improvement and systems able to automate or greatly accelerate AI research. Neither bill has advanced beyond committee referral, so neither changes current rules.

What has OpenAI said about automated AI research?

OpenAI said it had met its goal of an “automated research intern” according to its own measurements and aims for a full automated AI researcher by March 2028. The Decoder reported that longer tasks often still required human intervention and that OpenAI cited no independent review of its figures.

Source: ThorstenMeyerAI.com

You May Also Like

Anthropic’s November 3 Change Brings Text Watermarks To Earlier Claude Versions

A Mashable headline reports text watermarking will reach older Claude versions on Nov. 3, but gives no year, model list or technical details.

The Twelve Real Complaints About AI Tools in 2026 — A Reddit, Twitter, and GitHub Synthesis

A detailed report on the top twelve user complaints about AI tools in 2026, sourced from Reddit, Twitter, GitHub, and official reports, highlighting real-world friction points.

ChatGPT Pro Now: What OpenAI’s Changes Mean For Subscribers

OpenAI cancelled a planned Pro 200 reduction and doubled Work and Codex allowances on Pro 100 and Pro 500. Prices are unchanged.

How NVIDIA Kumo Tabular Balances Accuracy And Efficiency In Prediction

NVIDIA has released Kumo Tabular, an open model for table-based predictions. Its benchmark claims lack scores and independent validation in the supplied material.