AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: What The Evidence Can—and Can't—Tell Us About AI Self-Improvement on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get audio and creator gear 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 indicates that AI tools are helping humans conduct AI research and may be speeding up some development work. It does not document an AI system autonomously designing and building its own successor; a proposed U.S. bill to restrict such development remains in committee.

A viral essay arguing that AI is already accelerating the development of AI has renewed debate over recursive self-improvement, but the evidence it cites does not show systems independently designing and building their own successors. Anthropic reports that its Claude model leads a portion of the company’s AI research work, while an independent evaluator says AI assistance likely accounts for much of a measured increase in research speed; both describe humans as remaining involved.

Dwarkesh Patel’s November 13 essay, “The avalanche that hit math is coming for AI development,” draws on findings from Anthropic and the nonprofit evaluator METR, as well as figures published by OpenAI. Patel’s forecast that much faster progress may follow is his interpretation of those materials, not a new independent measurement. The distinction matters because “self-improvement” can refer either to AI assisting researchers or to a system carrying out the whole development cycle on its own, as explored in Anthropic’s evidence on recursive self-improvement.

Anthropic reported that, in August 2026, Claude led 26% of its AI research and development work. More than 90% of that work involved AI at or above the “collaborates” level on an automation scale, the company said. Anthropic also reported that no measured subset of its AI R&D work was operating fully autonomously. It defines the strongest form of recursive self-improvement as an AI system able to fully autonomously design and develop its own successor, and says that point has not been reached.

METR’s November 5 assessment examined Anthropic’s internal data from August 12 to October 29. It judged that, as of October 2026, Anthropic was likely improving model capabilities on well-specified tasks at around twice the rate recorded between March 2024 and October 2025. METR stressed that the estimate had high uncertainty: it said the true acceleration could be closer to 1.3 times or three times. The evaluator attributed most of the likely speed-up to AI assistance, while saying AI was not yet making highly significant, fully autonomous contributions to the most important parts of model training. It also said Anthropic redacted much of the evidence it considered most informative.

At a glance
reportWhen: As of November 16, 2026; the cited METR…
The developmentA viral essay has renewed debate about AI self-improvement, drawing on research that documents AI-assisted development but not fully autonomous systems building their successors.

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

The Difference Between Assistance and Autonomy

The distinction affects how readers interpret claims about an AI “improvement loop.” AI systems can help people write code, test ideas, or perform research tasks, potentially shortening parts of development. That is meaningful progress, but it is not the same as a system independently setting research goals, deciding how to train a successor, and completing that process without human intervention.

The reported acceleration could matter if it persists or expands to more complex work. But the evidence presented here does not establish how quickly that might happen, whether the reported gains apply across the industry, or whether current assistance will lead to autonomous development. For policymakers, the gap between present capabilities and a possible future system is especially relevant: a proposal aimed at broad categories of AI research could apply beyond the autonomous scenario prompting public concern.

There is also a source-quality question. Anthropic and OpenAI have access to internal operational data, but company-reported measures are not equivalent to independent, publicly reproducible evidence. METR provides an external assessment of Anthropic’s internal material, yet says much of the evidence behind its conclusion was redacted. The findings support a careful claim about AI-assisted work, not a definitive account of an industry-wide shift to self-directed AI development.

Products Worth Considering

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID (Midnight)

Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, Liquid Retina Display, and Touch ID (Midnight)

LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1, Black

LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1, Black

Samsung Galaxy Tab A11+ 11-Inch Tablet, 6GB RAM, 128GB Storage, Expandable, Gray

Samsung Galaxy Tab A11+ 11-Inch Tablet, 6GB RAM, 128GB Storage, Expandable, Gray

As an affiliate, we earn on qualifying purchases.

Research Findings and Proposed Restrictions

The discussion has a legislative dimension. The Ban Artificial Superintelligence Act of 2026 was introduced by Sen. Bernie Sanders as S.5493 on September 23 and Rep. Greg Casar as H.R.10538 on September 24. The proposed measure would pause training, modification, and fine-tuning of large AI systems until a new Department of Artificial Intelligence set rules. Its language includes recursive self-improvement and systems with the capacity to automate or greatly accelerate AI research and development.

As of the source report’s November 16 publication, the bills had been referred to committee and had not advanced. The House version had 12 Democratic cosponsors. A separate proposal from Rep. Ro Khanna, the “Human Control Over AI Act,” had been described in a September CNBC report, but the source found no bill text. The Associated Press reported that Democrats won both chambers of Congress on November 3; supporters of the restrictions see the result as a possible opening in the next Congress. That political outlook is not a change in the bills’ status.

OpenAI has also described progress toward research automation. The Decoder, which reviewed an OpenAI blog post that the source could not download, reported that OpenAI said it had met its “automated research intern” goal according to the company’s measurements and was aiming for a full automated AI researcher by March 2028. The Decoder reported that tasks 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 The Decoder noted that OpenAI did not cite independent review.

““We are not there yet, and recursive self-improvement is not inevitable.””

— Anthropic, in its essay “When AI builds itself”

Limits in the Available Evidence

How much AI assistance contributed to the reported acceleration remains uncertain. METR gave a broad range for the possible rate and said it had high uncertainty; its review also relied on internal Anthropic data, much of which was redacted. The assessment concerns one company and well-specified tasks, so it does not by itself establish a comparable trend across the AI industry.

OpenAI’s reported task results likewise come from company measurements, according to The Decoder’s account. The available material does not provide an independent evaluation of those figures. It is also unclear how Patel’s forecast would change if current task-level improvements fail to extend to the longer, more complex work involved in developing and training advanced models. The source material documents no AI system fully autonomously designing and developing its successor.

The legislative outcome is unsettled too. Both federal bills remain at the committee stage in the reported status, and the source provides no indication of a scheduled vote. The scope and practical effects of Khanna’s proposal cannot be evaluated from a summary alone because bill text was not available to the source.

Committee Action and New Measurements

The immediate next steps are further evidence and legislative action, not a confirmed change to AI development rules. Researchers and outside evaluators can test whether the reported gains hold over time, across companies, and on more complex work, while making methods and evidence available for scrutiny. Company statements about research automation should be read alongside the limits and review status of their measurements.

In Congress, the Sanders and Casar bills would need to advance from committee before lawmakers could consider their proposed restrictions. Any later version, amendment, or related proposal could clarify what counts as AI that “greatly accelerate[s]” research and how human involvement would be treated. As of November 16, 2026, the source report says neither bill has moved, and it reports no immediate change for businesses or users.

Key Questions

Has recursive self-improvement started?

AI is assisting with AI research, and Anthropic reports that Claude leads a portion of its R&D work. The evidence described does not show a system autonomously designing and building its own successor.

What did METR find about the pace of AI research?

METR estimated that Anthropic’s progress on well-specified model-capability tasks was likely around twice the rate recorded in an earlier period. It emphasized high uncertainty, with its estimate potentially closer to 1.3 times or three times.

What would the proposed federal ban cover?

The Ban Artificial Superintelligence Act of 2026 would pause training, modification, and fine-tuning of large AI systems until a new department set rules. Its text also refers to recursive self-improvement and systems that can automate or greatly accelerate AI research and development.

Has Congress passed the proposed ban?

No. As of November 16, 2026, the source report says the House and Senate bills had been referred to committee and had not advanced. The report says no immediate rules have changed for businesses or users.

Source: ThorstenMeyerAI.com

You May Also Like

Korean Courts Face Questions About Fake Precedents In The AI Era

A survey of 360 South Korean judges found widespread reports of fake or distorted citations and prompted calls for fines and AI-use disclosure rules.

The Scarcity Of Intelligence In An AI World: What’s Changing?

Thorsten Meyer’s 75,000-word book on AI, work and economic change entered ebook distribution on October 16, 2026.

The Search Shift: AI Answers, Publisher Earnings, And The Open Web

Google said fast-growing AI Mode queries earn less per query, while Search revenue rose 12%. Publisher traffic and the effect on clicks remain key concerns.

Six Patterns To Remove For More Readable AI Text

A preprint comparing 14 language models with human writing identifies six patterns linked to harder reading, but does not test reader comprehension.