🔍 Read the full analysis: How To Read The Evidence On AI Recursive Self-Improvement on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Evidence cited in a viral essay points to AI helping researchers develop AI faster, with people still setting goals and handling key decisions. Anthropic says no measured part of its AI research and development runs fully autonomously, while METR describes its acceleration estimate as highly uncertain. A proposed congressional ban remains in committee.
A viral essay by podcaster Dwarkesh Patel has renewed debate over whether AI systems have begun improving themselves, but the evidence it cites describes AI helping human researchers rather than independently designing successor models. Anthropic reports that Claude leads a portion of its AI research work, while saying no measured subset is fully autonomous; an independent assessment by METR estimates faster progress on certain tasks but gives that estimate with high uncertainty.
Patel’s November 13 essay, titled “The avalanche that hit math is coming for AI development,” draws on assessments and figures published by Anthropic and OpenAI. Its claims about what could happen next are Patel’s forecast, not new experimental results. The central distinction is between AI systems assisting or leading parts of research under human supervision and systems independently designing and developing their own successors.
Anthropic says that, as of August 2026, more than 90% of its AI research and development work was at or above a level where AI collaborates, and Claude led 26% of the work. The company says Claude was not operating fully autonomously for any measured subset. METR, which assessed Anthropic using internal data from August 12 to October 29, said the company was likely improving model capabilities on well-specified tasks at around twice its rate from March 2024 to October 2025. METR stressed that the estimate has high uncertainty.
METR attributed most of the apparent speed-up likely to AI assistance, while saying AI had not made highly significant, fully autonomous contributions to the most important parts of model training. The organization also said Anthropic redacted much of the evidence it considered most informative. OpenAI’s own figures, as reported by The Decoder after it reviewed a company post, describe progress toward an automated research assistant, but also say that longer tasks often still require human intervention. OpenAI’s published figures have not, in the material provided, been independently reviewed.
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
Anthropic: “We are not there yet.”
- 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.
- AI leads about a quarterIn 26%, Claude completes most of a task from a high-level prompt while a human supervises.
- No closed loopNo measured part runs fully autonomously. That is the step “recursive self-improvement” usually means.
- 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
Human Oversight Still Shapes AI Research
The distinction matters because AI-assisted research acceleration and fully autonomous self-improvement carry different implications. If AI helps researchers write code, test ideas or complete defined tasks, it may speed up development while people continue to choose research priorities and make decisions about training and deployment. That is not the same as an AI system independently creating a more capable successor and repeating the process.
The available evidence suggests that AI is already part of some companies’ research workflows, but does not establish that the human-free feedback loop at the center of many concerns is operating. METR’s estimate could inform debate about the pace of progress, but its uncertainty and limited visibility into company evidence make it a poor basis for treating a particular timeline as settled. The claims matter to readers because they inform discussions about research oversight and proposed regulation, even as the current findings remain narrower than some public descriptions of “recursive self-improvement.”
Products Worth Considering

Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite

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

Amazon Kindle Scribe (32GB) with Premium Pen, Tungsten (Like-New)
As an affiliate, we earn on qualifying purchases.
From Assistance to Full Autonomy
Recursive self-improvement, or RSI, generally refers to a loop in which an AI system helps create a better AI system, which can then help create the next one. Anthropic defines the endpoint as an AI system able to fully autonomously design and develop its own successor, and says that point has not been reached. The company’s framework distinguishes levels of involvement, from no AI participation to full autonomy; “leading” work means completing much of a task from a high-level prompt while a human supervises.
In a separate policy development, 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 proposals would pause training, modification and fine-tuning of large AI systems until a new Department of Artificial Intelligence established rules. Their text includes a ban on systems with the capacity to automate or greatly accelerate AI research and development. The bills are referred to committee and have not advanced, 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 material says no bill text was found. The policy debate has returned to attention after Democrats won both chambers of Congress on November 3, according to the source article. That political development does not change the current status of the introduced bills.
“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
Limits of the Available Evidence
The evidence does not establish a system that independently designs and builds its own successor. At the same time, the company-reported figures and METR’s assessment have limits: METR said much of the evidence it found most informative was redacted, and OpenAI’s figures, as described by The Decoder, came from the company without independent review. The reported measures also cover different tasks and time periods, so they do not amount to a single standardized measure of autonomy across the industry.
METR said it would not be surprised if the true rate of acceleration were closer to 1.3 times or 3 times the earlier rate. The source material does not provide enough detail to independently reproduce that estimate or judge how well it generalizes beyond the specified tasks. It also remains unclear how quickly AI assistance will expand into research decisions that currently remain under human control. Patel’s forecast about future progress should be read as commentary, not as a confirmed development.
Evidence and Bills Await Review
The next useful evidence would include clearer, independently reviewable measures of what AI systems do in research, how much human intervention tasks require, and whether models can take responsibility for increasingly important parts of development. Further evaluations could help determine whether reported gains persist across tasks and organizations, rather than reflecting company-specific workflows or internal measurement choices.
On the policy side, the Sanders and Casar bills remain in committee, so the source material identifies no immediate change for businesses or users. Their progress, any future bill text for Khanna’s proposal, and additional public assessments of AI research capabilities will shape what can be established next. Until then, the most supportable conclusion is limited: AI is assisting and, in some measured work, leading parts of AI research, while fully autonomous successor development has not been documented in the evidence described here.
Key Questions
Has recursive self-improvement started?
That depends on the definition. AI is assisting and leading some research tasks under human supervision, but the evidence described here does not document AI independently designing and developing its own successor.
What did METR find about Anthropic’s progress?
METR estimated that Anthropic was improving model capabilities on well-specified tasks at around twice its rate from March 2024 to October 2025. It described the estimate as highly uncertain and said the true acceleration could be closer to 1.3 times or 3 times.
What did Anthropic report about Claude?
For August 2026, Anthropic said Claude led 26% of its AI research and development work. The company also said no measured subset of that work ran fully autonomously.
Would the proposed federal ban take effect now?
No. The Ban Artificial Superintelligence Act of 2026 was referred to committee in both chambers and had not advanced, according to the source material. It is a proposal, not a current restriction.
How much of the evidence is independently verified?
METR is an independent evaluator, but it relied on Anthropic’s internal data and said much of the evidence it found most informative was redacted. OpenAI’s figures were company-reported, and The Decoder noted that the material it reviewed had no independent review.
Source: ThorstenMeyerAI.com
