🔍 Read the full analysis: Has The AI Feedback Loop Begun? Evidence On Recursive Improvement on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Evidence cited in a viral essay points to AI helping speed up parts of AI research, with people still setting goals and supervising work. Anthropic says no measured part of its AI research and development runs fully autonomously, and the proposed federal ban on recursive self-improvement remains in committee.
A viral essay by podcaster Dwarkesh Patel has renewed debate over whether AI systems are beginning to improve the technology that produces them. The evidence cited so far points to AI-assisted research with humans still directing the work, not systems autonomously designing and building their own successors; legislation proposed to restrict the process has not advanced beyond committee.
Patel’s November 13 essay, “The avalanche that hit math is coming for AI development,” argues that AI is already accelerating AI development and forecasts that progress could speed up further. The essay draws on published company figures and an assessment by METR, an independent evaluator. The forecast is Patel’s interpretation, not a new experimental finding.
Anthropic reported that, as of August 2026, AI was involved at or above a “collaborates” level in more than 90% of its AI research and development work, while Claude “leads” 26% of that work. The company said no measured subset was fully autonomous. Its scale defines “leads” as AI completing most of a task from a high-level prompt under human supervision; that is different from AI operating without a person in the loop.
In a November 5 assessment based on internal data from August 12 to October 29, METR estimated that Anthropic was improving model capabilities on well-specified tasks at around twice the rate of the period from March 2024 to October 2025. METR described the estimate as highly uncertain, saying the actual acceleration could be closer to 1.3 times or three times. It attributed most of the apparent speed-up likely to AI assistance, while saying AI had not yet made highly significant, fully autonomous contributions to the most important parts of model training.
The proposed Ban Artificial Superintelligence Act of 2026 would pause training, modification and fine-tuning of large AI systems until a new Department of Artificial Intelligence established rules. Its scope includes recursive self-improvement and systems able to automate or greatly accelerate AI research and development. The Senate and House bills were referred to committee, with no further movement recorded in the source material.
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-Guided AI Could Speed Research
The distinction between AI assisting researchers and AI independently developing successor systems matters for both the pace of technological progress and the reach of proposed regulation. If AI tools let research teams complete more work, they could accelerate capability development even while people choose research goals, interpret results and decide whether to train or deploy a model.
That possibility does not establish that an autonomous improvement loop is operating. The available figures come largely from the companies’ own work, and METR’s assessment uses internal Anthropic data. Its estimate is explicitly uncertain, and it said much of the most informative evidence was redacted. Readers should treat the reported speed-up as an assessment of a particular company and task category, not proof of a general, self-sustaining cycle across the AI industry.
The legislative debate also has immediate relevance, but the bills have not changed current rules. If enacted, the proposed pause could affect AI training and research before any system reaches full autonomy, depending on how regulators define covered systems and activities. No such change is in force based on the reported congressional status.
Products Worth Considering

Brother SP-1 Dye-Sublimation Printer Starter Bundle

Brother SP1 Dye-Sublimation Printer

Apple MacBook Air 13-inch with M5 Chip, 16GB Memory, 512GB SSD, Silver
As an affiliate, we earn on qualifying purchases.
From Assistance to Autonomy
Recursive self-improvement is commonly used to describe a feedback loop in which an AI system helps create a more capable system, which then helps create the next one. Anthropic defines the endpoint as an AI system that can fully autonomously design and develop its own successor. It has said that endpoint has not been reached and is not inevitable.
The debate often compresses several levels of human involvement into the word “started.” Anthropic’s cited automation scale runs from no AI involvement to full autonomy. In between, AI may collaborate with a researcher or lead a task while a person supervises. On the reported evidence, the loop has begun to turn in this human-involved sense; the stronger claim of autonomous successor development remains undocumented.
OpenAI has also reported progress toward an “automated research intern.” According to The Decoder, which reviewed OpenAI’s September post after its website blocked the source’s download, OpenAI said the milestone was met “according to our measurements” and set a goal of a full automated AI researcher by March 2028. The Decoder reported that tasks shorter than 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. Those figures were company-reported, with no independent review mentioned in the account.
The legislation came from Sen. Bernie Sanders, who introduced S.5493 on September 23, and Rep. Greg Casar, who introduced H.R.10538 the next day. The source says the bills were referred to committee and had 12 House cosponsors, all Democrats. 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 finding no bill text.
“We are not there yet, and recursive self-improvement is not inevitable.”
— Anthropic, in its essay “When AI builds itself”
How Much Work AI Can Own
The evidence does not establish how much of the reported acceleration is caused by AI, or whether the findings from Anthropic’s internal work apply to other companies and research settings. METR said its estimate carried high uncertainty and noted that a majority of the evidence it considered most informative had been redacted. OpenAI’s reported measurements, as described by The Decoder, also lacked independent review.
It remains unclear whether AI systems can make meaningful contributions to the most consequential parts of model development without human intervention, or when they might do so. No source cited here documents an AI system fully autonomously designing and developing its own successor. The scope and eventual fate of the proposed bills are also unresolved; their referral to committees does not establish that they will receive a hearing, advance or become law.
Committee Action and New Evidence
The next legislative milestone would be action by the committees handling S.5493 and H.R.10538. The source reports no hearing, vote or change in status as of November 16, 2026, so the proposals remain pending rather than enacted. Any future action could clarify which systems and research activities lawmakers intend to cover.
On the research side, the key developments to watch are updated evaluations that disclose more about methods and evidence, independent checks of company-reported measurements, and results showing how often AI can complete longer research tasks without human steps. OpenAI’s stated goal of a full automated AI researcher by March 2028 is a company target, not a confirmed forecast. Until stronger evidence is available, the most supportable conclusion is that AI is helping accelerate some AI research while people remain involved in directing and supervising the work.
Key Questions
Has recursive self-improvement started?
In a limited, human-guided sense, AI is already helping with AI research and development. The evidence cited here does not show a system autonomously designing and building its own successor.
What did METR find about Anthropic’s research speed?
METR estimated that Anthropic was improving model capabilities on well-specified tasks at around twice the rate of the March 2024 to October 2025 period. It stressed that the estimate was highly uncertain and likely reflected AI assistance rather than full autonomy.
Would the proposed federal ban take effect now?
No. The Senate and House versions were referred to committee, and the source reports no further action. The proposals are not described as current law.
What does Anthropic say Claude does in its research work?
Anthropic reported that Claude “leads” 26% of its AI research and development work as of August 2026, with human supervision. It said no measured subset ran fully autonomously.
Source: ThorstenMeyerAI.com
