Has Recursive Self-Improvement Started? A Look At The AI Evidence
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: Has Recursive Self-Improvement Started? A Look At The AI Evidence on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get movie nights 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 from Anthropic and an assessment by METR indicates that AI tools are helping researchers develop AI faster, while people still set goals and oversee key work. The stronger claim—that an AI system can autonomously design and build its successor—has not been documented in the material reviewed. A proposed U.S. bill to restrict such development remains in committee.

AI is helping speed up AI research, but the evidence cited in a viral essay does not show systems independently designing and building their own successors. The debate returned to public attention this week after podcaster Dwarkesh Patel argued that AI is already accelerating the development of AI, while a proposed U.S. bill that would restrict parts of that process remains in Congress.

The distinction is between AI contributing to work directed by people and a system carrying out the full cycle of improving itself. Anthropic reported that, as of August 2026, Claude led 26% of its AI research and development work. The company said more than 90% of that work involved AI at or above a collaboration level, but that no measured subset was fully autonomous. The figures are company-reported, not an independent audit of all AI research.

In a November 5 assessment, the nonprofit METR said Anthropic was likely improving model capabilities on well-specified tasks at around twice the rate seen from March 2024 to October 2025. METR attached high uncertainty to that estimate, saying the actual acceleration could be closer to 1.3 or three times. It judged that AI assistance likely accounted for most of the speed-up, while saying AI was not yet making highly significant, fully autonomous contributions to the most important parts of model training.

Patel’s November 13 essay, “The avalanche that hit math is coming for AI development,” draws on METR’s assessment and figures published by AI companies, amid accelerating efforts toward recursive self-improvement in AI. Its forecasts about what may follow are Patel’s analysis, not a new experimental result. Separately, legislation introduced by Sen. Bernie Sanders and Rep. Greg Casar would restrict training, modification and fine-tuning of large AI systems, including through recursive self-improvement, until a new federal agency sets rules. Both bills are still in committee.

At a glance
reportWhen: Debate renewed the week of November 16,…
The developmentA viral essay has renewed debate over recursive AI self-improvement, drawing on existing company figures and a METR assessment rather than presenting new evidence.

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

AI Gains Still Depend on Researchers

The evidence matters because AI-assisted research could shorten the time needed to improve models, even if researchers remain in charge. Faster work on well-specified tasks may affect how quickly companies develop and test systems. METR’s estimate points to a possible acceleration, but its stated uncertainty and reliance on internal Anthropic data limit how firmly readers can interpret the scale or cause.

That distinction also shapes the policy debate. The Sanders-Casar bill would reach beyond systems that autonomously build successors: its text would also cover systems with the capacity to automate or greatly accelerate AI research and development. That could make the boundary between ordinary research assistance and a prohibited capability consequential for laboratories and developers if the bill advances. No restrictions from these proposals are in force now, and the evidence presented does not establish that full autonomous self-improvement has arrived.

As an affiliate, we earn on qualifying purchases.

From Assistance to Autonomy

Recursive self-improvement, or RSI, describes a feedback loop in which an AI system helps develop a more capable AI system, which can then help with further development. Anthropic defines the endpoint as an AI system able to fully autonomously design and develop its own successor. It has said that endpoint has not been reached and is not inevitable.

The work between no AI involvement and full autonomy can be divided into degrees. Anthropic uses an automation scale in which AI can collaborate with people or lead a task under human supervision. Its August 2026 figures place Claude in the “leads” category for 26% of its AI research and development work, while stating that humans remain involved in every measured subset. That is evidence of substantial assistance, not proof of an autonomous self-improvement loop.

OpenAI has also described progress toward automated research. According to The Decoder, which reviewed an OpenAI blog post that could not be accessed directly for this report, the company said it had reached its measured “automated research intern” goal and aimed 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. The figures come from OpenAI’s own reporting, with no independent review identified by The Decoder.

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

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

Evidence Has Important Limits

Full autonomous self-improvement has not been documented in the evidence described here. METR’s assessment relied on internal Anthropic data, carried high uncertainty and said Anthropic had redacted a majority of the evidence METR considered most informative. Its conclusions therefore offer an outside evaluation, but not a fully transparent record that readers can independently reproduce.

The OpenAI measures are also company-reported. The Decoder noted that it found no independent review of those figures. The available information does not establish whether gains at either company generalize across the industry, how much acceleration can be attributed specifically to AI assistance, or when systems might take over more consequential research decisions. Patel’s predictions should be read as forecasts, not confirmed outcomes.

Legislative status is also distinct from policy in force. The Sanders-Casar proposals have been referred to committees, and the supplied status information lists 12 House cosponsors, all Democrats. A separate proposal associated with Rep. Ro Khanna was reported by CNBC as a summary; no bill text was identified in the source material. The timing and scope of any future action remain unknown.

Research and Bills Face Further Tests

The next evidence to watch is whether independent evaluators can inspect more detailed data and determine how much AI tools contribute to research progress over time. Further company disclosures could clarify what tasks models complete, how often people intervene and whether researchers retain control over priorities and deployment decisions.

In Congress, the Sanders-Casar bills would need to advance beyond committee before they could change rules for developers or users. The proposals’ current status does not establish that a vote is scheduled or that they will become law. For now, the grounded conclusion is narrower than the viral claim: AI is assisting and accelerating parts of AI research, while fully autonomous design and development of a successor remains unshown in the cited evidence.

Key Questions

Has recursive self-improvement started?

That depends on the definition. AI is helping people conduct AI research, but the cited evidence does not show an AI system autonomously designing and building 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 earlier rate. It described the estimate as highly uncertain and said AI assistance likely explained most of the acceleration.

Does Claude do Anthropic’s AI research autonomously?

Anthropic reported that Claude led 26% of its AI research and development work as of August 2026, but said it was not fully autonomous in any measured subset.

Would the proposed U.S. bill ban AI research assistance?

The Sanders-Casar bill would cover systems able to automate or greatly accelerate AI research and development, as well as training and modification through recursive self-improvement. The bill remains in committee and is not current law.

What is still unknown?

The evidence does not establish how broadly the reported acceleration applies, how much AI itself caused it, or when systems could take over key research decisions. Some underlying company data is internal or redacted.

Source: ThorstenMeyerAI.com

You May Also Like

Debunking The Myth: ‘Not American’ Isn’t A Valid AI Sovereignty Marker

Analysis clarifies that ‘not American’ isn’t a valid measure of AI sovereignty, emphasizing legal distinctions and the limits of proxy measures.

Cloud Costs and Budgets: The FinOps Playbook

Ineffective cloud cost management can drain budgets—discover essential FinOps strategies to optimize spending and stay in control.

One-click Employee Offboarding For Startups Without IT

A new tool enables small startups to revoke employee access across SaaS platforms instantly, streamlining offboarding without dedicated IT teams.

Anthropic Sets IPO Price Range At $70–$78 A Share: What Its Filing Reveals About Claude’s Maker

Anthropic’s amended SEC filing sets a $70–$78 IPO range and details its share sale, rapid revenue growth, large commitments and voting structure.