🔍 Read the full analysis: The Case For Asking Whether AI Can Improve Itself Recursively 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
A November essay by podcaster Dwarkesh Patel has renewed debate over whether AI systems are improving the process of building AI. Available evidence describes AI-assisted research and faster progress on some tasks, but does not document a system independently designing and developing its successor. Two proposed congressional bills would restrict parts of AI development, but neither has advanced beyond committee referral.
A viral essay by podcaster Dwarkesh Patel has reignited debate over whether AI is beginning to improve itself, as research cited in the discussion shows AI helping speed some model-development work but does not document a system independently building its successor. The debate also intersects with two proposed congressional bills that would restrict training and modification of large AI systems; both remain in committee.
Patel’s November 13 essay, “The avalanche that hit math is coming for AI development,” draws on assessments and figures published by AI companies and the nonprofit evaluator METR. Patel’s argument that much faster progress may follow is a forecast, not a newly reported technical demonstration. The distinction matters because “recursive self-improvement” can mean anything from AI tools helping researchers to a system autonomously designing and developing its own successor.
Anthropic reported that in August 2026, Claude led 26% of the company’s AI research and development work. The company said more than 90% of that work involved AI collaboration at some level, while no measured subset was fully autonomous. Anthropic defines the more advanced endpoint as a system able to design and develop its successor without human direction.
METR’s November 5 assessment, based on internal Anthropic data from August 12 to October 29, said the company was likely improving model capabilities on well-specified tasks at around twice its earlier rate. METR attached high uncertainty to that estimate, saying the true acceleration could be closer to 1.3 times or three times. It attributed most of the 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.
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
Why AI-Assisted Research Matters
The distinction between AI helping researchers and AI directing its own development affects how policymakers, companies and the public assess the pace and control of AI progress. Evidence of faster work with human researchers still setting goals and making key decisions does not establish the autonomous feedback loop often associated with recursive self-improvement. But it does suggest AI tools may already be contributing to the process of building more capable systems.
The policy debate is active even though the most sweeping technical scenario has not been demonstrated in the material cited here. The Ban Artificial Superintelligence Act of 2026 would pause specified AI training and modification until a new federal department set rules. The scope of such a measure could reach beyond self-improving systems: its text also covers systems capable of automating or greatly accelerating AI research and development. Neither bill has advanced, so the proposals do not change the rules for businesses or users today.
For readers, the practical question is not simply whether AI can “improve itself.” It is which tasks AI can perform, how much human supervision remains, and whether reported performance gains are independently verified. Those differences shape both the evidence for accelerating development and the reach of proposed regulation.
Popular right now

Vitamix 5200 Blender with 64 oz Container, Black

De’Longhi La Specialista Touch Semi-Automatic Espresso Machine with Grinder

Hisense 43-Inch E6 Cinema Series Hi-QLED 4K UHD Smart Fire TV (43E6QF)
As an affiliate, we earn on qualifying purchases.
From Research Tools to Autonomy
Anthropic’s framework distinguishes levels of AI involvement, from no AI participation to fully autonomous operation. At an intermediate level, AI collaborates with people; at a more advanced level, it leads tasks from high-level instructions while a human supervises. Anthropic’s August figures place some of its research work at the “leading” level, but the company said no measured part operated fully autonomously.
METR’s assessment concerns progress on well-specified tasks, not an across-the-board measure of AI research or proof that a model can replace its developers. The organization said its estimate carried high uncertainty and that Anthropic had redacted much of the evidence it considered most informative.
OpenAI also reported progress toward an “automated research intern” in a September blog post, according to a summary by The Decoder, which said it had reviewed the post. The reported figures included an 86% success rate on tasks taking under 15 minutes without intervention; for successful tasks estimated to require four to eight hours of human work, more than half needed at least one human step. The figures were company-reported, and The Decoder noted no independent review was described. OpenAI said people still set priorities and make decisions about scaling, pauses and deployment.
““We are not there yet, and recursive self-improvement is not inevitable.””
— Anthropic, in its essay “When AI builds itself”
How Far the Evidence Reaches
The available evidence does not establish that any AI system has independently designed and developed its own successor. It does show company-reported AI involvement and a METR estimate of faster progress on specified tasks, but the estimate has a wide uncertainty range and relies on internal Anthropic data. METR also said much of the evidence it found most informative was redacted.
OpenAI’s figures, as relayed by The Decoder, also came from inside the company, with no independent review described in that account. The sources cited do not settle how much of the reported acceleration comes from AI assistance, changes in research methods or other factors. Patel’s expectation of much faster progress remains commentary and forecasting, not confirmation of autonomous self-improvement.
The legislative status is clearer: the Sanders and Casar bills were referred to committees and had not advanced in the latest records described by the source material. Rep. Ro Khanna’s separate “Human Control Over AI Act” was reported by CNBC as a summary, but no bill text was found in the source material. Its detailed scope and status therefore remain uncertain.
Congressional and Research Milestones
The next developments to watch are further independent evaluations of AI’s role in model research, clearer disclosures about the tasks systems can complete without intervention, and any new evidence of autonomous contributions to training or successor design. Companies’ internal measurements may provide useful signals, but external scrutiny would help establish how broadly the results apply.
In Congress, the immediate milestone is whether the committees handling the Sanders and Casar proposals take them up or publish further action. Their present committee status means no new restrictions from those bills are in effect. If the proposals progress, the bill text and any amendments will determine which AI systems and research activities would be covered.
Until then, the most grounded description is that AI is assisting and, in some reported work, leading parts of AI research under human supervision. Whether that assistance will develop into a self-directed cycle—and on what timeline—remains an open question.
Key Questions
Has recursive self-improvement been demonstrated?
The cited material does not document an AI system independently designing and developing its own successor. It describes AI assisting with research and leading some work while people retain key decisions.
What did METR estimate about Anthropic’s progress?
METR said Anthropic was likely 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, with a possible range closer to 1.3 times or three times.
What would the proposed federal AI ban do?
The Ban Artificial Superintelligence Act of 2026 would pause training, modification and fine-tuning of large AI systems until a proposed Department of Artificial Intelligence set rules. It also addresses systems able to automate or greatly accelerate AI research and development. The bills remain in committee.
Do the proposed bills change what companies can do now?
No. According to the source material, the House and Senate bills have not advanced beyond committee referral, so they do not change current rules for businesses or users.
Source: ThorstenMeyerAI.com
