Anthropic Says Claude Now Leads 26 Percent of Model R&D – but What "Self-Improvement" Actually Means Remains Unresolved
In a single week in September 2026, the debate about AI improving itself moved from abstract speculation to concrete numbers: Anthropic states that Claude now leads 26 percent of the company's model R&D, OpenAI has announced an automated "research intern," and Alibaba CEO Eddie Wu declared that the path to artificial superintelligence via self-improvement is "increasingly clear." At the same time, technologist Ramez Naam published a data-driven counterargument: the loop is, by his estimate, 5–10 times too weak to sustainably explode. This analysis separates what is verified from what are corporate claims with commercial incentives – and from what are contested estimates.
The Week That Made the Debate Concrete
Between 20 and 27 September 2026, a series of announcements landed that together moved recursive self-improvement (RSI) from future vision to measurable work underway.
Anthropic stated that Claude now leads 26 percent of the company's model R&D, and that this means the model can complete most of a given task "end-to-end from a high-level prompt" – that is, from a high-level instruction – while still operating under human supervision (AP News). This is a company claim: AP notes that the models do not work fully autonomously "at least not yet," and that Anthropic does not say how close to fully autonomous improvement the company is.
OpenAI announced in September an automated "research intern" – a system that, according to the company's own definition, can carry out well-defined research tasks under human direction (same AP article). Here too, it is the company itself that describes what the system can do.
In China, Alibaba CEO Eddie Wu used the keynote at the 2026 Apsara conference to declare that the technical path to artificial superintelligence (ASI) is "increasingly clear," pointing to RSI as the process: models that engage with real tasks and feedback, identify their own limitations, and autonomously design experiments, synthesize data and evaluate results in a continuous self-improvement loop (Unite.AI). Wu also stated that Alibaba's Qwen team is exploring RSI with "meaningful progress," and that the company plans models with 5–10 trillion parameters. All capability and scale claims here are Alibaba's own; neither the RSI progress nor the parameter count is independently verified. (Note: Unite.AI dates the article to 21 September but places the keynote on the conference's 22 September – the exact date is unclear in the source material.)
Meanwhile, according to a Reuters report carried by Lee Enterprises on 20 September, the heads of America's leading AI labs gathered in a rare united call for a pause in development, warning that the technology "could soon improve itself" and slip beyond human control (madison.com). Anthropic CEO Dario Amodei has also written that recursive self-improvement "could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all" – cited in Ezra Klein's op-ed in the New York Times on 20 September (NYT). The op-ed is an opinion piece arguing for stopping the labs from RSI, not neutral coverage.
And then, on 27 September, Ramez Naam published the essay that forms the counterargument (rameznaam.com).
"RSI" Means Different Things to Different Companies
One of the most striking findings in AP's reporting is how divergent the definitions are: leading AI companies define recursive self-improvement differently. Some define it as any form of AI feedback on model improvement; others define it as AI working toward the goal fully autonomously (AP News).
That means when an executive says "RSI is near," it is often unclear what is actually being promised. Anthropic's 26 percent figure describes a model leading research work under human supervision – not a model improving itself without humans in the loop. OpenAI's "research intern" is, by definition, limited to "well-defined research tasks under human direction." Wu's description, by contrast, points toward a continuous autonomous loop – the most ambitious end of the spectrum.
Cornell researcher John Thickstun, assistant professor of computer science, points out that a supportive form of self-improvement has in fact been underway for years: "We have already, for years, been using these models in supportive roles for creating the next version of these models. So people use the past generation of models to write code for the AI systems that then create the next generation" (quoted in AP News). His point is that this has yielded modest results, not creative leaps. If "RSI" is already happening, and has been for a long time without running away, the question is not whether the loop exists – but whether it delivers accelerating returns.
Naam's Counterargument: 5–10 Times Too Weak
This is where Naam's essay comes in. He poses the question its title suggests: Can AI self-improvement overcome diminishing returns?
His main answer is quantitative: "Given our best current data, the AI self-improvement loop would need to be roughly 5–10× stronger to sustain itself, let alone run away." In other words, on the best available data, the loop would need to be roughly 5–10 times stronger just to sustain itself, and even more to run away (Naam, 27 September 2026). This is the author's own analytical estimate – not a verified measurement, and not a consensus. The essay builds on methods and data that are only partially visible in the available material, and the details behind the 5–10× figure must be read in full text before they can be fully assessed.
Naam also presents a taxonomy that is useful for sorting the claims: Type 1 is productivity gains – AI makes researchers faster, but humans are still driving. Types 2–4 are increasing autonomy under still-diminishing returns – the loop becomes more self-sufficient, but each generation yields less. Type 5, the actual runaway to superintelligence, requires something no one has demonstrated: accelerating returns, where each iteration makes the next iteration more than proportionally faster or better.
His conclusion is not that AI progress stops. Quite the opposite: "I expect incredibly rapid AI progress by the standards of nearly any other technology. But the evidence we have doesn't suggest a sudden explosion to incomprehensible superintelligence anytime soon." And he acknowledges he could be wrong.
It is worth noting what Naam does not claim. He does not claim that autonomy is not increasing – Anthropic's 26 percent figure and OpenAI's "research intern" fit neatly into Types 2–3 in his taxonomy. He claims that the increasing autonomy has so far been outweighed by diminishing returns in the research process itself, so the loop does not reinforce itself. The disagreement is therefore not about whether AI gets a larger role in its own development – that is, at bottom, undisputed – but about whether the returns per generation fall or rise.
The Weight of Evidence – Both Ways
A skeptical reading must weigh both sides against each other.
Naam's 5–10× figure is, as noted, an estimate built on a specific reading of data about loop weaknesses – in particular findings the essay links to OpenAI's own research. It is not peer-reviewed, the methods are only partially visible in public material, and the author himself flags uncertainty. At the same time, forecasts have repeatedly underestimated AI progress – something Naam himself concedes – so "diminishing returns have won before" is not proof they will win again.
On the other side: the claims of imminent RSI come from parties with clear commercial incentives. Anthropic and OpenAI compete for capital partly on the basis of precisely these capabilities (briefs.co reported in the same period on IPO discussions around Anthropic – a source with visible technical errors that should be corrected against CNBC/Reuters and is therefore given little weight here). Alibaba is selling a full-stack AI strategy and enormous model plans. "The path to ASI is increasingly clear" is an executive's claim from a conference stage, not a documented result. The Reuters report on the pause call is largely unreadable in the reproduction and can only confirm the main message, not the details.
Perhaps the most illuminating feature of the week is that both camps actually agree on two things: that AI systems are taking an ever-larger role in their own development (Anthropic and OpenAI document it, Naam categorizes it), and that the outcome – control and speed – is highly uncertain (hence the pause call, and hence Naam's caution). The real point of disagreement is narrower than the rhetoric suggests: whether returns per self-improvement generation are diminishing or accelerating.
What Could Settle the Question
The debate does not need to be settled with manifestos. It can be settled with measurable indicators, most of which are in principle already public:
- Autonomy share of R&D. Anthropic's 26 percent figure is a first data point of this kind. If the number rises steadily toward larger shares of the research without correspondingly increased human labor, it supports accelerating returns. If it plateaus at 30–50 percent with humans still driving direction, it fits Naam's Type 2–3 picture.
- Time per iteration loop. Accelerating returns require the time from hypothesis to evaluated next generation to fall faster than difficulty rises. If cycle time plateaus while the capability gain per generation declines, that is diminishing returns in practice.
- Returns per generation. The gain per model generation – on benchmarks, in research productivity, in cost per unit of capability – is the most direct measure of returns. This is the core of Naam's analysis and the area most open to independent verification.
The open questions remain large. None of the parties has published methodology that lets outsiders verify – in full depth – Anthropic's 26 percent, Alibaba's RSI progress, or Naam's 5–10× estimate. There is no independent measurement of returns per generation across the labs. And even if Naam is right about today's loop, that says little about where it lands two years from now – he himself concedes that history gives reason for humility.
What is new after the week of 20–27 September is that the argument is no longer about future visions, but about numbers that can be tracked: a percentage, a parameter scale, a multiplication estimate. The numbers are not necessarily in conflict with each other – most of them could be compatible with Naam's taxonomy – but they cannot yet be verified in full depth. It is the tension between them, not any one of them alone, that will show which direction the loop actually takes.