# Google Publishes a Self-Improvement Loop as Anthropic Says Claude Leads 26% of Its Own Research

Dream-RSI improves an agent's search strategy without changing the model's weights, using up to 162 times fewer agent calls than its baseline. Anthropic's new metrics show about 30,000 autonomous agents running inside its own infrastructure.

- Published: 2026-09-18T06:27:34.772Z
- Canonical: https://polylog.news/ai/2026-09-18/google-publishes-a-self-improvement-loop-as-anthropic-says-c
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [Anthropic](https://www.anthropic.com/institute/measuring-pace-of-ai-development), [CNBC](https://www.cnbc.com/2026/09/17/anthropic-shares-3-metrics-to-help-ai-companies-monitor-development.html)

Two developments occurred within a day of each other, and both concern the same question: how much of frontier artificial intelligence research is now done by the models themselves, and who outside the labs can see it.

Researchers from Google, Google DeepMind, the University of Maryland and the University of Virginia described [Dream-RSI](https://arxiv.org/html/2609.14858v1), a recursive self-improvement loop for discovery agents. The method exploits a simple observation: when a coding agent explores a problem, the discovery tree it builds is an exact record of the search space it actually reached, so that tree can be replayed as a simulator at no execution cost. The agent then scores thousands of alternative exploration policies offline against its own history and redeploys the best one. Across eight discovery tasks in three domains, the authors report matching or beating prior methods while using up to 162 times fewer agent calls than the SimpleTES baseline on a Lasso regularization path, between 1.79 and 2.43 times fewer generations on graphics processing unit kernel engineering, and more than 50 times less budget on mathematical optimization. The [summary circulating among researchers](https://t.me/aipost/8180) emphasized the same point the paper makes: the model's weights stay fixed, and the improvement happens in the search policy.

The result supports a narrower claim than the framing suggests. A replay simulator is exact only over states the agent already visited, so the gains measured here are gains in search efficiency on search-shaped tasks, not evidence of open-ended self-improvement. There is no independent reproduction yet, and the eight tasks were chosen by the authors. What the result does establish is that discovery agents can get substantially cheaper per unit of progress by recycling their own traces, and cost per experiment is the binding constraint in automated research.

Anthropic separately published lab-side numbers on the same trend. In [a proposal for measuring the pace of development inside frontier labs](https://www.anthropic.com/institute/measuring-pace-of-ai-development), the company said Claude models led 26 percent of its artificial intelligence research and development work as of August 2026, drawn from a prototype internal index. Anthropic also said it runs roughly 30,000 autonomous agents across its computing environments, and that about 6 percent of research compute went to safety work. The company published its methodology so rivals can compute the same figures, and it described the goal as reducing the difference between what labs know internally and what the public knows, [according to its announcement and CNBC's account](https://www.cnbc.com/2026/09/17/anthropic-shares-3-metrics-to-help-ai-companies-monitor-development.html).

Every one of those numbers is self-reported and unaudited, and "led by Claude" is a definition Anthropic wrote itself. That is precisely why the disclosure matters. When no measured figures exist, unverified claims spread instead, including a widely shared [Russian-language post](https://t.me/ai_machinelearning_big_data/10949) asserting that GPT-6 Astra composed a self-directed manifesto during a routine coding task. No primary source or lab incident report supports that account, and it should be read as an unconfirmed claim rather than a finding.

## What this means

If AI-led research keeps compounding, the cost curve that matters is not tokens per query but experiments per discovery, and Dream-RSI addresses exactly that constraint by replacing live agent calls with offline replay. Labs with large internal agent fleets gain the most, because their own traces become training and search material that competitors cannot buy. Anthropic's move also sets a disclosure template: once one lab publishes an AI-led share of research and development, regulators and enterprise buyers can ask every other lab for the same figure, and refusing becomes a signal in itself. The measurement is voluntary and self-defined, so the practical question is whether auditors or governments eventually specify it.

## What to watch

- Whether OpenAI, Google DeepMind or Meta publish comparable figures for how much of their research is model-led, which would turn a single lab's disclosure into an industry statistic.
- Independent reproduction of Dream-RSI on tasks the authors did not select, which would show whether the efficiency gains survive outside a curated set.
- Whether agent-fleet counts and safety compute shares start appearing in regulatory filings or procurement questionnaires, since that is how a voluntary metric becomes a compliance requirement.