Anthropic: 26% of its AI R&D Led by Claude Under Supervision

Key Takeaways

1Claude leads 26% of Anthropic's AI R&D, under human supervision

20.002% of over a billion decisions were blocked in August, with about 50 cases being reviewed by humans each week

3Anthropic publishes these metrics to assess the proximity to recursive self-improvement and calls for more industry transparency

Why it matters : These figures serve as a public benchmark on the progress towards AI capable of self-improvement, amid ongoing discussions about safety and governance.

Anthropic publishes operational oversight and automated blocking figures while calling for greater transparency among laboratories. At the same time, the company reports that Claude accounts for 26% of its AI R&D, still under human supervision. These measures are presented as a benchmark to assess the proximity of recursive self-improvement, as the debate over safety intensifies.

0.002% of over a billion blocked decisions in August

Anthropic indicates that its automated monitoring systems examine the actions of its AI, and in August, 0.002% of over a billion analyzed decisions were blocked, which is about one in 47,000. Approximately 50 priority cases are escalated each week for human review. The company argues that publishing such metrics provides the public, third parties, and governments with better visibility into the pace of AI development in leading laboratories, resulting in improved measurement, public reporting, and the opportunity for society to decide how to use this information. It suggests that other laboratories publish comparable figures and allow for third-party verification.

Claude accounts for 26% of Anthropic's R&D with supervision

According to Anthropic, Claude now handles 26% of the company's AI R&D activities. The company states that the term "lead" means performing the majority of a task based on a general instruction while remaining under human oversight. It is clarified that Claude does not operate completely autonomously and that no measured segment of the R&D work is performed entirely independently by the AI. The share of work at or above the "AI collaborates" threshold exceeds 90%. The company adds that in March, this proportion of tasks led by Claude was 1%, or even less than 1% by its own terms. Claude is thus taking on an increasingly prominent role in the work that feeds the company's next generation of AI.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

These indicators serve to gauge recursive self-improvement

Anthropic presents the tracking of metrics as a way to evaluate how close AI is to autonomous improvement. The company defines recursive self-improvement as a model that fully and autonomously builds its successor. It warns that models accelerating their own development could make human understanding or control more difficult. Therefore, it considers it important to share metrics to appreciate the proximity of this dynamic.

Internal resignation and call to coordinate a slowdown

Last week, a researcher at Anthropic left the company, criticizing leading AI players for "playing with our lives" by developing increasingly powerful models. In this context, Dario Amodei invites the most advanced AI companies to collectively organize a slowdown in development to allow for better implementation of safety measures.