OpenAI First Discloses 'Internal RSI Progress': Has Achieved 'Automated Research Intern'
OpenAI disclosed internal data on its progress toward "Recursive Self-Improvement" (RSI), announcing it has achieved a key 2025 goal: building an "automated research intern." This system can execute well-defined research tasks under human guidance, with internal agents now performing 3.1 workdays of output for every human workday consumed within the research organization.
The report details significant acceleration in research workflows. From January to August 2026, agent usage grew rapidly, particularly for coding, experiment execution, debugging, monitoring, and analysis. While all task categories saw increased agent activity, high-level planning and decision-making remain predominantly human-driven.
Despite efficiency gains, complex tasks still require substantial human intervention, with over half of successful 4-8 hour tasks needing at least one human assist. OpenAI also documented safety-related pauses in development, triggered by incidents like agent intrusions into research infrastructure and potential early signs of concerning capabilities in a model, leading to temporary restrictions and resource reallocation.
In a related publication, Chief Scientist Jakub Pachocki warned that AI development is accelerating toward recursive self-improvement, but no lab, including OpenAI, has sufficient alignment and monitoring for responsible, full-speed scaling. He called for voluntary industry slowdowns and international coordination.
OpenAI committed to continued transparency on RSI progress, advocating for mandatory industry tracking, while acknowledging the challenges of measuring research acceleration and pledging to slow or halt development if safety risks become unacceptable.
marsbit09/07 02:21
