OpenAI Reports Exponential Internal AI Use, Calls for Mandatory RSI Metric Disclosure
On Nov 19, OpenAI released a data report on internal "research acceleration," an update and expansion of its first such report in September, offering the first systematic look at how AI is transforming its internal research work. The report shows researchers' AI usage continues to grow exponentially, AI agents can successfully complete ever-longer tasks, yet human involvement remains central to research. OpenAI also called for AI labs to be required to track and publicly report recursive self-improvement (RSI) metrics.
Confirmed
- The report notes that time researchers spend on safety and alignment has increased significantly, with headcount growth in this area outpacing all other directions
- Per chart data relayed by @Elivnurotic, researchers' median per-capita spending on Codex, ChatGPT Work, and Dots (28-day rolling average) has grown exponentially since July 2026, approaching $2,500 per person per day
- Per a chart relayed by @kliu128, the success rate of tasks without human intervention has risen steadily month over month
- Per @daveholtz, agent success rates show an upward trend on EpochAI-related benchmarks
Why it matters
- This is OpenAI's first public disclosure of internal AI usage data, providing a rare quantified window into how much AI accelerates AI research itself
- The call for mandatory RSI disclosure signals OpenAI's position that the capacity for "AI accelerating AI research" should fall under regulatory and public oversight, potentially shaping industry debates on transparency standards
2026-11-19 ~ 2026-11-19 · 6 related posts
Primary sources
- [source] OpenAI report calls for labs to publicly track recursive self-improvement as AI use in research soars — OpenAI · 2026-11-19
- OpenAI researchers' AI spend per person surged exponentially since July, hitting ~$2.4K/day — Eliv_nurotic · 2026-11-19
- OpenAI Updates Its AI Research Acceleration Report: Agents Complete Ever-Longer Tasks — daveholtz · 2026-11-19
3 near-duplicate retellings: kliu128 · deanwball · Moh1tAgarwal