Skip to content
WeChat · AI Era (Xinzhiyuan)·Indexed by this site · Originally published· AI updates →· Its place in the scan record →SelectedAttention74

OpenAI report says recursive self-improvement has begun, agent usage has surged, and humans remain the R&D bottleneck

OpenAI自曝递归自我提升!智能体指数级爆发,人类已成瓶颈

AI summary

OpenAI released a research report stating that agents are accelerating AI's own R&D and that a closed loop of recursive self-improvement has taken shape, but humans remain the bottleneck in the R&D loop. The report says that, at API list prices, the median researcher's daily agent spending rose from $1,500 in September to $2,400 in October, with the top 10% of heavy users reaching $23,000, and that agent workdays per person per day increased from 0.5 to 5.7. In terms of capability, it shows a "jagged frontier": from March to October, the zero-intervention success rate on 2-to-4-hour complex R&D tasks rose from 23% to 46%, but the success rate for writing low-level inference and serving-stack code was only 25%. The report also says OpenAI models' internal capability index ECI has risen at a rate of 16.4 points per year since August 2025, above the industry average of 14 points, and calls for top AI labs to track and disclose core metrics of recursive self-improvement.

Why it was selected

The report also discloses the surge in agent usage alongside the wide gap in zero-intervention success rates across tasks, providing a data reference for judging the actual extent to which AI is accelerating R&D.

Read originalmp.weixin.qq.com