OpenAI 数据:智能体可独立完成 46% 的 2-4 小时研究任务
OpenAI 内部真实用量数据挺难得的:智能体独立完成研究任务的比例从 3 月的 23% 涨到 46%,研究员一天烧 2,400 美元 API 额度当同事用。想知道智能体到底能不能干长活儿、离 RSI 有多近,看这组一手数字比看营销稿靠谱。
OpenAI 公布的最新数据显示,其智能体在零人工干预下,已能完成 46% 需要人类 2 至 4 小时才能搞定研究任务,而今年 3 月这一比例仅为 23%,四个月左右翻倍。
使用强度同样在攀升:OpenAI 研究员每日智能体用量中位数按 API 价格计为 2,400 美元,前 10% 的重度用户达 23,000 美元一天。GPT-6.
2 Astra 的一位预训练研究员还提到,Codex 在共同推演想法的过程中帮忙捕捉到了核心洞见。作者认为这组数字说明我们离 RSI(递归自我改进)的距离比想象中更近:智能体正在从辅助工具变成研究流程中能独立承担长任务的角色。
This is a pretty good insight on how close we are to RSI, read these new numbers from OpenAI:
- Agents now finish 46% of research tasks that take a human 2 to 4 hours, with zero human help. In March it was 23%.
- The median OpenAI researcher uses $2,400 worth of agents per day (at API prices).
The top 10% are at $23,000. And for GPT-6.2 Astra, a pretraining researcher credits Codex with spotting core insights while they worked through an idea together.
pretty impressive
This report is categorized under AITOP's "Industry" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison and AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include OpenAI, Codex, GPT-6.2 Astra, agents.
Source attribution: curated from kimmonismus (@kimmonismus).