AI recursive self-improvement might not come so quickly after all
Source: https://www.technologyreview.com/2026/08/18/1142188/ai-recursive-self-improvement
Author: MIT Technology Review
Published: 2026-08-18
Platform: news
中文概要
MIT Technology Review 转述 Princeton 大学 Peter Kirgis 与 Sayash Kapoor 牵头的多机构研究：当前 AI Agent 已能完成 AI 研究所需的工程子任务，但仍缺少判断力与创造力，难以产出达到顶级 ML 会议录用标准的研究同期社论配合 Arxiv 2607.27191（Can AI agents conduct open-ended AI research?），指出 AI 自动化研究的乐观时间线缺乏证据支撑这是对近期 RSI 加速叙事的直接降温，与本文收录的 PAST-BenchThe Last AI Built by Humans 等递归自改进基准形成对比
English Summary
MIT Technology Review editorial covering the Princeton-led multi-institution study by Peter Kirgis and Sayash Kapoor: current AI agents can complete the engineering sub-tasks needed for AI research but still lack the judgement and creativity to produce work that meets a top ML conference bar. The piece is paired with arXiv 2607.27191 ('Can AI agents conduct open-ended AI research?') and serves as a direct cooldown on aggressive RSI timelines. It contrasts with recently indexed RSI benchmarks such as PAST-Bench, AI4AI-Bench, and 'The Last AI Built by Humans'.