谷歌DeepMind称AI自我改进仍遥远
Google DeepMind 工程师 Haydn Belfield 表示，目前没有观察到 AI 自身推动 AI 研究的可测量加速，自我改进型 AI 依然遥远。但自动化已相当普遍：AI 现在编写超过 75% 的生产代码，约 88% 的代码改动仍需人工监督或审查，完全交由 AI 处理的改动比例从 7.5% 升至 11.9%。在 1,389 次内部 Argon 智能体运行中，任务越复杂，无人工干预的成功率下降越明显；在包含 74 项真实机器学习工程任务的内部基准 RE-Bench 上，Argon 约能完成一半，高于 Gemini 3.7 Flash 的 27%。团队构建的通用能力指数（GCI）覆盖 123 个模型与 92 项基准，显示 Gemini 能力以每月约 0.1 个 GCI 点稳定增长、未见显著加速，Gemini 4 Argon 略高于趋势线被归因于算力与组织投入而非 AI 驱动的加速。Belfield 称目前对 AI 驱动研究自动化风险的评估为低，但正在建设测量基础设施；与此同时，包括 Geoffrey Hinton、Yoshua Bengio 和 OpenAI 的 Jakub Pachocki 在内的 20 多名研究者警告可能出现“智能爆炸”。
这场关于自我改进 AI 是否临近的公开分歧，为判断 AI 研发自动化速度与 AGI 时间表提供了来自头部实验室的内部数据。
Google says self-improving AI isn't anywhere close
Matthias Bastian View the LinkedIn Profile of Matthias Bastian
Oct 14, 2026
Google engineer Haydn Belfield says self-improving AI remains far off, even as significant automation is already taking hold.
Google Deepmind has seen no measurable acceleration of AI research driven by AI itself, according to Belfield. That doesn't mean significant automation isn't happening already: AI now writes more than 75 percent of production code. But about 88 percent of code changes still need human oversight or review, and every change is reviewed by a human before submission, according to Belfield. The share of code changes fully delegated to AI rose from 7.5 to 11.9 percent.
#Google DeepMind#Gemini 4 Argon#Gemini 3.7 Flash#Argon#RE-Bench#General Capability Index#Haydn Belfield#Geoffrey Hinton#Yoshua Bengio#Jakub Pachocki#OpenAI