谷歌：AI 自我改进还很遥远
论证要点
- DeepMind 未观察到 AI 驱动的 AI 研究可测量加速；AI 已写超 75% 生产代码，但约 88% 代码变更仍需人工监督或审查，完全委托 AI 的代码变更占比从 7.5% 升至 11.9%。
- 任务越复杂耗时，AI 越难独立完成：1389 次内部 Argon 智能体运行中，无人工输入的成功率随复杂度上升而急剧下降。
- 人类对创造性研究、新实验设计、诊断细微训练不稳定性和长期规划仍不可或缺；对 12 位 DeepMind 资深研究员的调查证实，自动化在定义明确、可验证的任务上最高，在开放式科学构思上最低。
- 在包含 74 个真实 ML 工程任务的内部基准 RE-Bench 上，Argon 解决约一半，远高于 Gemini 3.7 Flash 的 27%，但距完全自动化所需的近乎完美表现仍很远。
- 团队构建覆盖 123 个模型和 92 个基准的通用能力指数（GCI），Gemini 能力以约每月 0.1 GCI 点的稳定速度增长，无显著加速；Gemini 4 Argon 略高于趋势线归因于更多算力和组织投入，而非 AI 驱动的加速。
- 团队目前将 AI 驱动研究自动化的风险评估为低，但正在建设测量基础设施以追踪未来加速迹象；Belfield 的发现与近期研究一致，但包括 Geoffrey Hinton、Yoshua Bengio 和 OpenAI 的 Jakub Pachocki 在内的 20 多位 AI 研究员警告可能出现“智能爆炸”。
对实践的启示
对做 AI 应用的人：当前 AI 适合定义明确、可验证的任务，复杂开放式工作仍需人类主导；别指望 AI 自我加速研究，应投资于人类监督和评估基础设施。
作者承认的反方与局限
20 多位 AI 研究员，包括 Geoffrey Hinton、Yoshua Bengio 和 OpenAI 的 Jakub Pachocki，警告可能出现“智能爆炸”，认为 AI 系统可在几年内自动化所有 AI 研发，将数年进展压缩到几个月。
原文摘引
Google Deepmind has seen no measurable acceleration of AI research driven by AI itself, according to Belfield.
The more complex and time-consuming a task gets, the less likely AI is to handle it on its own.
More than 20 AI researchers, including Geoffrey Hinton, Yoshua Bengio, and OpenAI's Jakub Pachocki, recently warned of a possible "intelligence explosion" in a research paper.
▸ 展开原文全文（11 段，抓取自 The Decoder）
Google says self-improving AI isn't anywhere close
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.
The more complex and time-consuming a task gets, the less likely AI is to handle it on its own. Across 1,389 internal Argon agent runs, the success rate without human input dropped sharply as complexity increased. Humans remain essential for creative research, novel experiment design, diagnosing subtle training instabilities, and long-term planning, Belfield said. A survey of twelve senior Deepmind researchers confirmed the pattern, with automation running highest for well-defined, verifiable tasks and lowest for open-ended scientific ideation.
On Deepmind's in-house benchmark RE-Bench, which includes 74 realistic ML engineering tasks, Argon solves about half of them. That's a big jump from Gemini 3.7 Flash at 27 percent, but it's still far from the near-perfect performance full automation would require.
Google's own capability index shows no speedup in AI research
To track progress systematically, the team built a General Capability Index (GCI) covering 123 models and 92 benchmarks. Gemini model capabilities are growing at a steady pace of about 0.1 GCI points per month, with no significant change in speed. The team attributes Gemini 4 Argon sitting slightly above the trend line to more compute and organizational investment rather than AI-driven acceleration.
The team currently rates the risk from AI-driven research automation as low but is building the measurement infrastructure to track signs of acceleration going forward.
Belfield's findings align with recent research, but not everyone in the field agrees. More than 20 AI researchers, including Geoffrey Hinton, Yoshua Bengio, and OpenAI's Jakub Pachocki, recently warned of a possible "intelligence explosion" in a research paper. They argued that AI systems could automate all AI research and development within a few years, compressing what would normally be years of progress into just months.
AI News Without the Hype – Curated by Humans
Subscribe now