人工智能在软件工程中的整合正在迅速扩大，Anthropic的Claude聊天机器人在5月份生成了该公司五分之四以上的已发布代码，而此前这一比例仅为个位数。此外，Anthropic的联合创始人估计，到2028年底，人工智能有60%的概率能够进行自主递归自我改进（RSI）。

人工智能在优化自身开发方面已经展示出显著能力，例如谷歌的AlphaEvolve节省了0.7%的计算能力，并将Gemini的训练速度提升了1%。此外，一个人工智能代理在短短七天内将一个相当于GPT-2的模型的训练时间缩短了18%，超越了顶尖人类研究员的工作成果。

虽然这些进步可以将研发效率提高千倍，但它们也引发了人们对人类可能失去对自我改进模型控制的安全担忧。然而，诸如数据中心计算能力有限以及生成高质量训练数据的困难等物理制约因素将继续限制自主递归自我改进的速度

The integration of artificial intelligence in software engineering is expanding rapidly, with Anthropic's Claude chatbot generating more than four-fifths of the company's published code in May, compared to low single-digits previously. Furthermore, Anthropic's co-founder estimates a 60% probability that AI will be capable of autonomous recursive self-improvement (RSI) by the end of 2028.

AI has already demonstrated significant capability in optimizing its own development, such as Google's AlphaEvolve saving 0.7% of computing power and speeding up Gemini training by 1%. Additionally, an AI agent reduced the training time of a GPT-2 equivalent model by 18% in just seven days, surpassing the work of a top human researcher.

While these advancements could boost R&D productivity thousand-fold, they also trigger safety concerns regarding a potential loss of human control over self-improving models. However, physical constraints like limited data-center compute capacity and the difficulty of generating high-quality training data will continue to throttle the pace of RSI.

Source: How artificial intelligence got better at building itself

Subtitle: What does “recursive self-improvement” mean for the technology?

Dateline: 6月 11, 2026 04:21 上午