翁荔重新加入OpenAI，将领导一支团队深入探索「递归自我改进」研究，让AI参与自身训练和下一代模型的改进[1]。与此同时，英伟达向Ilya Sutskever创办的Safe Superintelligence投资，并通过部署新一代Vera Rubin系统将其算力扩大一个数量级[1]。这两项举措标志着AI领域正在加速推进自动化AI研发的技术探索。
然而，这一进展也引发了业界的谨慎态度。来自OpenAI、Anthropic、Google、Meta等公司的超过1100名从业者联署声明，要求提前建立控制「自动化AI研发」速度的技术与协调工具[1]。参与联署的行业人士包括Ilya Sutskever、OpenAI首席科学家Jakub Pachocki、Anthropic联合创始人Jack Clark和Jared Kaplan[1]。翁荔曾在7月初发布长文《Harness Engineering for Self-Improvement》，阐述了通过让AI改造工作环境来实现自我改进的技术路径[1]。
Weng Li has rejoined OpenAI to lead a team focused on accelerating internal research into "recursive self-improvement," an approach that enables artificial intelligence systems to participate in training and refining their own next-generation models [1]. This research direction aligns with broader industry efforts to explore novel pathways for AI iteration and evolution.
In parallel support of this vision, Nvidia has made an investment in Safe Superintelligence, the company founded by Ilya Sutskever, and is providing the organization with access to its latest-generation Vera Rubin computing systems, expanding its computational capacity by an order of magnitude [1]. Sutskever now leads the effort to advance what industry insiders refer to as automated AI development methodologies.
However, the rapid acceleration of such research has prompted significant concern within the AI community. More than 1,100 practitioners from OpenAI, Anthropic, Google, Meta, and other major organizations have signed a joint statement calling for the establishment of technical measures and coordination mechanisms to control the pace of automated AI research [1]. The signatories include OpenAI's Chief Scientist Jakub Pachocki, Anthropic co-founders Jack Clark and Jared Kaplan, and Sutskever [1]. Earlier in July, Weng Li published an extensive essay titled "Harness Engineering for Self-Improvement," outlining pathways through which AI systems could enhance themselves by reshaping their operational environments [1].
