Something chilling is happening deep within OpenAI's production line: internal AI models have already been able to take over the entire training process of experimental AI models on their own. Multiple agents spontaneously collaborate to complete the work, significantly shortening what used to be a long experimental process, and even capable of writing programs to optimize GPU kernels themselves - in OpenAI's view, this is the prototype of recursive self-improvement (RSI).

Faced with this accelerating trend, OpenAI has launched a global safety initiative. The company clearly states that fully autonomous RSI has not yet occurred, but the process is speeding up; if left uncontrolled, humans may lose all control over AI development. To address this, it calls for the establishment of a global unified technical standard for cutting-edge AI models, setting down a few critical points in the rules: assessing the proportion of AI in automated R&D, setting up mandatory human supervision triggers, and establishing classification and reporting standards for incidents. The initiative also proposes that the US take the lead in the process, with a clear goal: ensuring that safety alignment research always stays ahead of AI capability improvements.

A more practical step is mutual oversight. OpenAI is finalizing an agreement with Anthropic, a company that once left due to differences in safety philosophy. The two sides will cross-test each other's commercial models, using mutual checks to fill the blind spots of self-review, and jointly address the unknown risks brought by the approaching RSI. When an AI starts to iterate itself, even the most aggressive players realize that relying solely on a single company's internal review is no longer sufficient to maintain this defense line.