Chinese AI vendors advance RSI without rewriting model weights
雷峰网 AI surveyed Chinese model vendors’ recursive self-improvement efforts across inference infrastructure, training frameworks, GPU kernels, sandboxes, agent harnesses and reinforcement learning. The report says these projects form partial engineering feedback loops, while leading systems generally remain between levels one and two of a five-level RSI framework. It also reports that none of the surveyed vendors had allowed AI to rewrite underlying model weights autonomously.
The survey distinguishes practical AI-assisted optimization of AI-building tools from the still-unachieved goal of AI autonomously designing and rewriting its successor models.
This source is a digest of several news items (“一天造300万个沙盒,10万张卡自己写系统:2026年国产 RSI 到底有多野?”). Find this story in the original.