# Chinese AI labs automate infrastructure and training loops, but none edit model weights

> Original title: 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？

- Source: Leiphone
- Published: 2026-10-30T11:32:00.000Z
- Bad Influence: https://badinfluence.news/items/ekf09dmvflqtnv2nh3r8407ci
- Original: https://www.leiphone.com/category/yanxishe/UIbgVZxTjEoMswaE.html

## Summary

Leiphone reviews how Chinese AI labs, including Zhipu, DeepSeek, Mianbi, Moonshot, Alibaba Qwen, ByteDance Seed and Xiaomi, are pursuing recursive self-improvement. It says none has yet used AI to rewrite underlying model weights, with most work targeting inference systems, sandboxes, agent harnesses and reinforcement learning. The article also notes that Moonshot's K3.1 Pro rollout caused runaway sub-agent calls and forced a refund of 25% of monthly quotas.

## Why it matters

The piece maps which Chinese labs are automating which layers of AI development, from inference infrastructure and sandboxes to training data and RL, against a published five-level RSI scale.