Survey finds Chinese AI labs aim recursive self-improvement at training and serving systems, not model weights

Pandaily, summarising a Leiphone survey of recent disclosures, lists Zhipu's GLM-5.3 Infra Agent building GLM-5.3-Flash serving in two weeks, DeepSeek's DSec producing about 3 million agent sandboxes a day, Moonshot using more than 51 million sandboxes for Kimi K3 and Alibaba saying Qwen3.8-Max ran its own training pipeline for 33 iterations; these are company claims and none involve AI rewriting weights.