刚刚，唐杰发布智谱RSI首个成果

TL;DR - Zhipu reports that a GLM-5.3-powered infrastructure agent helped build and optimize the production inference stack for GLM-5.3-Flash on a cluster of more than 100,000 domestic accelerators, raising end-to-end throughput to 3.2 times the initial baseline in under two weeks. The company frames this as an early, human-supervised form of recursive self-improvement—not autonomous RSI.

• The agent used “dense feedback” from correctness tests, traces, runtime events, microbenchmarks, and service metrics to form hypotheses, modify code, and iteratively validate results.
• It identified a Python GIL bottleneck that prevented KV-transfer work from overlapping with DeepEP operations, reducing the Prefill-plus-transfer overhead from over 20% to below 1%.
• It fixed accumulated numerical errors in a context-parallel KDA path by using higher-precision tf32x3 operations.
• By reorganizing repeated normalization and gating computations in a KDA Decode kernel, the agent achieved a 1.71× kernel-level speedup.