[GLM-5.3 Begins Optimizing Its Own System, Zhipu Tang Jie: The Smallest RSI Closed Loop Has Emerged] Beating AI Newsflash: Zhipu has disclosed that the Infra Agent powered by GLM-5.3 has started participating in optimizing the inference system of GLM-5.3-Flash. It took less than two weeks to transition from the first successful run to handling all production traffic, with end-to-end throughput increasing to 3.2 times the initial level. Tang Jie, co-founder and chief scientist, mentioned during a review that the most common bottleneck for the Agent is not its inability to write code, but rather its lack of understanding of where the problem lies. For instance, a modification might cause throughput to drop by 20%; the Agent knows it made a mistake but doesn’t know which layer is problematic or what to investigate next. To address this, Zhipu has transformed the troubleshooting methods typically used by senior engineers into a set of tools that the Agent can directly utilize. After modifying the code, the Agent can independently check whether the results are correct, identify where time is being consumed, and determine which solution is faster, then continue making adjustments based on the findings. Using this approach, the Agent identified issues such as long-context computation errors, KV transmission blockages, and redundant kernel computations during decoding. One kernel rewrite resulted in a 1.71x speed improvement. Tang Jie predicts that the role of engineers will increasingly resemble that of "feedback designers." Humans will be responsible for setting goals, defining boundaries, and reviewing high-risk modifications, while the Agent will propose hypotheses, modify code, and run experiments independently. Although a complete RSI is still far off, the smallest closed loop of "model optimizing system, system serving model" has already emerged. [Original Link]