XNews

GLM recursive self-improvement infrastructure article

Zhipu AI announced GLM-5.3-Flash runs on over 100,000 domestic accelerators, with throughput improved 3.2x via self-optimization. An agent identified and fixed three real engineering issues (precision drift, GIL bottleneck, redundant computation), achieving 1.71x speedup on one kernel. The company calls it a minimal loop toward recursive self-improvement.

Key insights

GLM-5.3-Flash deployed on mass domestic accelerators, throughput up 3.2x

All production inference for GLM-5.3-Flash now runs on over 100,000 domestic AI accelerators. From first run to production took under two weeks, with end-to-end throughput improved by 3.2 times. The optimizations were performed by an AI agent driven by GLM-5.3 itself.

Tough deployment conditions on domestic accelerators

The domestic accelerators have limited memory and bandwidth, with immature software ecosystems and missing documentation. Optimizations included memory-compute tradeoffs (ReplaySSM), communication-compute tradeoffs (tensor parallelism), precision tradeoffs (INT8/FP8/BF16 hybrid cache), and architecture separation (encode-prefill-decode split). Final token cost matched mainstream NVIDIA GPUs.

Dense feedback mechanism enables agent to debug like a senior engineer

The agent collected dense feedback via hierarchical verification interfaces: correctness, system behavior (time spent), and performance (which strategy is better under what conditions). This allowed the agent to localize problems where sparse rewards (e.g., throughput drop) alone would not suffice. The approach mimics the 'process reward' intuition of experienced engineers.

Three real engineering issues fixed by the agent

The agent discovered and fixed three problems: (1) precision drift in KDA kernels due to TF32 rounding errors accumulating over long contexts; (2) a concurrency bottleneck where KV cache transfer and computation did not overlap because DeepEP's intra-node path did not release the GIL; (3) duplicate normalization in a decoding kernel that recomputed the same normalization four times. The fix for (3) achieved 1.71x speedup.

Human engineers still set goals and review high-risk changes

Zhipu CEO Tang Jie emphasized that human engineers remain responsible for goal setting, feedback environment construction, and review of high-risk changes. However, the role of engineers is shifting from problem solvers to feedback designers. The agent's completed tasks become training material for the next generation of models.

Minimal loop toward recursive self-improvement achieved

Tang Jie stated that while they are far from full recursive self-improvement, the smallest closed loop now exists: the model optimizes the system, and the system runs the model. The layered verifiable feedback environment built on real infrastructure tasks is exactly what training next-generation models requires.

Source posts