Inside Zhipu's dense-feedback InfraAgent: an early blueprint for recursive self-improvement
量子位 · wechat · 2026-09-17
QbitAI relays Zhipu chief scientist Tang Jie's technical blog in full, detailing how a GLM-5.3-driven InfraAgent helped build and optimize a production inference system on 100k+ domestic accelerator cards. GLM-5.3-Flash ran anonymously as Ox-Alpha on OpenCode and OpenRouter, handling 62+ trillion tokens in six days.
The core technique is "dense feedback": organizing correctness tests, execution traces, and micro-benchmarks into the Agent's iteration loop, with feedback that is local, timely, and objectively verifiable—solving why end-to-end metrics alone can't explain regressions. Three case studies:
- Correctness: the Agent found TF32 default precision causing error accumulation in KDA context-parallel paths; fixed with tf32x3, merged upstream to FlashLinearAttention.
- System behavior: the Agent traced a Python GIL bottleneck in DeepEP v1.2.1's intranode dispatch/combine blocking MooncakeTransfer threads; after the fix, Prefill+KVTransfer overhead dropped from >20% to under 1%.
- Performance: the Agent distilled "optimization skeletons" from SGLang and FlashLinearAttention, restructured a Decode Kernel to eliminate four redundant normalizations for a 1.71x speedup.
Zhipu stresses RSI hasn't been achieved—humans still set goals, boundaries, and review high-risk changes—but each Agent-completed engineering task may become training data for the next model generation.
More from AGI Musings
- "Causal patterns are all you need": MLStreet Talk debates whether the hard problem exists — burny_tech · 2026-11-02
- Dev says blog traffic fell from 700k to 7k reads as AI-generated job apps flood market — ericelliott_ · 2026-11-02
- Master foundational math to stay irreplaceable, argues mathematician — TivadarDanka · 2026-11-02
- Narayanan: labs racing toward Recursive Self-Improvement may hit Recursive Self-Extinction of their business model — random_walker · 2026-11-02
- Editor fully resigns from JMLR, arguing AI reviews beat rushed humans and journals need a rethink — roydanroy · 2026-11-02
- Alpha School investigation: 17 campuses on probation, Bill Ackman as shadow lead, Holmes involved — benjaminjriley · 2026-11-02