GLM Builds Recursive Self-Improving Inference Infrastructure

Zhipu AI revealed that GLM-5.3-Flash's production inference runs on over 100,000 domestic AI accelerators, with throughput improved 3.2x in under two weeks, largely by an AI agent powered by GLM-5.3 itself. The agent used a dense feedback system to fix three real engineering issues: precision drift, concurrency bottleneck, and redundant computation.
Key insights
GLM-5.3-Flash Deployed on 100,000+ Domestic Accelerators
Zhipu AI disclosed that all production inference for GLM-5.3-Flash now runs on over 100,000 domestic AI accelerators. From first model run to production deployment took less than two weeks, with end-to-end throughput improved by 3.2x. The optimization work was performed by an AI agent driven by GLM-5.3 itself, not by human engineers.
Dense Feedback System Enables Agent to Debug Complex Systems
The AI agent used a layered verification interface called 'dense feedback' to identify issues. It provided correctness feedback, system behavior feedback (where time was spent), and performance feedback (which solution is better under what conditions). Each signal was local, low-cost, and objectively verifiable. This allowed the agent to discover and fix three real engineering problems.
Three Engineering Issues Fixed by the AI Agent
First, precision drift in the KDA kernel due to TF32 rounding errors accumulating through chain state matrix merging, fixed by contributing to Flash Linear Attention. Second, a concurrency bottleneck where KV cache transfer never overlapped with computation due to GIL not being released in DeepEP's intra-node path, reducing overhead from over 30% to under 1%. Third, redundant computation in a decoding kernel that repeated normalization four times, accelerating it by 1.71x.
CEO Tang Jie: Model Optimizing Its Own System, Recursive Loop Exists
Zhipu CEO Tang Jie emphasized that the main workforce was the model itself, helping optimize the system that runs it. He noted that the agent rarely got stuck because it couldn't write code, but because it didn't know why performance degraded. The dense feedback system turned implicit engineer experience into explicit signals. He stated that the smallest recursive self-improvement loop now exists: the model optimizes the system, and the system runs the model.
AI Self-Improvement Levels Framework: 5 Levels Defined
A separate paper by over 30 researchers defined a 5-level framework for AI self-improvement: Level 1 (human decides what to improve) to Level 5 (AI rewrites the process of creating next-gen AI). A review of 491 papers found 75% at Levels 1-2 and less than 6% at Level 5. This framework provides a taxonomy for tracking progress toward recursive self-improvement.
Engineers Shift from Problem-Solvers to Feedback Designers
Tang Jie emphasized that while goal setting, feedback environment construction, and high-risk change review remain human responsibilities, the engineer's role is changing. Engineers now design feedback systems rather than directly solving optimization problems. Each task completed by the agent becomes training material for its successor, creating a virtuous cycle.