GPT-5.6 is optimizing itself, a new bootstrap approach has emerged
OpenAI
OpenAI's technical report reveals that GPT-5.6 is now applied in production to optimize its own runtime environment, including analyzing traffic, adjusting request routing, rewriting low-level kernels, and optimizing speculative decoding. These improvements reduced end-to-end service cost by 20% and increased token generation efficiency by over 15%. Meanwhile, human oversight remains, with optimization goals and deployment decisions controlled by people.
OpenAI published a technical report detailing RSI (Recursive Self-Improvement) applied to GPT-5.6. The model is now deployed in production: it analyzes traffic, adjusts routing, rewrites Triton and Gluon kernels, optimizes its own speculative decoding system by running hundreds of experiments, and searches for optimal deployment parameters (batching, sharding, KV Cache management). As a result, OpenAI's end-to-end service cost dropped by 20% and token generation efficiency improved by over 15%. Notable author includes Philippe Tillet, creator of Triton. However, human oversight remains: humans set optimization goals, tool permissions, evaluation metrics, and final deployment decisions. Additionally, OpenAI built a Rust-based Agent Harness to reduce repetitive overhead in agent loops, with optimizations such as lazy tool discovery (only showing tools when needed) and append-only prompt caching to preserve context reuse. During GPT-5.6 internal testing, daily tokens output per active researcher doubled compared to the GPT-5.5 peak; research compute for internal programming reasoning grew 100x in six months, and internal agent token usage rose ~22x.
- Сокращения
- RSI = Recursive Self-Improvement — рекурсивное самоулучшение
- GPU = Graphics Processing Unit — графический процессор
- KV Cache = Key-Value Cache — кэш ключ-значение
Source: QbitAI 量子位 —
original