GPT-5.6 Builds New Kernel, Achieving 9.7x Speedup on TPU

HuaxiuYaoML · x · 2026-08-27

In a TPU v6e Masked GQA kernel optimization task, GPT-5.6-sol demonstrated recursive self-improvement capabilities. It generated 145 versions: the first 23 tuned the official kernel to 7.775x speedup, while the agent then switched strategies to build a new Pallas kernel.

The final artifact achieved 9.758x speedup on the visible set and a hidden-set score of 0.708, beating the reference. This shows the agent's ability to recognize the ceiling of one approach and search beyond it.

Original post →

More from coding & agent

coding & agent channel →