KDA Blog

Notes from the
kernel design loop

Results, failure modes, and lessons from building agents that write, verify, and tune GPU kernels.

Subscribe via RSS Request a kernel

LATEST POST

· 12 min Read

KDA²: Kernel Design Agents (KDA) optimize Kimi Delta Attention (KDA)

Our agents wrote Kimi Delta Attention kernels that run up to 2.96× faster than FlashKDA on B300 with a tenth of its state error. Here is how, and how the agents tried to cheat along the way.

  • Results
  • Kimi Delta Attention
  • Reward hacking
Read the post

ALL POSTS

1 post

2026

  1. KDA²: Kernel Design Agents (KDA) optimize Kimi Delta Attention (KDA)

    Our agents wrote Kimi Delta Attention kernels that run up to 2.96× faster than FlashKDA on B300 with a tenth of its state error. Here is how, and how the agents tried to cheat along the way.

    Results · Kimi Delta Attention · Reward hacking