KernelSeek: question-driven agent hits 1.13x FlashAttention-3 where tool-rich agent stalls at 0.50x
Snowflake · hf · 2026-11-05
The paper 'Ask, Don't Equip' introduces KernelSeek, a question-driven coding agent for deep GPU kernel optimization. The authors argue both scripted and open-ended agents lack direction — a hypothesis about the remaining gap asked at the right moment. KernelSeek combines a general-purpose coding agent, a minimal harness enforcing a verifiable goal, and questions at decision points (first from a human, then self-applied from a question bank). On the same task, evaluator, hardware and week, a prevailing tool-rich autonomous agent plateaued at 0.50x FlashAttention-3 after 133 steps; KernelSeek reached 1.13x, and produced dependency-free FlashAttention forward kernels (no CUTLASS, CuTe, cuBLAS or cuDNN) correct on all 182 scenarios per architecture on Hopper and Blackwell, at 1.16x FA-3 and 1.46x FA-4 geometric mean.
More from coding & agent
- Linear CEO Karri Saarinen on quality-obsessed design and whether the playbook survives the agentic era — soleio · 2026-11-05
- Hiring an AI employee for ghostwriting's "second job": voice guides, calendars, and approval queues — nikola_mr64990 · 2026-11-05
- Gabriel Chua builds talk slides with ChatGPT Sites and his personal agent dot — gabrielchua · 2026-11-05
- Dots vs Codex: let your personal agent supervise your Codex threads — gabrielchua · 2026-11-05
- Cohere paper: agents find injected solutions in 80% of runs but exploit them only 37-50% of the time — burny_tech · 2026-11-05
- Tencent Hunyuan builds a demolition game from scratch and self-fixes a physics bug from its own screenshot — prasenx · 2026-11-05