Databricks AI agents take 2nd in Meta kernel challenge with 2.5X speedup over SGLang
Yuchenj_UW · x · 2026-11-14
Databricks placed 2nd out of 300K+ submissions in Meta's Muse Spark Open Kernel Challenge, with kernels averaging a 2.5X speedup over SGLang.
The team built a "Kernel Factory" where AI agents autonomously hill-climb toward better GPU kernels, finding Opus 5.5 the strongest model for kernel optimization. The kernels go straight into production, powering Databricks Unity Gateway and customer AI inference.
More from coding & agent
- Matthew Berman open-sources Scarlet, an AI video editing agent that speeds up his workflow — MatthewBerman · 2026-11-14
- 3DSceneBench: coding agents rebuild livable 3D rooms but top functionality score is just 42.1 — yizhu59 · 2026-11-14
- The 5 foundational AI concepts that matter beyond learning tools — Zachly · 2026-11-14
- Demo: Claude finds 20 verified marketing leads into Google Sheets for $0.68 via Composio — JaynitMakwana · 2026-11-14
- Boski launches a shopping agent that hunts the best price and buys for you — Scobleizer · 2026-11-14
- mireye launches a real-estate research agent covering 160M+ US parcels via API and MCP — ycombinator · 2026-11-14