Kimi K3 kernel researcher nathancgy4 leaves Moonshot AI

SonglinYang4 · x · 2026-10-15

Nathan Gao (@nathancgy4) has left Moonshot AI, calling his year at Kimi the best of his life. His work included optimizing KDA (Kimi Delta Attention) kernels, co-developing Attention Residuals, deep dives into optimizers and training dynamics, and building test cases for the Kimi K3 release. The retweet chain also cites a blog post exploring a future-dependent rounding effect in KDA, plus TorchTitan/Attention Gym work on training primitives for hybrid GDN/KDA attention architectures.

Related event: Kimi Core Researcher Nathan Departs Moonshot AI(2 posts)→

Original post →

More from Companies & People

Companies & People channel →