AI article
One Update, One Quarter-Turn: The Attention Layer That Learned to Rotate
Community description: In October 2025, Moonshot AI made a startling claim: its Kimi Linear architecture — built on a new...
Dev.to | Oct 6, 2026 | Daniel Sam Pete Thiyagu
Automated excerpt
Nonnegative per-channel gate (KDA): breaks the symmetry, but the eigenvalues stay real. Signed per-channel gate + β ∈ 0, 2: the spectrum goes complex. A scalar gate can never rotate; a signed per-channel gate can.
Selected automatically from source text; not independently written or fact-checked. Read the original for full context.