1 comment

[ 3.9 ms ] story [ 12.4 ms ] thread
Most AI models blend meaning — but this one selects.

GD-Attention is a provably nonlinear attention mechanism with: ・ No softmax ・ No averaging ・ A unique semantic jump point $s^*$

Verified independently by Gemini, GitHub Copilot, and GPT-4. → Softmax isn't just suboptimal — it's structurally incapable of what this model does.