Forget model names, look at the ops: the classic math behind attention, RoPE and diffusion

techNmak · x · 2026-09-19

A refresher thread arguing the best way to study AI is to ignore model names and look at the operations: attention is just dot products, scaling, masking, softmax and weighted sums; RoPE is geometry (rotated coordinates whose dot products depend on relative position); convolutions and PCA rest on equally old foundations; training is gradient descent, and score-based diffusion folds log-probability gradients into its dynamics. Topics usually split across linear algebra, optimization, probability and geometry keep reappearing in modern AI.

Original post →

More from Research

Research channel →