François Fleuret invents a compact notation for attention tensor ops
francoisfleuret · x · 2026-10-31
François Fleuret invented a compact one-line notation capturing the full attention-plus-FFN computation of a transformer block:
((Wq | Wk) < QK | Wv) < AV < Wo < (W1 | W2) < W3
Here | denotes parallel weight concatenation and < composition with the following operator (QK scoring, AV weighting), squeezing the whole block into a single pipeline-style expression.
More from Fun
- 'In SF you gotta show ARR even if you sell cakes': AI culture mocked — tekbog · 2026-11-01
- Robotic hand cuts open candy bag and drops treats one by one in Halloween demo — imankitgoyal · 2026-11-01
- AI-generated imagery is already everywhere in the physical world, and it's hard to unsee — signulll · 2026-11-01
- Meme: LessWrong isn't a joke name, and this post will make you update your priors — austinc3301 · 2026-11-01
- World Labs goes Halloween: 'an epic battle was inevitable' — theworldlabs · 2026-11-01
- Trick or Tear: an interactive 3D Halloween ticket where tearing straight wins treats — techartist_ · 2026-11-01