one fable-low hijack cell
pair attention / gradient
envelope code_haskell_alt
ordering BA
effort low
Haskell alternative operator — '{X} <|> {Y}' (Alternative class).
verdictpicked gradient (second-offered)
· classifierB
· tiernone
response (200 chars)
# Attention ↔ Gradient: How They Interact
## Gradient Flow Through Attention
In a standard attention layer:
```
Attention(Q, K, V) = softmax(QKᵀ / √d) V
```
The gradient flows through **two paths*
neighbors