Definition
Plain language
The step that turns a model's raw scores into probabilities that add up to one, so it can weigh or pick its next word.
As stated in the literature
A function that exponentiates and normalizes a vector of logits into a probability distribution; the normalization step inside attention and in token sampling, and a component whose saturation can throttle gradients at the output head.
Why it matters: It is the step that turns raw scores into usable probabilities inside attention and word selection, and when it saturates it can choke off learning.
For example, it takes a model's raw scores for possible next words and converts them into percentages that add up to a hundred so one can be chosen.
Heard on the show
“If you take softmax attention and train it to convergence on a retrieval-style objective, what it converges to is, mathematically, a classical statistical estimator.”Episode 033 — Echo: The Paper Arguing You Never Needed a KV Cache for Retrieval