Skip to main content

The Transformer

This video presents the same text shown beside it, spoken and on screen. It adds nothing the text does not say.

State

A transformer stacks attention layers with ordinary weighted-sum-and-bend layers, mixes in position information so order matters, and trains end-to-end by descent with backward blame.

Show

Attention reads all positions at once — a shape fitting parallel hardware, and why the 2017 design won the scale era.

Watch for

Trainability at scale won as much as capability — three dials, and a machine built to spin them.