Decoder
Domain: LLM Concepts
🎯 Motivation
The decoder generates output tokens one at a time, using both the encoded input and the tokens it has already generated.
📋 Concrete Example
Generating Bonjour: decoder starts, uses encoder’s understanding of Hello, produces Bon. Then uses encoder output + Bon to produce jour. Then stops.
🔗 Analogy
A writer drafting a story - they write one word at a time, each word depends on what they’ve already written (previous tokens) and their overall plan (encoder output).