Catherine Olsson and Nelson Elhage: Anthropic, Understanding Transformers
The Gradient: Perspectives on AI
The Basics of Attention Only Models
In this paper, we look at attention only models. Attention only models are mostly only exist like in literature as objects of study. They're not very practically useful, but they contain some of the core complexity of the core value of a transformer while being easier to reason about than MLPs. We talk about query key or value composition depending on where the intermediate value from the first layer feeds into the second layer. And so I think having that framework in mind allows you to ask CRISPR questions about how exactly combinations of different layers are handing off one information to another.
00:00
Transcript
Play full episode
Remember Everything You Learn from Podcasts
Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.
