note
Latent spaces as maps
A model is not a brain. It is a map of relationships, and you only see the territory by walking it.
- latent-space
- interpretability
- metaphor
A model is not a brain. It is a map — and like any map, it tells you about the territory only because somebody walked it first.
Every direction in the latent space corresponds to a way of moving between ideas. Some of those directions are interpretable: gender, sentiment, formality. Most are not. The ones we name are the cities. The unnamed directions are the roads between cities, and there are many more roads than cities.
When I write notes about The shape of a thought I am almost always asking the same question: which directions in this space matter to me, and why?
related
- The shape of a thought
- the-geometry-of-intelligence (essay)