Glossary · Term

latent space

← all terms

Definition

Plain language

The internal high-dimensional space where a model's representations live, before they get turned into output.

As stated in the literature

The continuous representational space inside a neural network where intermediate computations occur, distinct from observable inputs and outputs.

Also called: latent compute

Why it matters: Almost everything interpretability and probing tries to study lives in latent space, so the concept anchors how we describe what models 'know.'

For example, two sentences with similar meaning end up near each other as points in the model's latent space, even if their words are entirely different.

Heard on the show

“… You compress images into a continuous latent space with a variational autoencoder, you run a plain transformer over those latents — that's the Diffusion …”
Episode 127 — What Diffusion Language Models Were Missing: A Map, Not an Algorithm

Mentioned in 4 episodes

  1. 127
    What Diffusion Language Models Were Missing: A Map, Not an Algorithm
  2. 115
    Teaching a Phone Agent to Reason Silently, And Keeping It Honest
  3. 041
    When the Iteration Teaches the Model to Skip the Iteration
  4. 032
    A Sticky-Note for Every Layer: Letting Transformers Remember What They Were Just Thinking

Related concepts

Related terms