AI/ML · Tasks

Language

Teaching machines to read and write: from word vectors to large language models.

Embeddings Large Language Models

Language and vision are fundamentally different

Language

a mind one word after another →
Generated by the human brainWe invented it: symbols that mean something only because people agree they do.
One-dimensionalA single line of words, sounds or characters.
SequentialOrder carries meaning: "dog bites man" is not "man bites dog".

Vision

the real world a whole image at once
Not generated: a phenomenon of the real worldLight bounces off things into an eye or a camera. Nobody wrote it.
Two-dimensionalA flat grid of pixels: a 2-D picture of a 3-D world.
Spatial, all at onceEvery pixel arrives together; what matters is what sits next to what.

So language models learn what people produced; vision models must learn how the world looks.