Seems Meta have been doing some research lately, to replace the current tokenizers with new/different representations:
- Large Concept Models: Language Modeling in a Sentence Representation Space [Github] (December 11, 2024)
- Byte Latent Transformer: Patches Scale Better Than Tokens [Github] (December 12, 2024)
Does this use the same attention architecture as traditional tokenisation? As far as I understood it each token has a bunch of meaning associated with it encoded in a vector.
Uh, I’m not sure. I didn’t have the time yet to read those papers. I suppose the Byte Latent Transformer does. It’s still some kind of a transformer architecture. With the Large Concept Models, I’m not so sure. They’re encoding whole sentences. And the researchers explore like 3 different (diffusion) architectures. The paper calls itself a “proof of feasibility”, so it’s more basic research about that approach, not one single/specific model architecture.