Comment by torginus
11 hours ago
My understanding is that current LLMs aren't really well suited to do this - tokens are predetermined, and while embeddings are learned, they are learned from an existing corpus of text, which presumably comes from a human language. After this point the language is locked in. There really isn't a kind of training which could efficiently change its embedding representation. I mean, you could probably instruct an LLM to design a more compact language, generate synthethic data and train a new gen on that, but that would be a fairly explicit process and not something that would emerge during training.
> tokens are predetermined, and while embeddings are learned, they are learned from an existing corpus of text, which presumably comes from a human language
That's not true since are least multimodal models - token space is broader now, encompassing visual and audio signals. Tokens are more like sensory/perception units now, not digitized pieces of writing.
I imagine LLMs exhibit this tendency for compressed communication in post-training/RL phase. Particularly with CoT, until interpretability became baked in as grading criteria.
What you said doesn't contradict me, and doesn't refute my point.
For images and audio, you still need to predetermine an encoding, then pretrain to learn an embedding. This embedding will try to replicate the input distribution - so if you trained it on Google Street View and scanned documents, its representation will be grounded in only those.
I would even claim that this approach is somewhat counterproductive, as images are far more information dense, containing tons of concepts
While LLMs do have some ability to learn to use their embedding space in non-predetermined ways, they still lack the ability to pick an efficient embedding.
So I guess, a nice thing is that interpretability is baked into this approach to some degree, and humanity has proven through its existence, that you can do a lot with just text, but this approach is still predetermined.
I guess this is what LeCun's JEPA is about, that the AI gets to learn the representation on its own as well.