Post

It does seem like latent space is the most efficient place to do deliberation - like deep consideration before putting your thoughts into words.

it does. but every time i start thinking in that direction, i can’t convince myself i’m not engaging in magic thinking. like maybe i’m just hoping the algorithms will save me without really understanding the problem. and i’m not sure **how** to understand latent space at scale

I got a little hazy on the internals. So the coprocessor is fed the embeddings directly or is it doing something else. The “thinking model” is frozen right.

yeah, that part was hard to digest. the coprocessor generates embeddings that are fed into the kv cache (attention blocks) of the main model. yes, the main model is frozen. well, both are, but the expert is trained after the main model

What I understand is the the expert inserts embeddings in the kv cache of the main model. The final answer is taken from the main model and losses back propagated through the coprocessor. This way the coprocessor knows what to insert inside the main model.

Thanks for sharing your blog post! But I also would so be up for a long thread form of this XD When you mentioned working in the latent space, I instantly thought of this late 2023 paper by Google Research arxiv.org/abs/2310.04475

interesting, i’ve heard of CAVs, but they always seemed more of a gray box approach. it seems like you could train a model to translate an embedding back into text. doesn’t seem like it would be too hard. also seems inevitable if we go all in on latent space

I believe the latent space almost always means that there's potential for more accessible information than just text with the space being continuously differentiable. I think of early days variational autoencoders that could be used to generate beyond training data

The big notes here are: * Use an "expert" LLM to augment a generalist * Yes, it's like RAG, but it doesn't replace RAG

Cool! The idea of generalist + expert is very intuitive, curious how it works in practice

im also curious if it could be combined with the byte level encoders, have an expert in understanding video, another for x-rays & MRIs, another for poetry, etc.

It doesn't really look like RAG at all, is it? It looks much more like a different way to handle internal thoughts in embedding space rather than forcing them to pick concrete tokens to give it space to calculate...

a lot of people see RAG as a cheaper way to get a domain specific version of a model. atm RAG is actually a reasonable approach, but this might be better