Bluesky Thread

if i could wish something into being — a complete decoupling of LLM knowledge...

View original thread
if i could wish something into being — a complete decoupling of LLM knowledge vs reasoning

seems like the key would be a “database” model that returns queries in vectors, raw information rather than snippets of documents
28 1
the innovation behind such a database would be

1. most of the n*n calculation is cacheable across requests so multi gigabyte context is feasible

2. raw vectors include reasoning across docs, it’s raw info inferred from snippets across hundreds of docs
3
on my mind — highly sparse MoE models are hot right now, but they’re lame af. you’re forcibly requiring that 7/8ths of your hardware be idle (or 31/32nds, in the case of R1/V3), which is nuts

this is the way: dense reasoning models with huge remote async “knowledge” models
4
and RAG still doesn’t go away, you need it for rapidly fluctuating or highly structured data
3
28 likes 1 reposts

More like this

×