Damus
Open in Damus
utxo the webmaster 🧑‍💻 · 5d
I think orchestrating small models that are highly specialized and can be quickly loaded makes more sense then giant models Like I don't need astrology knowledge in my coding llm but I might need it...
 Pyrohawk profile picture
Pyrohawk @Pyrohawk
I think the goal was to make a model that could keep learning and improving. It's not panning out that way, so with energy constraints we are learning small models make more sense now.
I do think there might be a place for large model to discover connections we might not have seen/noticed.
1
Matt 🛸 · 5d
I'm imagining something like modules where model domain can be expanded or reduced as needed. So you basically have a very basic component, call it module 0 (some basic knowledge or ability), that can be "taught" different domains.