Interesting! But knowledge is still best instilled via pre-training. "Relatively" cheap still remains very expensive and cumbersome, even for 7B models, compared to just adding some more embeddings to a database.

→ View original post on X — @rasbt,