I think we need to increase the data and its quality, and as in LMs do fine tuning (aka grounding). The model already has compression bottlenecks, which we could improve sure, but the concepts are already there.

→ View original post on X — @nandodf,