Comment by joshjob42 on AI #5: Level One Bard · 2023-03-31T08:06:19.619Z · LW · GW

Crossposting from your substack: 
Re: “Llama is this so easy”, this repo https://github.com/ZrrSkywalker/LLaMA-Adapter seems to suggest that actually you may need *very* few additional parameters and a tiny amount of fine-tuning of a network that you bolt on top of an LLM to get pretty good instruction following. I haven’t experimented with it though.