Early attempts at instruction tuning relied entirely on demonstrations from humans. This made the model easier to prompt, but the approach was limited by the inherent difficulty of manufacturing new humans.

→ View original post on X — @goodside