My intuition is it's somewhere in between. Somewhat bigger than GPT-4, more training data, and some additional training methodology.

→ View original post on X — @rasbt