2C. Since we haven't tested all possible abilities, we don't know the full range of abilities that have emerged in large language models.
2D. We're likely to see more emergent phenomena as we continue to scale up models (and implicit argument for more scaling).

→ View original post on X — @_jasonwei,