> Why would anyone expect it could just be repurposed for a domain specific task, without additional fine tuning? That was the unexpected result of training GP

> Why would anyone expect it could just be repurposed for a domain specific task, without additional fine tuning? That was the unexpected result of training GPT-3 (zero-shot learning). Finetuning in theory would give better results, though.