Exciting progress from the Superalignment team:
“When we supervise GPT-4 with a GPT-2-level model using this method on NLP tasks, the resulting model typically performs somewhere between GPT-3 and GPT-3.5.”

→ View original post on X — @romainhuet,