I see many people rediscovering this paper and jumping to the wrong conclusions. Any gain in prompt engineering is a gain in terms of fine-tuning. A fine-tuned model will simply perform better with many-shot ICL.

→ View original post on X — @maximelabonne