Comment by supermdguy

The direct prompt comparison isn't quite fair due to the instruction tuning on GPT-3.5 and 4. It'd be interesting to see examples with prompts that would work better for the raw language models.