[dongobread]

This isn't apples to apples - they're taking the optimal prompting technique for their own model, then using that technique for both models. They should be comparing it against the optimal prompting technique for GPT-4.