I've seen models learn heuristics that are harmful in real performance, and I wonder how much is accuracy directly transferrable to actually good drafting.
A question, when GPT-4 contradicts in explanation, how much of them were in fact correct?
A question, when GPT-4 contradicts in explanation, how much of them were in fact correct?
It was mostly when a card is good in a vacuum but not as good in a specific set. WOE (which this was trained on) skewed pretty aggressive, so GPT-4 was tended to overvalue strong expensive cards (compared to what good players thought at least).