[danielmarkbruce]

A large model well trained on good data will have logits so negative for something like "1+1=" -> 3 that they won't come up in practice unless you sample in a way to deliberately misuse the model.