Sure, emergent properties can arise as parameters increase. Everyone knows that. That’s a much less specific claim than to say that the benefit of modifying sof

Sure, emergent properties can arise as parameters increase. Everyone knows that. That’s a much less specific claim than to say that the benefit of modifying softmax can only arise as an emergent property after N parameters, and therefore the benefit can only be evaluated on models above a certain size. To my understanding the author of TFA isn’t suggesting the same issue as the one in your linked paper.