Even then, the presence of a next token is just as informative as the absence of another during training. That information also gets backpropagated. And during

Even then, the presence of a next token is just as informative as the absence of another during training. That information also gets backpropagated. And during inference, good luck identifying which weights were responsible for a given next token (assuming we’re using greedy decoding, don’t even get started on beam search) let alone which samples contributed to that weight (hint: they all did).