Comment by oergiR

I agree except for (6). A language model assigns probabilities to sequences. The model needs normalised distributions, eg using a softmax, so that’s the right way of thinking about it.