[t0mpr1c3]

I agree with the author's overall conclusion about pricing LLMs, although I question some of the reasoning. He does not talk enough about one important input: the data used to train the models. To a first approximation, their utility is a function of the quantity of data available. Currently, the bulk of the training data is public (the internet) and that is a major reason why the performance of different models is not diverging more. One huge future area for AI is modeling the physical world and in that arena people are going to be creating their own empirical datasets. Waymo has achieved FSD and Tesla has not. It is not coincidental that Waymo has built a sophisticated platform to simulate the physical world.