The universe is no narrow thing and the order within it is not constrained by any latitude in its conception to repeat what exists in one part in any other part. Even in this world more things exist without our knowledge than with it and the order in creation which you see is that which you have put there, like a string in a maze, so that you shall not lose your way. For existence has its own order and that no man's mind can compass, that mind itself being but a fact among others. 1
It is often asked: how much better can LLMs get? This piece started as an answer to that question, although it has since ended up concerning a more specialised aspect. My sense of language models is that—apart from any broader limitations in terms of their exclusively modelling language—their limits will ultimately be those of language itself. Here I want to explain what I mean by this, in which I will consider specifically the difference between solidifying and extending performance. 2 The former is more important for integration into existing society, the latter for more radical visions.
Supposing that unsupervised learning works by acquiring primitives, the key is that they acquire reliable access to a flexible set of these. This allows for their sampling more or less creatively from the space of all possible sentences. There are two desires in terms of improving performance here, either to ensure their reliability for tasks which are already within their ambit—thus solidifying their performance—or to extend the tasks of which they are capable, an aim which may take several forms.