It's a little frustrating when the author calls out "technofetishistic" faith in large language models[1] and then goes on to misunderstand their capabilities i

It's a little frustrating when the author calls out "technofetishistic" faith in large language models[1] and then goes on to misunderstand their capabilities in the opposite direction. I'm no "true believer", they're just useful tools for modeling text generation, but to underestimate them is a shame. > Ideally, the essay writer has first observed their subject, then drawn some sort of analytical conclusion about that subject, then organized their evidence in a way that expresses the logical connections between various pieces of evidence, before finally communicating that to a reader in a way that is clear and persuasive. Of the elements discussed about "writing an essay" above, none are clearly impossible for a large language model to do at the present, despite the author's insistence that some of them are impossible. If the subject is something in writing (and you can pass it in the prompt), or is written about in the training set, then the model has indeed "observed" it, and to get it to draw analytical conclusions all you have to do is ask nicely and correctly, maybe using those words. Again, with organizing evidence, you merely must ask for it by name, rather than hoping that the model understands what you mean when you say "write an essay" - ask for "organize the evidence in a way to support/contradict the conclusion X", and finally of course communicating cogently to the reader is something the writer correctly understands is possible even with a trivial prompt. Sure, it moves the challenge from "write an essay" to "design a prompt to get the language model to write an essay", but that's "not writing an essay", or is it? It is immensely frustrating that people feel like a large language model is "cheating", but also don't consider using word processing software with grammar and spelling correction "cheating", even though they use fundamentally the same processes behind the scene and differ only in the user interface, really. [1] I don't like the term AI, I prefer large language model because it is unambiguous about the properties of the system - AI implies thought processes, which is a distraction at best from the useful properties of a complex statistical model.