A very good point from the KiwiPyCon 2022 talk by @chrisjrn via @glyph's The Futzing Fraction: if you produce code by writing natural language telling an LLM what to do, you are thinking in much different terms than if you used a high-level formal language. And it feels intuitively correct that this engages a different mode of thinking.

My research interest is exactly how we can learn things about mathematics by being made to express them much more precisely in a formal language. So using an LLM, even (assuming for the sake of discussion) if it were mostly useful for writing code in general, is quite orthogonal to what i am interested in.
"Saturday Keynote: Convention and Construction" - Christopher Neugebauer (Kiwi Pycon XI)

YouTube
@anya ok, how did you find my talk from @glyph's piece? (Just interested)
@chrisjrn @glyph Your thread where you discuss the topic and link to the keynote is linked in footnote 11! :D
@anya @glyph oh that is a deeeeeeeeeep cut. I'd missed that Glyph had included my toots in a footnote; and then the whole damn thread :)
@anya @glyph I am giving a talk that's explicitly about LLMs and thinking about problems at #NBPy next month; but it's nice to know that I've been thinking about this stuff for longer than I can remember.