Post #560019

@shriramk@mastodon.social The ideological problem I have with LLMs is that they're not pitched as coding aides per se in that they don't just make coding faster, but also as being able to span the problem solving space as well. I see this as being a required property of something that can translate natural to machine language like LLMs can due to the need to fill the precision gap - but it's also one that makes it profoundly uncomfortable to argue for one's validity or differential utility as a problem solver. Furthermore, I find that it's very easy for the LLM to silently fill in gaps in the specification that I didn't know existed and do so in ways that I didn't intend, leading to accumulation of cognitive debt. I find that hand-coding in codebases that I was previously writing with LLMs is very akin to using a library after extensively reading the documentation but have never personally used. In my experience, I rapidly lose the ability to critique the LLM's decisions (both as presented directly to me and as implicit in the code) in meaningful non-superficial ways as the amount of "fill in" from the model increases.