Correct Arity

terminator vs rubber duck

Asking a human vs asking a robot

(maybe this is obvious) Asking a human for help is very different from asking an AI for help.

Often IRL or on discord or probably many other places for asking humans questions, someone will purely by asking their question, find their own answer. i.e. rubber duck debugging — it doesn't matter if it's a literal toy listening, just talking through question is enough to illuminate it. But I do think it helps to have the credible threat of a real human listening, because a human listening kind of forces you to get your affairs in order. You don't want to waste human time or look like an idiot, like it's ok to not have thought of or known something but you want to look at the basics again before roping in another human. I don't feel this as much for an AI, they won't even point out if you ask a poorly formed question. And of course the cost of asking them and the general friction is so low, so there's less need to form an intelligible question. So there's much less of the process of thinking through it that helps you answer your own question; it's much more the AI answering it for you. Because AI is so general, I don't think you can just say "I'm moving up a level of abstraction"; it has to be "I'm ok with the AI thinking for me".

The other big thing is AI seems much less inclined to reject the premise of the question, especially that what it's asking for is uneeded or too complicated. For example I spent a very long time with GPT-6.1-Sol high trying to generate code to generate decoders for FHIR slices by first checking the element's discriminator, which says which of multiple potential structures it has, and then trying to parse it into that structure. The problem with generating code for this is there are a lot of ways to do discriminators, including patterns on values which can be something like one out of an array of a subelement has a code field and system field belonging to valueset. The point is you probably could in any given resource, check the discriminator somehow and then parse into the appropriate slice, but that "somehow" does a lot of heavy lifting, it's not a simple case statement. And I was really banging on the AI to do this for a long time and it would just add thousands of lines of code to the generator; never complain. By contrast, someone on the Gleam discord immediately suggested the standard decode.one_of, which technically does not return an error when the discriminator matches a slice but the rest of the structure does not parse into the discriminator's slice and instead falls through to open slice, but like, who cares? And for that matter, who even needs a Gleam FHIR code generator that creates decoders from arbitrary profiles? Imo current AIs are ok-ish at code itself, maybe sometimes they won't use the best approach to lists or whatever, but what really kills them in terms of piling on unjustified complexity is they never ask to zoom out and ask who needs this, what is it all for. You can of course tell the AI "talk like a caveman" "STE100" "keep it simple" "reduce lines of code" etc., but I at least have not seen them narrow down/decide what do we actually need to do.

Socially, asking a human is kind of scary and an imposition on their time, but also potentially friendly and community forming. So I don't want to undersell AI, it continues to improve, often it does give a correct answer and sometimes even to the right question. But it does kind of lead to vibecoding alone. Not making any claims about why this is, I just feel like empirically AI software tends to be one guy's thing rather than community driven. Not that that's necessarily bad, maybe it helps realize a specific vision. It could be better to work with other people when they bring a valuable different perspective or experience though. And in a "maybe the real open source was the friends we made along the way" sense.