Josh Simmonsone of the major recurring themes i pick up reading people talk about their usage of llms for <insert whatever> is that it allows people to skip past the easy but very difficult part where they have to make a bunch of small decisions while typing over an extended period of time. essentially you're removing this sustained effort of constant micro-decisions with bursts of effort followed by waiting and review. it's a work around for "writers block" or i suppose, "programmers block".
Per Vognsen@dotstdy I've used it in a few cases for random throw-away explorations that I otherwise would have not have bothered doing because of the initial setup and effort required. And at least for me, I get so annoyed at the bad code quality, even after corrections, that it provokes me into throwing it all away and rewriting it from scratch. So it can work as a weird sort of "get off your ass and do it yourself" provocation, but I suspect that's not what most people get out of it.
@dotstdy Like, I don't understand the append-only GitHub repo spew that I see so much of. My experience is that if you have any opinions about how code should be written, it's harder to get Claude to do the right thing than it is to rewrite the whole thing yourself. I think that's why the hardcore adopters quickly stop trying to do anything about that and just let it rip.
@dotstdy Also, the latency involved in interactive prompting and some of the self-driving bug fixing becomes a nightmare, and that's another thing that those hardcore adopters seem to leave behind quickly, but that just means you have even less control. So all the positive and negative affordances of these tools push you towards not even looking at the code anymore.
@pervognsen yeah there seems to be a pretty broad divide between i guess, peripheral users, and the core prompters. most of what i see where people try to integrate it into existing codebases with standards and whatnot devolves into a lot of prompting and hand editing (which i suspect people downplay when talking about it as well). then with people operating outside their usual expertise or cares they see value in the other mode, where you try to get as hands off as possible.
Matt Sutherlin@dotstdy @pervognsen that last bit is probably true but also so weird to me. Like, I can see the appeal of using the llm to fast track an idea outside of my wheelhouse where I’m never going to invest the time to learn it from scratch but I do not understand not wanting to understand the thing at all. I want to learn enough to not have to use the llm in the future!
Lesley Lai@pervognsen @dotstdy This matches my experience using those things as well. The fact that I can produce something quickly, but producing high-quality work is more exhausting than doing it manually, combined with the fact that people forcing me to use those tools clearly don't care about high-quality work, has pushed me gradually toward more YOLO mode.
No wonder how many ex-responsible people seem to be producing slops after they adopted LLMs heavily.
Paul Khuong@pervognsen @dotstdy I keep the scope of each change small, and remain ready to make some changes directly: the bot doesn't learn, no point trying to make it fix its mistakes.