chapotraphouse

Having a profession centered around writing and being in favor of the gibberish machine really is telling on yourself lol

"what if we had a guy who was like 'im the ideas guy' but it was a shitty algorithm"

And even if you use it as an idea generator, the ideas you get are mediocre boring ideas. You never get something unique and interesting.

I remember trying it for some very basic TTRPG campaign prep and it gave me the most hokey by-the-numbers boring derivative shit I've ever seen. You have to keep in mind TTRPG writing is a genre of writing that mostly consists of ripping off books and movies you liked, anyway. And it couldn't even reach that very low bar. Genuinely concerned idea scroungers never heard of hitting the "random" button on TVTropes.

I imagine using it as an "idea generator" is more like a programmer's rubber duck. It's not actually giving you anything you'd use directly, it's a way to think out loud and maybe jog your memory, or potentially connect ideas you hadn't thought to connect.

It's trash as an idea generator.

The only useful thing I've gotten out of a (text) AI is asking it to guess functions of keyword mechanics in games. Like I was designing personality traits for AI leaders in a strategy game, and had a dozen bad candidates for "over produces defenses." So I told ChatGPT to try to guess the meanings of bunkerist, hoxhaist, prepper, turtle, protectionist, survivalist, isolationist, guardian. Which did narrow it down to bunkerist, turtle, and protectionist (note that this is literally wrong in the case of protectionist). Normally I'd try to poll a bunch of random people for this sort of thing, and try to avoid anyone who's trying to be clever. So it did save some work there.

It won't come up with anything useful going the other way around though ("list some possible names for traits of AI leaders in a strategy game"). Like I said, it doesn't work as an idea generator.

I guess in general it's probably useful if you're in a situation where you need to make sure your writing is very very clear. If ChatGPT can correctly summarize what you wrote, it's probably safe for people who are distracted or bad at reading or whatever.

I've looked into using it to save time, the way this interviewer recommends, and I must say "I would rather put a gun in my mouth" is a pretty accurate response to how it makes you feel as a writer. The only thing it is good for is when I'm feeling incompetent, I plug my ideas into an AI and the garbage it generates makes me feel way better about my own writing skills.

I like this place, I see it on /all/ often with some good stuff (I’ve never listened to the podcast… yet?) but this reads like a foreign language to me, or maybe I’m having a stroke?!

It is very useful for coding because that is one of the few places where unoriginal repetitive solutions are often desirable. But even with coding you have to know what to tell the LLM to do and you have to be able to read and understand the output to make sure it works as intended.

LLM's are a useful too for programmers to automate repetitive tasks but it is nowhere near bearing able to produce usable applications by itself. I am not worried that I'll be replaced by a robot anytime soon.

Those who should be worried about their jobs are people in places like customer support or government services directed at people who doesn't matter to the ruling class. In these cases the powers that be have little holding them from replacing human interactions with significantly worse interactions with a LLM. Nobody important gives a shit if some schmuck can't cancel their cable subscription or gets their employment benefits cut because the computer had a hiccup.

IMO no, for two reasons:

• reading code is harder than writing it. If the AI writes you a standard implementation, you still have to read it to make sure it's correct. So that's more work than just doing it yourself
• AI will produce code that looks right. Since it can't understand anything that's all it does, next most likely token == most correct-looking solution. But when the obvious solution is not the right one, you now have deceptively incorrect code, specifically and solely designed to look correct.

I've never used Copilot myself but pair programmed with someone who used it, and it seemed like he spent more time messing with the output than it would have taken to write it himself.

More like helping programmers write e-mails to needy project managers who need a status update on that feature ticket every 11 hemiseconds

I use JetBrains "local LLM" thingy and it's good at suggesting the very obvious, trivial code that I would write anyway, so it just saves me keystrokes

It's clearly become a crutch for some programmers. I remember talking to someone who does ai research and openly admitted that most of the people in their lab couldn't code and that the outputs from chatgpt where sufficient to do their work.