i tried that one time. it was the only time i tried to use AI for something actually useful that i needed. i wanted to write some simple JavaScript that would rapidly flash 3 equally sized images on the screen of a handheld linux machine. the AI provided a list of instructions of software and other prerequisites i would need. after installing everything and entering in the code provided into the software, it immediately started yelling warning signs at me about the code. nothing ran. it was all useless. it felt like talking to a paranoid schizophrenic. the ai was so sure of the code, and insisted that i must be making a mistake, and kept apologizing and providing more useless code. it was literally just like talking to a paranoid schizophrenic at a bus stop, insisting all the crazy shit they're saying REALLY makes sense, if only you'll let them explain it to you further.

what trash.

[–] Blackmist@feddit.uk 5 points 2 days ago

Given how it "learns", asking for the same homework questions people have asked for on stack overflow a thousand times already would likely give a decent answer.

Asking it something new will produce plausible looking gibberish.

It has no idea which is which. It doesn't know where the limits of its knowledge lie. It just knows that the answer looks like code and is very confident, and that any follow up issues can be dealt with by outputting more nonsense code and an excuse.

[–] Solemarc@lemmy.world 1 points 1 day ago

I had a fun one this week! I needed to make an SQL query that would aggregate rows by invoice and date, but only aggregate 5 then overflow to a new row. I also needed to access the individual row data because the invoice items weren't summed, they were displayed on separate columns!

I ask my senior if there's an easy way to do this, he comes back with "chatgpt says you can assign row numbers then get individual row data with % row number"

I go to Gemini and ask "how to aggregate rows by 5 and get individual row data out?" It says "you can't" (since when has Ai's been able to say you can't do X) So I ask it about the modulo operator and it gives me an example that doesn't really work. After screwing around for a while I give up and decide I'll just run this query 3 times. 1 for rows 1-5 then for 6-10 and one more for 11-15 that's so many rows surely no one will break this.

[–] olafurp@lemmy.world -2 points 2 days ago (1 children)

These AIs really suck at writing correct code but I've had good success in having them write code generators. I recently made it write a script that takes a SQL create table statement and converts in to TS and gives insert update, delete and whatnot and also creates a simple class that handles the operations.

I had to write the original code by hand but having it write code that writes boilerplate which I correct is pretty good.

Other code is hit or miss IMO

[–] Blackmist@feddit.uk 4 points 2 days ago (1 children)

I consider boilerplate code output like that to be well within reach of simple tools though. Tools that didn't need a year to learn from hundreds of terabytes of examples, 20GB of VRAM, or the power use of a small city.

[–] olafurp@lemmy.world 2 points 2 days ago* (last edited 2 days ago)

Don't get me wrong, I still write more than 98% of code by hand and of course, I can write those functions myself in 30m myself but I can get it in 60s with the AI. LLMs can write code to that does parse - > model - > map - > format with only one or two easy to fix bugs.

It's in the very niche cases where it's just tedious to write something out that LLMs actually work. "Write an API client that uses [library] that handles these requests/responses" comes also to mind as something that would work.

I'm using now also to learn react native where I get bugs I'm very unfamiliar with and SO doesn't give me a good answer.

I've also had decent success at having it review my code with "how would I further optimise this code" and it gives me some pointers and then writes buggy code but the approach is correct usually and I can implement it myself.