I tried to get ChatGPT 4 to write an application to brute-force a PDF file. 5 times in a row it provided code with functions that didn’t even exist in the library. It also provided versions that doesn’t exists at all for some libs. Feels like it has become worse over time.

Reply to this note

Please Login to reply.

Discussion

You’re supposed to write those functions. It provides a skeleton for the core of the program.

I think copilot has gotten worse too. I thought it was just me but it used to suggest what I wanted a lot of the time and now it’s giving me a bunch of nonsense.

That sucks.. I haven’t used Copilot yet.

i’ve never wanted to sign up to anything, so i’ve just been using phind(dot)com

essentially it acts as a better, more tailored search engine for me, pretty much assisted in helping me setup my rpi4 bitcoin core node “the hard way” (i.e. completely over SSH using Ubuntu Core)

i’ve been getting it to explain to me what all the different parameters in the bitcoin.conf that people tell you to set in their guides, are actually doing

but when something doesn’t make clear / direct sense i’m almost always googling it to confirm (i.e. doing a bit of my own research to double check)

for me, it’s just a waaaaay better search engine essentially

Copilot really shines when your are using a language you aren’t that familiar with or for boilerplate code which is really common in enterprise settings.

it's sandbagging

Yep. Same experience with rust. It pulls libraries out of its ass and functions that don’t exist or often dead code as it’s a few major versions behind.

The fog of certainty is real - the responses are so confident - and then you have to reply with… you make this dumb mistake, that doesn’t exist, that won’t compile, etc. Can you rewriting using functional programming instead of if else.

LLMs are generally speaking terrible at code gen for anything important.