ngl code-llama has been better

even claude-2 sometimes

gpt-4 is not as good as it used to be

janitor for /zk and some more stuff here linktr.ee/randomishwalk

when you say it is not as good as it used to be, what do you mean specifically? for example the code doesn't compile now, the logic behind the code is wrong, or something else?

IMO, sadly, ChatGPT is getting worse at coding across all dimensions. Code is wrong, code is incomplete, answers have gotten lazier (code omitted), etc

Blockworks Research. @_RyanRConnor on twitter/X. 

is there any number of lines of code that you reasonably trust GPT to generate still? i.e. before code is omitted or the answer/logic goes off the rails

I used to trust the system to write complex logic, now the more boilerplate the better

idk if it’s lines of code or specific types of code — i’m not sure i can diagnose it well given i’m not particularly technical :/

if i had to say where id draw the line — anything beyond simple python scripts for menial task automation or boilerplate CSS / HTML, id be very wary of GPT tbh