[bugglebeetle]

I hope there’s still room for improvement. I haven’t been able to get ChatGPT to do anything coherent for stuff with > 5 functions or logic that is even the slightest complex. It likewise frequently stumbles/hallucinates when trying to integrate revisions. This is all using GPT-4 as well.

What’s it’s absolutely fantastic at is the small, one-off scripts that were tedious and time-consuming to write, integrating well-specified changes into existing code, tests, and boiler plate config stuff.

I know this is all “for now” talk, but it will be interesting to see how quickly or if at all it can get to production-ready code in the absence of pretty thorough review by someone who knows how stuff actually works. Natural language is a fantastic new interface, but (for now) you still have to know how to describe stuff that actually works.

[discordance]

My experience has been similar. Not to say it hasn't been helpful in 'larger' apps, but I've needed to scaffold/stub out the functions, and then verify GPT-4 generated code passes the appropriate tests.