[qumpis]

How do you define extrapolation at the scale LLMs operate? Even if you work with unseen-to-model software, it seems sufficient to "understand" the documentation and code examples to orient itself to helpful context.

That's why I'm scared. Once embedding whole codebases becomes a viability, I expect many opinions to change too

[RamblingCTO]

I invite you to play around with the code it generates. For example, yesterday I tasked it to generate a gpt-3.5-turbo API client. In the time I needed to get it running I could've wrote it myself. And that's < 100 lines. Don't even get me started on architecture, contextual decisions, clean code etc.

[qumpis]

Interesting that it's so insufficient for you, maybe indeed stuff you do is novel and would require lots of instructing before helping you.

Personally, my usecases involve quite standalone applications. copy pasted from another thread (using gpt4):

My personal use of gpt4 (also daily) is: correct, rephrase spelling from my brain dump, make python plots (stylize, convert, add subplots, labels, handle indexing when things get inverted), make short shell scripts (generated 2FA, login vpn through console using 2fa, make script of disabling keyboard etc), and help debug my code (my situation is this, here's some code, what do you suggest?).

[RamblingCTO]

I would agree with summarization and NLP-driven tasks and I will actually add chatGPT to my side-gig for that. But code-wise it's not as big of a help as I'd like.

Good for you that it actually helps with coding! I like copilot a lot for auto-completion tho, but the rest of my work apparently is too complex, yes.