Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

The problem with these generative systems is that they're synthesizing existing code and their output isnt therefore always valid. Ie., they arent programming (which involves following discrete constraints of the language).

They are, therefore, largely tools for power-users. You need to understand why codepiolt/etc. is suggesting a completion before its actually useful.

If you prefer: Intelligence is knowing what question to ask, and knowing how to validate the answer.



Oh god no. No please don't let "power-users" anywhere near a Low Code framework. It doesn't take long for them to fuck up and go ask a real developer what is wrong. I've seen things...


> It doesn't take long for them to fuck up and go ask a real developer what is wrong. I've seen things...

Worse is when it doesn’t take them long to fuck up, but it does take them (or, often, other people in the org, possibly after they’ve left) long to bring in a dev...

I’ve seen things, too...


The sooner the better. Would you prefer stage one cancer or stage four cancer?


> The problem with these generative systems is that they're synthesizing existing code and their output isnt therefore always valid

Actual programmers have this problem too, though they often use non-AI tooling to catch and prompt them to correct invalid code.

No reason a system using an LLM for code generation couldn't do that, too.

Anticipating the “but even if it is valid it won’t always be correct” followup: well, valid human-written code also often has bugs.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: