It’s the same cycle since the '70s. Whether it’s COBOL or VB.NET or vibe coding, the premise hasn’t changed.
There’s three broad categories of code:
Monkey code (random applets that are almost entirely business logic and non-critical)
Actual code (most things)
Crazy shit like kernel or browser code.
I can see vibe coding, situationally, lower the barrier to entry of (1). But also that’s no different from COBOL or VB.NET which both promise “MBAs can now write code”, which conveniently never extends to maintaining said code. And vibe coding doesn’t help with that either, ChatGPT is an awful debugger.
Your boss thinks ChatGPT will help with (2), but it either won’t or only very slightly as an advanced autocomplete. For any problem-solving that requires more specific domain knowledge than can automatically find its way into their tiny context windows, LLMs are essentially useless.
… So I’m not worried. Today’s vibe coders are yesterday’s script kiddies.
the amount of mistakes and and hallucinations ai has makes it actually take longer to code.
it’s the same old garbage in, garbage out….
it can kinda help you get started but that only saves you 10 minutes of reading documentation that you have to read anyway to make sure it didn’t make something up.
It seems OK at spewing out a bit of code it found on StackOverflow, or even joining two bits of code together, but it really falls apart when you poke at the edges of it’s knowledge.
And the problem is, neither you nor it knows where those limits are, and it very quickly goes from confident copy and paste to confident bullshit.
It even knows what excuses smell like, so it’ll give you one at random when you call it out.
yep. i’ve tried it a bit and the errors are blended in so well and seem so plausible, it’s worse than stack overflow….
even when just getting default arguments for a function it makes stuff up.
i do see it getting better at errors like that, but not much better….
It’s the same cycle since the '70s. Whether it’s COBOL or VB.NET or vibe coding, the premise hasn’t changed.
There’s three broad categories of code:
I can see vibe coding, situationally, lower the barrier to entry of (1). But also that’s no different from COBOL or VB.NET which both promise “MBAs can now write code”, which conveniently never extends to maintaining said code. And vibe coding doesn’t help with that either, ChatGPT is an awful debugger.
Your boss thinks ChatGPT will help with (2), but it either won’t or only very slightly as an advanced autocomplete. For any problem-solving that requires more specific domain knowledge than can automatically find its way into their tiny context windows, LLMs are essentially useless.
… So I’m not worried. Today’s vibe coders are yesterday’s script kiddies.
the amount of mistakes and and hallucinations ai has makes it actually take longer to code.
it’s the same old garbage in, garbage out….
it can kinda help you get started but that only saves you 10 minutes of reading documentation that you have to read anyway to make sure it didn’t make something up.
It seems OK at spewing out a bit of code it found on StackOverflow, or even joining two bits of code together, but it really falls apart when you poke at the edges of it’s knowledge.
And the problem is, neither you nor it knows where those limits are, and it very quickly goes from confident copy and paste to confident bullshit.
It even knows what excuses smell like, so it’ll give you one at random when you call it out.
yep. i’ve tried it a bit and the errors are blended in so well and seem so plausible, it’s worse than stack overflow….
even when just getting default arguments for a function it makes stuff up.
i do see it getting better at errors like that, but not much better….
Debugging is the hardest part, and now you get to spend all your time doing it