3
12bitfloat
233d

Why I still don't fully believe in vibe coding: Even the frontier models are constantly wrong!

Claude says c# structs with LayoutKind.Auto can have differing layout for each usage or JIT compilation. ChatGPT says they can't (and that's what I assume is true)

Claude says to "Multiply by a large prime", 0x9e3779b97f4a7c15
...That's NOT a prime!

AI generated code looks really convincing but they always make these little mistakes. And even if you inspect the AI code yourself (which let's be honest, we're all too lazy for that :P), will you be able to spot these? I really doubt it

Comments
  • 4
    Yeah, basically that‘s my rule of thumb for using AI:
    If it being wrong wouldn’t be that big of a deal, go for it. Otherwise stay the fuck away.
  • 1
    @whimsical > 'Before AI, I didn't do docs. Now it became one of my favo things. Projects are more finished than ever. Even when I write something myself, at the end it will go trough AI for finishing touch, without exception.'.

    Well... just remember that you could be an outlier in this whole thing... also, the docs could be wrong, or incomplete... like, for example, one of your dwn manuals directed to something generic like 'git clone https://www.my_company.com/my_repo' or something similar...
  • 0
    Haven't tried Claude, but with Kiro it feels like instructing a team of newly graduates. But instead of it taking a weak to get result I barely have time for a coffee before they are done. And I can pint out the faults and get them going again.
Add Comment