I just had a lovely afternoon of trying to unwind a mess someone vibe coded their way into with Opus 4.6… Which I have to mention model and version every time or else someone comes along and says ‘Well Opus 4.6 won’t make mistakes!’ when it absolutely does and I’m so tired of people overstating how good Opus is.
So far they create the mistakes quickly and when it happens they generally are at a loss on what to do, and it becomes messy because the vibe coder went way past the point of failure and now I have to extricate what the hell they were doing…
Though I will say today the ‘review my commit’ feature did catch one issue legitimately, though it flagged about 5 others erroneously and in a couple of cases really really wanted to change correct code to do exactly the wrong thing for some reason.
The codegen can be useful if watched carefully, but things will happen like after porting dozens of functions mostly fine to a replacement library, for some reason it decided to delete a function and replace it with an if/else chain of every test case and hard code passing results rather than porting the function to the new library (which upon manual inspection was just a two line change)…
Which is why you use it simply as a proofreader and then evaluate the mistakes it brought up. You don’t blindly trust it or worse, use it to write everything from scratch, because then you’re just going to deliver shit code.
I see people hating on AI for things that are far more the fault of how people use it than of the technology itself. AI is extremely over hyped, but if you understand it for what it is, it’s a fairly useful tool.
I just had a lovely afternoon of trying to unwind a mess someone vibe coded their way into with Opus 4.6… Which I have to mention model and version every time or else someone comes along and says ‘Well Opus 4.6 won’t make mistakes!’ when it absolutely does and I’m so tired of people overstating how good Opus is.
So far they create the mistakes quickly and when it happens they generally are at a loss on what to do, and it becomes messy because the vibe coder went way past the point of failure and now I have to extricate what the hell they were doing…
Though I will say today the ‘review my commit’ feature did catch one issue legitimately, though it flagged about 5 others erroneously and in a couple of cases really really wanted to change correct code to do exactly the wrong thing for some reason.
The codegen can be useful if watched carefully, but things will happen like after porting dozens of functions mostly fine to a replacement library, for some reason it decided to delete a function and replace it with an if/else chain of every test case and hard code passing results rather than porting the function to the new library (which upon manual inspection was just a two line change)…
Which is why you use it simply as a proofreader and then evaluate the mistakes it brought up. You don’t blindly trust it or worse, use it to write everything from scratch, because then you’re just going to deliver shit code.
I see people hating on AI for things that are far more the fault of how people use it than of the technology itself. AI is extremely over hyped, but if you understand it for what it is, it’s a fairly useful tool.