Comment by bigstrat2003
4 days ago
LLMs are worse at programming than any dev I've ever worked with. Yes, even $latest_model. They have no understanding or ability to reason, and they make mistakes no human would make. They are, in short, bad at programming.
I've seen developers pushing code repeatedly to car engine firmware with for-loops that weren't even valid C syntax, have yet to see a modern LLM do this kind of mistake
If I had coworkers with the ability to program like Sonnet 3.7 did, I'd be extremely happy. In reality, I can probably count 1-3 people like that, out of many over the years. Honestly, pair programming with Opus is better than any pair programming I had with a human, especially at understanding the reasons behind my designs.
You've not worked with average developers then, or this is a purely reactionary/emotional statement.
In my experience it's always been easier to work with terrible human programmers because the terribleness of their code is inherently bound to human comprehension. Of course I've seen some absolutely massive messes created by humans. But generally if a human wrote it then a human can understand it. Also it takes an amount of time and effort for a human to create this kind of mess which tends to actually be more effort than what a more experienced person will expend to fix what is wrong with it. LLMs flip all this on it's head. Plus there's the thing the Godot statement talks about. I have a completely different willingness to help a struggling noob rise through the same trajectory I did, compared to fixing someone else's LLM code.
I know what you mean regarding the class of mistakes, but strong disagree on the "worse" part.
Or rather I envy you for your experience with humans so far.