alt Hacker News> This should be at minimum 8–10 separate modules.
Can't really say that for sure. The way humans structure code isn't some ideal best possible state of computer code, it's the ideal organization of computer code for human coders.
Nesting and cyclomatic complexity are indicators ("code smells"). They aren't guaranteed to lead to worse outcomes. If you have a function with 12 levels of nesting, but in each nest the first line is 'return true', you actually have 1 branch. If 2 of your 486 branch points are hit 99.999% of the time, the code is pretty dang efficient. You can't tell for sure if a design is actually good or bad until you run it a lot.
One thing we know for sure is LLMs write code differently than we do. They'll catch incredibly hard bugs while making beginner mistakes. I think we need a whole new way of analyzing their code. Our human programming rules are qualitative because it's too hard to prove if an average program does what we want. I think we need a new way to judge LLM code.
The worst outcome I can imagine would be forcing them to code exactly like we do. It just reinforces our own biases, and puts in the same bugs that we do. Vibe coding is a new paradigm, done by a new kind of intelligence. As we learn how to use it effectively, we should let the process of what works develop naturally. Evolution rather than intelligent design.
Replies
> One thing we know for sure is LLMs write code differently than we do.
Kind of. One thing we do know for certain is that LLMs degrade in performance with context length. You will undoubtedly get worse results if the LLM has to reason through long functions and high LOC files. You might get to a working state eventually, but only after burning many more tokens than if given the right amount of context.
> The worst outcome I can imagine would be forcing them to code exactly like we do.
You're treating "code smells" like cyclomatic complexity as something that is stylistic preference, but these best practices are backed by research. They became popular because teams across the industry analyzed code responsible for bugs/SEVs, and all found high correlation between these metrics and shipping defects.
Yes, coding standards should evolve, but... that's not saying anything new. We've been iterating on them for decades now.
I think the worst outcome would be throwing out our collective wisdom because the AI labs tell us to. It might be good to question who stands to benefit when LLMs aren't leveraged efficiently.
I’ve heard this take before, but if you’ve spent any time with llm’s I don’t understand how your take can be: “I should just let this thing that makes mistakes all the time and seems oblivious to the complexity it’s creating because it only observes small snippets out of context make it’s own decisions about architecture, this is just how it does things and I shouldn’t question it.”
I think this view assumes no human will/should ever read the code. This is considered bad practice because someone else will not understand the code as well whether written by a human or agent. Unless 0% human oversight is needed anymore agents should still code like us.
I don't buy this. Claude doesn't usually have any issues understanding my code. It has tons of issues understanding its code.
The difference between my code and Claude's code is that when my code is getting too complex to fit in my head, I stop and refactor it, since for me understanding the code is a prerequisite for writing code.
Claude, on the other hand, will simply keep generating code well past the point when it has lost comprehension. I have to stop, revert, and tell it to do it again with a new prompt.
If anything, Claude has a greater need for structure than me since the entire task has to fit in the relatively small context window.