Hacker News Reader

Comment by notepad0x90 | Hacker News Reader

Big part of coding is understanding the code and making decisions on what needs to be added/removed/changed. LLMs can code, even if they generate perfect code every single time, someone still needs to read the code and make decisions. Others speak about understanding business logic, interfacing with clients and stakeholders,etc... I get that, but even without that, someone will always need to decide how things should be done, improved,etc.. LLMs are not going to benchmark your code and they will never understand the developer/client's intent perfectly.

Why are LLMs in these context being viewed as more than really powerful auto-complete/intellisense? I mean, correct me if I'm wrong, but aren't LLMs still years away from auto-generating a complete complex codebase that works without fixing a lot of bugs/assumptions?

View on HN

It's funny how we'd like to chop off the head of that junior developer who's behind that pull request of a perfectly working but not enough clean code, but we happily welcome spaghetti and buggy LLM code because sisterhood and brotherhood claim it's trendy. Herd mentality.

There's an uncomfortable truth backing this, at least for me: revising the code of a LLM and changing my cursorrules is easier than fixing up a junior's code and teaching them _why_ I'd want to do something a particular way. After all, it's a tool, not a person, and you can be much rougher with your tools.

In my experience, you kinda get used to other people's mindset and after a while it becomes second nature to look for their mistakes or to understand their approaches.

Well I, for one, have infinitely more patience for my juniors than whatever nonsense an LLM spits out. Far more pleasant to teach a person something new than to fix autocomplete output.

> LLMs are not going to benchmark your code

If I may focus on this particular line without invalidating the rest of the content -- why aren't they going to benchmark the code?

You hardly need more than a simple script to just run a generic benchmark tool. But to measure performance of your code and figure out what needs to be tuned, I don't think LLMs are there, am I wrong? They can't understand the code they generated well enough to make small adjustments and improve the benchmark score. For a dev to do that on LLM generated code, it would have to understand the codebase as intended by the LLM at least.

I had them write me performance tests. I guess I could go back to that and ask it to improve the code based on the output. You can feed the output automatically now with Roo Code, Aider, Cline, windsurf, cursor, augment, etc.

Reply on news.ycombinator.com