Hacker News new | past | comments | ask | show | jobs | submit login

I can see how you got that impression from my comment (if you ignore how I mentioned architecture), so let me elaborate:

It's the opposite. FizzBuzz and getting the syntax right is what LLMs are good at... but there's so much more nuance at being experienced with a language/framework/library/domain which senior engineers understand and LLMs don't.

Being able to write Elixir assisted by an LLM does not mean you can produce proper architecture and abstractions even if the high level ideas are right. It's the tacit knowledge and second-order thinking that you should hire for.

But the thing is, if someone cannot write Elixir without syntax errors unless using an LLM, well, that's a extremely good proxy that they don't know the ins and outs of the language, ecosystem, best practices... Years of tacit knowledge that LLMs fail to use because they're trained on a huge number of tutorial and entry-level code ridden with the wrong abstractions.

The only code worse than one that doesn't work is one that kinda works unless your requirements change ever so slightly. That's a liability and you will pay it with interests.

To give a concrete example: I am very experienced with React. Very. A lot. The code that LLMs write for it is horrid, bug-ridden, inflexible and often misuses its footgun-y APIs like `useEffect` like a junior fresh out of a boot camp would, directly contradicting the known best practices for maintainable (and often even just "correct") code. But yeah it superficially solves the problem. Kinda. But good luck when the system needs to evolve. If it cannot do proper code that's <500 lines how do you expect it to deal with massive systems that need to scale to 10s of KLOC across an ever-growing twine?

But management will be happy because the feature shipped and time to market was low... until you can no longer ship anything new and you go out of business.




Ah, sorry, I read your comment as disagreeing with me, now I see it's the opposite. Exactly, LLMs (for now) are good at writing low-level code, but we need someone to work on architecture.

I had an idea the other day of an LLM system that would start from a basic architecture of an app, and would zoom down and down on components until it wrote the entire codebase, module by module. I'll try that, it sounds promising.




The deadline for YC's W25 batch is 8pm PT tonight. Go for it!

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: