Remix.run Logo
tripledry 2 hours ago

Fun example I've had (some weeks ago), Agent completely dismissed the lack of strong consistency in our db system, this tiny bug would have caused a massive problem in the future.

A bug that is not immediately obvious, no syntax error, logic is sound, works as expected on local and QA envs. But let it run in prod for 2 months and you have a massive problem.

I'm not dismissing LLM's here btw, just pointing out that there is probably a lot of things that go into the category "you don't know what you don't know". These kinds of issues may or may not be a problem depending on the business you are in.

Most of the time the latest models perform better than I expect.

ethin an hour ago | parent [-]

Ouchy. That isn't as bad as the slop I've seen: one example I saw a few weeks ago from vibe-coded software had this absolutely mind-blowing bit of Rust code:

pub(crate) fn zeroed_safe<T>() -> T { unsafe { std::mem::zeroed() } }

pub(crate) fn read_unaligned_safe<T: Copy>(src: const T) -> T { unsafe { std::ptr::read_unaligned(src) } }

pub(crate) fn box_from_raw_safe<T>(ptr: mut T) -> Box<T> { unsafe { Box::from_raw(ptr) } }

pub(crate) fn isize_to_wndproc_safe(value: isize) -> WNDPROC { unsafe { Some(std::mem::transmute::< isize, unsafe extern "system" fn(HWND, u32, WPARAM, LPARAM) -> LRESULT, >(value)) } }

No, I'm not joking either, this is actual code that Claude wrote (and I'm pretty sure it's the latest models too but I don't actually know for certain, since the "dev" never specified). This idiom is repeated about 238 more times through that one file. I have tried repeatedly to help this person out but they're the kind of vibe-coder who thinks they know best, and who will take your advice and drop it into CLAUDE.md verbatim, and then Claude will go off and do the most literal interpretation of that text and not best practice.

To be clear, I don't hate LLMs. They're really useful for very specific things or where you know the domain very well. But then I see slop like this (and the dependency many people have on them -- I have literally heard my own close acquaintances say that they could not see themselves without a Claude subscription) and it really hammers home that software quality is not at all going to get better because of these things unless something changes.

Edit: grammar/spelling

a2ff6eeb0 29 minutes ago | parent [-]

Yeah, trying too hard to guide with hard rules, instead of asking the LLM for self-review, is going to go wrong. The LLM is much better at code review than it is at writing code.

Use a high effort model and let the llm review it's output and figure it out on its own.

The code still won't be great, but it'll be good enough.