Dmitrii Aleksandrov

    My Take on LLMs for Coding

    Posted on Last edited 2 mins

    Tech

    Originally written as an HN comment on “The Rise of Whatever” :

    I disagree on LLMs being “worse than useless”.

    Sure, “vibe coding” an entire app from a short prompt will always give you fragile, subtly broken nonsense. Code is the spec. In most cases, you can’t meaningfully “compress” your requirements into a short informal prompt. We need better formal languages for expressing requirements concisely and declaratively! Think: Prolog, Haskell…

    LLMs are good at small tasks that you can review much quicker than doing it yourself. Something tedious, like doing some local refactoring, writing ad-hoc Bash scripts, SQL queries, FFmpeg commands. I use Bash and SQL regularly, but somehow I always have to google the exact syntax. I already use ShellCheck , by the way. It’s a must, and it helps a lot when reviewing LLM output.

    I like the autocomplete feature too. It often saves time when writing repetitive or obvious code. if bad_stuff { usually autocompletes return Err(BadStuff) for me. MyStruct { initializer usually autocompletes the list of fields for me. I know that incorrect suggestions piss off some people and make it a net-negative for them. Incorrect suggestions are common, but they don’t bother me in practice.

    To clarify where I’m coming from. I’ve been using Copilot (Claude) in VSCode for a month (full-time), and ChatGPT (sparingly) for two years before that. I mostly write Rust.

    Don’t jump to conclusions after reading this post, or any other posts online. Just try “vibe coding” yourself. Try autocomplete yourself. Try various tasks yourself. It really doesn’t take that much time to get the hang of it and form your own opinion that’s actually based on reality.

    Discuss

    Comments