Post

DO temporarily, it's reliable and easy to setup. For my EU clients, they'll be moving over to boxes on Hetzner. For my Canadian clients, Robson Inc in Vancouver. Yes, these three hosts have GPUs and talk about LLMs - But they aren't going out of their way to glorify them, and chastise those against.

I've used LLMs - running models locally, integrated into my editors (gemma3:4b is pretty alright in this situation). Overall they're fine. For small things, good autocomplete but somewhat annoying when they're way wrong.

0 reposts0 likes

For my larger projects, they're abysmall or don't work at all - recently working with 2+ million lines of PHP which is not enjoyable on the best days, I can't imagine the LLM thought any better.

I vibe coded with cursor on Friday - properly - no peaking at the code, just build the thing to see if I like overall concept. It wasn't anything better than I can sketch quickly in a notebook and plan (or even the stack I would chose - loves tailwind 2 for some reason...).

Overall, I don't find much of an improvement in speed over snippets and lsp autocomplete, and sometimes after relying on it too heavily I feel stupider and takes me longer to figure out what I want to do.