Post #1926513

2026-03-14 12:34 UTC

@fvsch there are definitely reasons to be concerned about code quality if trusting these models too much, and code reviews only have immediate effects, they don't learn good practices like a human would. agents files and skills can help, but no avoiding knowing the codebase. there are free options, like qwen and deepseek (chinese) if you don't want to use american companies, or run localy (but expensive hardware needed for even moederately sized models, the frontier ones are out of reach.)

Replies (2)

@kotaro@kotaro.me 2026-03-14 12:48

@tshirtman @fvsch Using cloud LLMs like always-on, autonomous agents gets expensive. While they've gotten better at following instructions compared to before, you can't say they're cost-effective by any means. And if you try to do the same thing with local LLMs, they demand 128GB or more of VRAM. Otherwise, they can't even use their full context window and just get stuck in a loop of small failures. In short, it's just a toy for people with six-figure salaries.

Open ##1926514
@tshirtman@mas.to 2026-03-14 12:48

@fvsch reading through a codebase and orienting is definitely a thing they can help with, using local tools like grep to look for pointers because generally can't keep a whole codebase in their context, be it codex, claude, amp, or others, it should take a lot less tokens to do that than generating code, so you can probably rely on free tiers, or local models, if you have a decent enough gpu and are patient.

Open ##1926515