I think it’s telling that llama.cpp, the interface engine for LLMs on desktops, has blanket banned vibe coded PRs, as they simply cannot deal with them with the manpower they have.
…An LLM inference project. Banning LLM vibe coding.
I remember a collective groan from the devs when Google’s TurboQuant paper trended on Twitter, and they knew there’d be a flood of Claude PRs to “implement” the feature, even though most of it already existed in the engine.
vllm (the premiere enterprise server inference engine) is quite careful about them too, though they use bots to filter them and have more manpower to cull what slips through.