At Local-First Conf in Berlin, I witnessed a striking dissonance: engineers applauded talks criticizing LLMs while running Claude Code on their laptops. Armin Ronacher, creator of Flask and founder of Earendil (Pi.dev), admitted his team auto-closes nearly all LLM-generated PRs to protect their project from "slop," yet encourages human contributions.

I agree with almost every critique: LLMs ingest copyrighted material, harm the environment, concentrate power in NVIDIA/OpenAI bubbles, and erode trust in open-source contributions. Projects like Zig and Gentoo now reject LLM PRs entirely. Junior engineers face a crisis — seniors can't verify effort, and mundane tasks once used for mentoring are outsourced to models. Geopolitical risk is real: Anthropic abruptly disabled Fable 5 and Mythos 5 for non-US users due to export controls. Models also silently homogenize opinions toward training-data majorities.

Yet LLMs are here to stay. Local, open-weights models running on personal hardware offer independence from corporate APIs and government cutoffs. When the bubble bursts, these models remain. The conference's AI talks took local inference seriously — a "Star Trek" computer on your laptop, uncontrollable by any single entity.