Agreeing With the LLM Critics and Using LLMs Anyway

Writing from Berlin’s Local-First Conf, the author names a dissonance he saw everywhere: engineers who openly criticize LLMs—and applaud the criticism—while keeping Claude Code open on their laptops. The tell was Armin Ronacher, Flask’s creator, admitting on stage that his own coding-agent project auto-closes nearly all PRs and issues to survive the flood of machine-generated submissions. The people building tooling for LLMs are drowning in their own creation.

The author concedes the critics are largely right. LLMs produce slop, are trained on copyrighted material, carry environmental and ethical costs, and sit inside an investment bubble he expects to burst. His sharpest worry is trust: a well-formed PR used to signal real human effort, and that filter is gone, which he thinks could kill open source unless contribution trust is rebuilt (verified contributors, real-life meetups). He adds the collapse of the junior-engineer pipeline—seniors can no longer gauge effort behind a junior’s code and have little incentive to teach when mundane work goes to a model—plus geopolitical exposure, citing the US cutting non-citizens off from Anthropic’s frontier model, and the quiet homogenization of opinion that comes from conversing with a machine trained on the majority view.

His case for using them anyway is pragmatic: they aren’t going away, so the better move is to shape the flow rather than fight it. Local, open-weights models are the hedge—they keep large vendors honest on price, can’t be switched off overnight by a government, and will outlast any company that topples when the bubble pops. He notes even the AI-skeptical conference took local models seriously, and frames the piece as an invitation for others to describe the same contradiction openly.