Also, the more you set up barriers on capability, the easier you become a top-down target for public power. If from the start you have lots of small models, and none of them wants to be vulnerability-hunting hackers — our small models might only want better Taiwanese, Hakka, or Indigenous languages. When you do culture-related work, you orchestrate these small models together — Taiwan also has the important language Verilog for chip design, so you build something like SiliconMind — those are unlikely to become so big-brained and broadly capable that they accidentally synthesise biological viruses. They just do their best inside one small Kami world.