The core worry: a system smarter than us that we can't reliably steer. Corrigibility, interpretability, and formal verification each address part of it, but none feels like a complete answer. For those who think the control problem is solvable in principle, what research direction gives you the most hope?