Jenn had no peanuts at home this morning. She walked through the grocery store mulling over her options, brushing her hand against her bloody head. She had just bumped into a large metal rod, but continued to fix her attention on the peanuts.
Moments later, Jenn collapsed, and was rushed to the hospital. She arrived a corpse.
Jenn was unreasonable here, if she values life. Not having peanuts might make her life slightly less enjoyable, but immediate attention to the wound gives her a shot at many more years.
Say you can take a year to improve the water quality of your million person city, on average making everyone's life there a year longer. The interventions are pretty specific, and probably won't generalize well to other cities, but you can still spend a year to save 100,000 years of life.
But could you be saving more? Imagine the number of human years lived in the course of future civilization. For now, lay aside crucial issues like the quality of those years, and whether human years are the metric you most care about.
The universe is huge. And there are even a lot of resources in it, resources that could be used for building humans who have live many years. For estimates of the universe's potential for human lives, see Astronomical Waste. Let's just say for now that there are 10^20 years of human life could reasonably happen. Then it would be enormously valuable to take a year to get a 1/10^5 shot of saving a 10^5 fraction of these people; worth ten billion lives in expectation. If we were able to identify such an action with high confidence, and were interested in maximizing the amount of human life, we'd want to do this instead of treating the water. (Scope-intensivity, difficulties estimating probabilities, and the computational difficulty of seeing all the implications of a given action should make us doubt our estimates on such actions, but the principle shines through.)