I think we're all basically agreeing with each other here now. Language is indeed a difficult tool.

"there's a difference between hypothetical probability and the actual benefits of the specific bed nets in the specific locations that are funded by your donation."

I don't think there really is, from a decision standpoint. According to previous statistics from Givewell, $2,500 is approximately what is needed to save a life via malaria nets (this figure is now outdated. Now it seems very different, but I think that Tom made the claim from when that was the case).

If this is true, than $5,000 would, on average, save 2 statistical lives.

Yes, it is possible that it wouldn't save anyone. But it's also possible it would save more than 2. Future unpredictable outcomes don't matter for decision making outside their probabilities at the time.

It's possible that any intervention wouldn't help anyone, even the GiveDirectly donations (though it is quite unlikely). That said, when making decisions it really seems like it's best to go on the best estimates, especially if benefits scale linearly (which they should do at this level of spending).

That said, I would definitely agree that the other benefits of giving to AMF are quite considerable and may be greater than the 'saving a life' part. 'Saving a life' in itself is a pretty poor metric for many reasons.

* I may be misunderstanding what you said, but to me it sounded as an argument of 'we can't make decisions on things we can't know for sure, because what if X happens?' This kind of argument is incredibly common and incredibly frustrating. Douglas Hubbard wrote a nice summary of this argument coming from who are supposed to be 'risk managers' in his book 'The Failure of Risk Management' here: http://www.amazon.com/Failure-Risk-Management-Why-Broken-ebook/dp/B0026LTMAU/ref=la_B001JSJHIS_1_2?s=books&ie=UTF8&qid=1390915736&sr=1-2