Improving on Effective Altruism II

Socialism gives us the criticisms, that certain classes of strategies will not fucking work. That having an impact is fucking hard, because look at these examples of people who tried to, and then systemic problems undid their work.

Which you can use to realize that most of what EA is doing is going to be useless. Like, completely predictably. An easy example is how Animal Charity Evaluators just assumes that corporations will keep their promises about treating caged hens better. And when someone asks them about it, they just say, “we did an expected value calculation that says the utility of getting the corporations to follow through on this promise is x, so since they might not follow through, we’ll just guess that there’s a 50% chance they’ll follow through. So, we’ll take x/2 as the actual utility of getting corporations to promise to improve conditions.”

But the issue here is that you’re assigning a probability, to something you can know. If you dig deep enough, you can figure out that of course corporations won’t keep their words on this, because incentives.

You can make the same argument about AI safety being the most effective thing to work on just on first principles, and the correct response is again to dig into the actual details, and figure out if this is true for any given approach to AI safety.

By dig into the actual details, I mean something like, imagine you’re researching whether to take a supplement, and so you dig into the literature for 30 hours about it. Except that when you’re doing this for “is Animal Charity Evaluators correct, or is this line of AI safety work useful”, you have to use all your knowledge about how people are likely lying to you to weed out persuasion attempts cloaked as information. Maybe in practice there’s no relevant existing work you can look at, and you have to do everything yourself, like being an independent scientist. So you read this and work out a potential way around some of deep learning’s shortcomings and pseudocode it out and think about the ways alignment could go wrong in this particular system rather than in general. The point is that this process needs to be grounded in digging through reality.

If you’re operating on information someone gave you about what’s likely to be effective, or how many years it’s going to be before someone makes an artificial general intelligence, I would say: you are going to have to re-evaluate those claims from the ground up if you want to make an impact. Yeah your time is limited, but if you’re going to do better, you have to actually reason through all the steps in the path you’re planning to take. Because blind spots and things you’re not checking are not made at random, they’re actually systematically likely to contain mistakes.

Next post: Improving on Socialism