[darawk]

> In terms of public policy you have to decide what you're optimizing for and that decision can't be made with data alone because it does not help resolve questions of value and fairness.

Sure, but once you decide, then you should use data to optimize it. Seems pretty straightforward, and I don't think anything in this article would disagree with that.

[bluGill]

Even after I know what I'm optimizing for I cannot use data to optimize for it. I know from experience that unintended results will happen. Some of them will be very base which will force me to come up with a whole new set of things to optimize for.

You are also assuming we can agree on what to optimize for. In fact we do not, and will not.

[darawk]

> Even after I know what I'm optimizing for I cannot use data to optimize for it. I know from experience that unintended results will happen. Some of them will be very base which will force me to come up with a whole new set of things to optimize for.

You can't use data to optimize anything? Google and Facebook must be wasting their time collecting all that user data, then.

> You are also assuming we can agree on what to optimize for. In fact we do not, and will not.

We agree on lots of things to optimize for. There are cases of disagreement, but very very broad agreement as well.

[bluGill]

>You can't use data to optimize anything?

No, I can use data, the problem is I don't know what all the effects of that optimization will be. So I constantly have to change what I'm optimizing for.

> We agree on lots of things to optimize for.

Broadly, but there are limited resources and each thing effects the other. So even though we agree that A and B are worth optimizing for, we will disagree on which is more important. Worse in many cases we will agree on A and B, but the data shows you cannot optimize for one without pessimism the other.

That is the broad agreement isn't really enough to do anything with, we need the details and there we disagree.

[darawk]

> No, I can use data, the problem is I don't know what all the effects of that optimization will be. So I constantly have to change what I'm optimizing for.

What are you trying to say here? Optimizing things with data is hard?

> Broadly, but there are limited resources and each thing effects the other. So even though we agree that A and B are worth optimizing for, we will disagree on which is more important. Worse in many cases we will agree on A and B, but the data shows you cannot optimize for one without pessimism the other.

I think we even agree broadly enough on the relative weights of many things to optimize for them. And I think we at the very least agree enough on things to partially optimize them, or pareto optimize them. In many cases there are low-hanging fruit to be picked that can optimize a metric that we all agree is good without sacrificing another.