Discussion about this post

User's avatar
Tim's avatar

We need to stop concerning ourselves so much with individual behavior and carbon footprints and start focusing on the small collection of giant corporations (not to mention the US military, the biggest polluting force on the planet) that contribute over 70 percent of all global carbon emissions

Expand full comment
Sam B Johnson's avatar

This is a poorly conceived article.

The premise is that when one optimizes for thing X, one is implicitly ignoring everything else that's not explicitly X. Of course this is correct; this underspecification problem is well known in computer science. The issue I have with Trembath is that all those externalaities are equally unmeasured in the original algorithm as in the eco-friendly algorithm. Once one notices this, Trembath's piece collapses to an argument that changing things can lead to new negative consequences; the issue is that he gives no justification (and there is essentially none that can be given*) that these consequences will be worse than the original unmeasured consequences. For example: the argument that Google is optimizing for speed alone could be used to conclude that Google might often direct people through neighborhoods where the speed limit is too high and therefore dangerous to pedestrians. This consideration exists equally for the speed-only algorithm and for the eco-friendly algorithm.

For this reason, I find that the main objective point of the piece is incorrect. The rest is opinion. He brings up some fine points. It's just that they're all organized around an incoherently presented thesis. Here is my condensed version of this article which captures what I find valuable in it: "We should attempt to quantify side-effects of algorithms so we better understand their net utility."

* In order to be friendly to Trembath, I can imagine one making the heuristic argument that the eco-friendly algorithm is pursuing two objectives at once and is therefore more constrained in its solution space; thus, the compromises it makes on unspecified externalities will be more extreme. This is a natural outcome to suspect, but it is hard to quantify--and certainly Trembath makes no effort to do so. In this case, the solution space is so large to begin with that my hunch is that the additional constraint is a drop in the bucket.

Expand full comment
4 more comments...

No posts