This is technically just a link post for those who didn't see the book yet. The main idea, as I understand it, is that while some physical theories are beautiful, other are complex and "ugly". Beauty should not be taken as evidence for truth.

It looks like sometimes in AI safety research aesthetics may be taken as an evidence (I will not provide examples, as in each case it may be just my interpretation), and thus possibility of such beauty bias should be taken into account.

7 comments, sorted by Click to highlight new comments since: Today at 10:05 AM
New Comment

"Beauty should not be taken as evidence for truth."

Shouldn't it? Isn't taking "beauty" as evidence just a way of phrasing Occam's razor?

Beauty is a hint and an inspiration, not evidence. Sometimes it guides you some place useful, and sometimes it leads you completely astray. Like it has with the String Theory and countless unified field theories.

Yes, but there is a problem of what I called "median complexity of the world descriptions", which is probably answered somewhere but I don't know where to look.

In other words, Occam razor doesn't mean that the simplest explanation is true. It means that the simpler explanations are more probable to be true than more complex ones. The difference between the two definitions is the way how the truth is distributed over complexity of the explanations.

In the first case, the distribution is very steep, so the simplest explanation is more probable than all more complex explanation combined. In the second case, the truth(complexity) function declines slowly, so may be first 100 explanations combined have 0.5 probability, - in that case, it is unlikely that the simplest explanation will be true.

This article on Solomonoff induction goes over a lot of the related considerations.

Yes, it is a good post, but doesn't cover the problem of median complexity directly.

Sure, but the fact that the probability distribution is skewed in favor of simpler (i.e. more "beautiful") explanations by Occam's Razor is equivalent to saying that there should be such a bias -- after all, bias is essentially just a skewing of one's probability function. Of course this bias shouldn't be taken to the extreme of assuming that just because one hypothesis is more beautiful than others, it automatically qualifies as the correct explanation. But discrediting such an extreme mindset doesn't mean that a mild bias in favor of "beauty" is discredited.

Haven't read the book, but her blog is one of half a dozen sites I follow regularly. It talks about how pushing for subjective "beauty" over other considerations is not the most useful approach.