The three posts on Selection Theorems could generally use some distillation and better marketing; the “selection theorems” name is quite bad, and the empirical aspects should be emphasized more
There’s been afewposts and a lot of comments on those posts between myself, Evan, and Abram arguing about the right way to think of “outer” vs “inner” alignment. (My comment on that last linked post is the best current summary of my thoughts.)
Generalized Koopman-Pitman-Darmois is probably a very hard one to distill, but it would probably be valuable if someone could explain the argument more intuitively. (Really, the right way to do it is to figure out a proof which explicitly routes through an entropy maximization problem, but that’s more a technical goal than a distillation goal.)
More generally, two big categories:
I’ve written a ton of material on general background world models, and a ton of material on alignment, but I’ve written relatively little explaining how the background world models narrow down the search-for-alignment-progress to the sort of work I’m doing. Or, to put it a different way: a lot of my technical posts could use good explanations of why the results are interesting and how they fit into the big picture.
Important stuff is often buried in comment threads. I’m not sure if LW currently has a way to rank a user’s comments by Karma, but that would be useful to find such threads.
Finally, since it’s Thomas Kwa asking this question: no, I am not going to create bounties for distillations on these right now, because I don’t want to deal with the overhead. Fortunately, the target audience for distillations of my work is everyone except me, so people other than me are quite well qualified to set up their own distillation bounties.
(Ordered by priority)
The Pointers Problem and Variables Don’t Represent The Physical World
The three posts on Selection Theorems could generally use some distillation and better marketing; the “selection theorems” name is quite bad, and the empirical aspects should be emphasized more
There’s been a few posts and a lot of comments on those posts between myself, Evan, and Abram arguing about the right way to think of “outer” vs “inner” alignment. (My comment on that last linked post is the best current summary of my thoughts.)
How To Think About Overparameterized Models. Also a distillation of the relevant parts of the Mingard et al work would go well with this.
My review of Coherent Decisions
Abstractions as Redundant Information
Anything in the Big Picture of Alignment talks
Generalized Koopman-Pitman-Darmois is probably a very hard one to distill, but it would probably be valuable if someone could explain the argument more intuitively. (Really, the right way to do it is to figure out a proof which explicitly routes through an entropy maximization problem, but that’s more a technical goal than a distillation goal.)
More generally, two big categories:
I’ve written a ton of material on general background world models, and a ton of material on alignment, but I’ve written relatively little explaining how the background world models narrow down the search-for-alignment-progress to the sort of work I’m doing. Or, to put it a different way: a lot of my technical posts could use good explanations of why the results are interesting and how they fit into the big picture.
Important stuff is often buried in comment threads. I’m not sure if LW currently has a way to rank a user’s comments by Karma, but that would be useful to find such threads.
Finally, since it’s Thomas Kwa asking this question: no, I am not going to create bounties for distillations on these right now, because I don’t want to deal with the overhead. Fortunately, the target audience for distillations of my work is everyone except me, so people other than me are quite well qualified to set up their own distillation bounties.