niplav @niplav@schelling.pt

guy who takes mdma at a techno rave and says “so this is what it's all about”, but not just about raving culture, but, like, life in general

this firefox isn't just leaking memory, it's gushing as if someone had performed shechita on it

use known inconsistencies of human preferences as value-learning trip-wires: if the value learning algorithm hasn't learned them yet, it's operating at the wrong level of abstraction.

i wonder if miri is looking out for people reinventing parts of their hidden agendas, as a kind of validation set

a vector field is just a kind of graph, mathematicians are silly for inventing unnecessary stuff

Oh, that's nice, 6 different people cloned the repo for my website

1½ thousand times

my reply guys have reply guys, and they have reply guys too, and deeper still; together we implement an HCH tree.

Gilgamesh girls

Hm, so this can't really encode all inconsistent preferences (e.g. the allais paradox can't be).

Maybe a continuous graph in the probability simplex over the possible subsets of options?

And transform that into a graph over the probability simplex over the options?

Look! I made a physical object! How quaint.

https://niplav.github.io/notes.html#Building-a-Cuboid

annoyance: terminology of "maximin"/"minimax" has some deep symmetry breaking

Me: "Now this is *truly* cuboidal. Don't you think, old chap?"
Cuboid: *bwwwwwwwwwww*
Me: *slams table* "Ha, yes!"

pump and dump? no, more like cuddle and befuddle

Sure, the big five model is okay, but I prefer a *real* personality model: dishwasher totalitarians versus dishwasher anarchists

What does an enthusiast for upper body clothing say when they see a new piece?

“Intewesting”