Stuart_Armstrong

Tech­ni­cal model re­fine­ment formalism

Model splin­ter­ing: mov­ing from one im­perfect model to another

Learn­ing hu­man prefer­ences: black-box, white-box, and struc­tured white-box access

AI safety as feather­less bipeds *with broad flat nails*

Learn­ing hu­man prefer­ences: op­ti­mistic and pes­simistic scenarios

Strong im­pli­ca­tion of prefer­ence uncertainty

“Go west, young man!”—Prefer­ences in (im­perfect) maps

Learn­ing Values in Practice

The Gold­bach con­jec­ture is prob­a­bly cor­rect; so was Fer­mat’s last theorem

Why is the im­pact penalty time-in­con­sis­tent?

Dy­namic in­con­sis­tency of the in­ac­tion and ini­tial state baseline

Models, myths, dreams, and Cheshire cat grins

Re­sults of $1,000 Or­a­cle con­test!

Com­par­ing re­ward learn­ing/​re­ward tam­per­ing formalisms

Prob­a­bil­ities, weights, sums: pretty much the same for re­ward functions

Learn­ing and ma­nipu­lat­ing learning

Re­ward func­tions and up­dat­ing as­sump­tions can hide a mul­ti­tude of sins

How should AIs up­date a prior over hu­man prefer­ences?

Dist­in­guish­ing lo­gis­tic curves

Dist­in­guish­ing lo­gis­tic curves: visual