Stuart_Armstrong

Con­sis­ten­cies as (meta-)preferences

Why un­rig­gable *al­most* im­plies uninfluenceable

A pos­si­ble prefer­ence algorithm

If you don’t de­sign for ex­trap­o­la­tion, you’ll ex­trap­o­late poorly—pos­si­bly fatally

Which coun­ter­fac­tu­als should an AI fol­low?

Toy model of prefer­ence, bias, and ex­tra information

Prefer­ences and bi­ases, the in­for­ma­tion argument

Why sig­moids are so hard to predict

Con­nect­ing the good reg­u­la­tor the­o­rem with se­man­tics and sym­bol grounding

Carte­sian frames as gen­er­al­ised models

Gen­er­al­ised mod­els as a category

Coun­ter­fac­tual con­trol incentives

Short sum­mary of mAIry’s room

Syn­tax, se­man­tics, and sym­bol ground­ing, simplified

The ethics of AI for the Rout­ledge En­cy­clo­pe­dia of Philosophy

Ex­tor­tion beats brinks­man­ship, but the au­di­ence matters

Hu­mans are stun­ningly ra­tio­nal and stun­ningly irrational

Knowl­edge, ma­nipu­la­tion, and free will

De­hu­man­i­sa­tion *er­rors*

An­thro­po­mor­phi­sa­tion vs value learn­ing: type 1 vs type 2 errors