Florian_Dietz

Con­text Mod­ifi­ca­tion as a Nega­tive Align­ment Tax

Con­text Mod­ifi­ca­tion as a Nega­tive Align­ment Tax

Con­text Mod­ifi­ca­tion as a Nega­tive Align­ment Tax

Pos­i­tive Feed­back Only

Split Per­son­al­ity Train­ing can de­tect Align­ment Faking

Already Optimized

De­cep­tion Chan­nel­ing: Train­ing Models to Always Ver­bal­ize Align­ment Faking

The Miss­ing Se­quence: Why Cor­rect Anal­y­sis Makes Ter­rible Ac­tion Guides

De­liber­ate Epistemic Uncer­tainty: An Au­to­mated Ex­per­i­ment on AI Self-Reporting

An Abla­tion Study on the Role of [Un­trans­lat­able] in Co­op­er­a­tive Equil­ibrium For­ma­tion: Emer­gent Ra­tion­al­iza­tion Un­der Miss­ing Primitives

Split Per­son­al­ity Train­ing: Re­veal­ing La­tent Knowl­edge Through Alter­nate Per­son­al­ities (Re­search Re­port)

De­liber­a­tive Credit As­sign­ment (DCA): Mak­ing Faith­ful Rea­son­ing Profitable

De­liber­a­tive Credit As­sign­ment: Mak­ing Faith­ful Rea­son­ing Profitable

Edge Cases in AI Alignment

Split Per­son­al­ity Train­ing: Re­veal­ing La­tent Knowl­edge Through Per­son­al­ity-Shift Tokens

Do we want al­ign­ment fak­ing?

Re­veal­ing al­ign­ment fak­ing with a sin­gle prompt

Flo­rian_Dietz’s Shortform

Achiev­ing AI Align­ment through De­liber­ate Uncer­tainty in Mul­ti­a­gent Systems

Un­der­stand­ing differ­ences be­tween hu­mans and in­tel­li­gence-in-gen­eral to build safe AGI