Quintin Pope

My Ob­jec­tions to “We’re All Gonna Die with Eliezer Yud­kowsky”

The shard the­ory of hu­man values

Evolu­tion pro­vides no ev­i­dence for the sharp left turn

Quintin’s al­ign­ment pa­pers roundup—week 1

QAPR 5: grokking is maybe not *that* big a deal?

Evolu­tion is a bad anal­ogy for AGI: in­ner alignment

Re­search agenda: Su­per­vis­ing AIs im­prov­ing AIs

Quintin’s al­ign­ment pa­pers roundup—week 2

QAPR 4: In­duc­tive biases

The Case for Rad­i­cal Op­ti­mism about Interpretability

QAPR 3: in­ter­pretabil­ity-guided train­ing of neu­ral nets

Meta learn­ing to gra­di­ent hack

Hy­poth­e­sis: gra­di­ent de­scent prefers gen­eral circuits

[Linkpost] A Gen­eral Lan­guage As­sis­tant as a Lab­o­ra­tory for Alignment

[Question] What’s the “This AI is of moral con­cern.” fire alarm?

New GPT-3 competitor

[Question] How good is se­cu­rity for LessWrong and the Align­ment Fo­rum?

[Question] Any prior work on mu­ti­a­gent dy­nam­ics for con­tin­u­ous dis­tri­bu­tions over agents?

Idea: build al­ign­ment dataset for very ca­pa­ble models

A sim­ple way to make GPT-3 fol­low instructions