Buck(Buck Shlegeris)

Notes on con­trol eval­u­a­tions for safety cases

Toy mod­els of AI con­trol for con­cen­trated catas­tro­phe prevention

The case for en­sur­ing that pow­er­ful AIs are controlled

Manag­ing catas­trophic mi­suse with­out ro­bust AIs

Catch­ing AIs red-handed

Mea­sure­ment tam­per­ing de­tec­tion as a spe­cial case of weak-to-strong generalization

Scal­able Over­sight and Weak-to-Strong Gen­er­al­iza­tion: Com­pat­i­ble ap­proaches to the same problem

AI Con­trol: Im­prov­ing Safety De­spite In­ten­tional Subversion

How use­ful is mechanis­tic in­ter­pretabil­ity?

Un­trusted smart mod­els and trusted dumb models

Pro­gram­matic back­doors: DNNs can use SGD to run ar­bi­trary state­ful computation

Meta-level ad­ver­sar­ial eval­u­a­tion of over­sight tech­niques might al­low ro­bust mea­sure­ment of their adequacy

A fresh­man year dur­ing the AI midgame: my ap­proach to the next year

One-layer trans­form­ers aren’t equiv­a­lent to a set of skip-trigrams

Try­ing to dis­am­biguate differ­ent ques­tions about whether RLHF is “good”

Causal scrub­bing: re­sults on in­duc­tion heads

Causal scrub­bing: re­sults on a paren bal­ance checker

Causal scrub­bing: Appendix

Causal Scrub­bing: a method for rigor­ously test­ing in­ter­pretabil­ity hy­pothe­ses [Red­wood Re­search]

Multi-Com­po­nent Learn­ing and S-Curves