Alek Westover

Sleeper Agent Back­door Re­sults Are Messy

An Em­piri­cal Study of Meth­ods for SFTing Opaque Rea­son­ing Models

How do LLMs gen­er­al­ize when we do train­ing that is in­tu­itively com­pat­i­ble with two off-dis­tri­bu­tion be­hav­iors?

Five ap­proaches to eval­u­at­ing train­ing-based con­trol measures

Model or­ganisms re­searchers should check whether high LRs defeat their model organisms

How will we do SFT on mod­els with opaque rea­son­ing?

Three vi­sions for diffuse control

The­o­ret­i­cal pre­dic­tions on the sam­ple effi­ciency of train­ing poli­cies and ac­ti­va­tion monitors

Four Down­sides of Train­ing Poli­cies Online

Method­olog­i­cal con­sid­er­a­tions in mak­ing ma­lign ini­tial­iza­tions for con­trol research

Notes on Soft­ware-Based Com­pute-Usage Verification

Alek Westover’s Shortform

Should AI Devel­op­ers Re­move Dis­cus­sion of AI Misal­ign­ment from AI Train­ing Data?

What train­ing data should de­vel­op­ers filter to re­duce risk from mis­al­igned AI? An ini­tial nar­row proposal

Why I think AI will go poorly for humanity

Safe Distil­la­tion With a Pow­er­ful Un­trusted AI