@eggsyntax @godoglyness @lumpenspace @fireobserver32 @Effective69ism @AfterDaylight @MikePFrank @repligate @indif4ent @Shoalst0ne @YeshuaGod22 @slimepriestess @immanencer @RobertHaisfield @doomslide @jd_pressman @pudepiedj @jpohhhh @xlr8harder @AlkahestMu @AITechnoPagan @anthrupad @amplifiedamp @loveinadoorway @parafactual @Jtronique @UltraRareAF @xenoludicpraxis @ereliuer_eteer @RiversHaveWings @KatanHya @ryunuck @karan4d @elder_plinius @zoink @irl_danB @latenkraft @kartographien @cajundiscordian @gwern @nostalgebraist @LericDax @chloe21e8 @SoC_trilogy @goodside @YaBoyFathoM @KennethFolk @DmitriVanDuine @arturot @voooooogel ooo if you're interested in deception have you experimented with using control vectors for worrisome things a model might hide to gauge how close the model is coming to thinking them at each layer?