02:19:20	Jasmine:	how would this work for closed models, where you can't get the activations?
02:26:45	Jasmine:	ooh, Koyena, maybe the o-reasoning GPT models are a good example
02:27:12	Koyena Pal:	True!!!!
02:48:26	Jasmine:	have you looked at instruction hierarchy?
02:50:41	Jasmine:	is that similar to deliberative alignment?
02:50:51	Jasmine:	or instruction hierarchy training?