This is a lot more blatantly capabilities than your last two capabilities posts. Consider deleting this and not publishing or sharing any further work like it.
Capability elicitation is also a “make thing more capable” technique, which I would prefer not get published when developed. Note that I don’t feel that I have enough insight into your research plans to suggest not doing the work, just not sharing it indiscriminately.
This is a lot more blatantly capabilities than your last two capabilities posts. Consider deleting this and not publishing or sharing any further work like it.
FWIW, I did consider whether this work would non-trivially advance AI progress, in particular advance the scarier parts of AI progress.
I think the most concerning aspect is hype and this seemed not-that-bad to me.
I’m curious what you’re refering to by “my last two capabilities posts”.
Capability elicitation is also a “make thing more capable” technique, which I would prefer not get published when developed. Note that I don’t feel that I have enough insight into your research plans to suggest not doing the work, just not sharing it indiscriminately.
Ah, I see. In particular, my prior work is about making elicitation robust to adversaries which feels pretty importantlly different to me.