That link goes to your previous comment instead of the Michael Blume example. Perhaps you mean his othello role?

Fixed, thanks.

I must admit that I'm somewhat confused, here. I make no claims that the described practices are safe, and in fact I make a number of explicit disclaimers stating that they are not safe. It is dangerous to be half a rationalist, and I readily admit that these tools will bite you hard if you misuse them. This, in fact, is something that I assumed was captured by the "Dark Arts" label. I continue to be baffled by how some people complain about the label, others complain about the danger, and still others complain about both at once.

I completely agree that you shouldn't go around compartmentalizing at every opportunity, and that you should have a deep understanding of the problem at hand before doing any So8res!DarkArts. Prefer other methods, where possible.

I get the impression that my mental model of when self-deception is optimal differs from your own. I don't currently have time to try to converge these models right now, but suffice to say that your arguments are not addressing the divergence point.

Regardless, I think we can both agree that self-deception is optimal sometimes, under controlled scenarios in which the agent has a strong understanding of the situation. I think we also agree that such things are dangerous and should be approached with care. All this, I tried to capture with the "Dark Arts" label -- I am sorry if that did not make it across the communication gap.