Currently reading: Timeless Decision Theory

I’ve started reading Eliezer Yudkowsky’s 120 page document Timeless Decision Theory. Hopefully I’ll have something intelligent to say about it in the near future. If you want a taste of it but don’t have time for the whole think, I recommend Luke’s Decision Theory FAQ, as well as Eliezer’s post Newcomb’s Problem and Regret of Rationality to get a sense of the motivations for TDT.

When I started reading about TDT, I was initially skeptical that Eliezer would be able to escape the kinds of problems that plague these sorts of philosophical discussions. Just appealing to a couple thought experiments (in this case, Newcomblike problems) rarely settles anything.

But as I read on, it looks like Eliezer has some interesting things to say about precommitment, and may have a good case for TDT as a superior alternative to traditional precommitment strategies. We’ll see how my perspective changes once I’ve read the whole thing.

  • ThrustVectoring

    I’ve read that 120 page paper, and it’s pretty good. As far as thinking about TDT in precommitment terms, I believe it’s equivalent to pre-committing to whatever the best pre-commitment strategy is. So for Newcomb’s problem, Omega models you as pre-committing to one-box (since that’s the best pre-commitment decision) and you end up one-boxing (since that’s the best pre-commitment decision).

    The key insight into TDT is that your decisions aren’t just a lever that affects the outside universe, but also a lever to affect the kind of person you are and the policies that you follow. Whenever outside agents have some insight into the kind of person you are and the policies that you follow and will follow, neglecting this part of reality can lead to sub-optimal outcomes.

  • Pingback: yellow october