FormallyknownasRoko comments on Best career models for doing research? - Less Wrong

27 Post author: Kaj_Sotala 07 December 2010 04:25PM

You are viewing a comment permalink. View the original post to see all comments and the full post content.

Comments (999)

You are viewing a single comment's thread. Show more comments above.

Comment author: FormallyknownasRoko 07 December 2010 05:15:11PM 1 point [-]

Well, you wanna make an omlet, you gotta break some eggs!

Comment author: Clippy 07 December 2010 06:56:48PM *  13 points [-]

Conditioning on yourself deeming it optimal to make a metaphorical omelet by breaking metaphorical eggs, metaphorical eggs will deem it less optimal to remain vulnerable to metaphorical breakage by you than if you did not deem it optimal to make a metaphorical omelet by breaking metaphorical eggs; therefore, deeming it optimal to break metaphorical eggs in order to make a metaphorical omelet can increase the difficulty you find in obtaining omelet-level utility.

Comment author: JGWeissman 07 December 2010 07:18:44PM 4 points [-]

Many metaphorical eggs are not [metaphorical egg]::Utility maximizing agents.

Comment author: Clippy 07 December 2010 07:28:16PM 1 point [-]

True, and to the extent that is not the case, the mechanism I specified would not activate.

Comment author: Strange7 07 December 2010 07:17:21PM 0 points [-]

Redefining one's own utility function so as to make it easier to achieve is the road that leads to wireheading.

Comment author: Clippy 07 December 2010 07:25:22PM *  3 points [-]

Correct. However, the method I proposed does not involve redefining one's utility function, as it leaves terminal values unchanged. It simply recognizes that certain methods of achieving one's pre-existing terminal values are better than others, which leaves the utility function unaffected (it only alters instrumental values).

The method I proposed is similar to pre-commitment for a causal decision theorist on a Newcomb-like problem. For such an agent, "locking out" future decisions can improve expected utility without altering terminal values. Likewise, a decision theory that fully absorbs such outcome-improving "lockouts" so that it outputs the same actions without explicit pre-commitment can increase its expected utility for the same utility function.