Eltragdors

[Eltragdors @eltragdors]

@dhadfieldmenell @jttiehen Which suggests to me that your disagreement comes down to wondering whether alignment isn't in reality easier than they imagine?

[Eltragdors @eltragdors]

@dhadfieldmenell @jttiehen And I would imagine that the ability to cooperate effectively across a large distribution of positive sum games would increase with an increasing amount of the "decision-theoretic intelligence", although the same would happen for other goals.

[Eltragdors @eltragdors]

@ESYudkowsky @eshear Furthermore, it seems plausible that this thing that is being modeled could also be inconsistent, so in distilling it down into a coherent utility function can be an ongoing process that might only get better with more computation but never quite converge in finite time.

[Eltragdors @eltragdors]

@ESYudkowsky @eshear what if the real utility function is a black box which is too complex to describe, and can only be queried in limited ways? Then perhaps it might be necessary to 'model it' just to handle the complexity?

[Eltragdors @eltragdors]

@MarcoMatthies @francoisfleuret How are francois' transformations above related to the max plus semiring precisely?

[Eltragdors @eltragdors]

@SuryaGanguli I hope there will there be a video posted online...?

[Eltragdors @eltragdors]

@francoisfleuret I wondered for a bit if I'm starting to lose my mind ... ;)... *uncomfortable smile*

[Eltragdors @eltragdors]

@nickcammarata What's the overall template or structure for this master doc that serves all these various goals as a single source of dynamic state that's easy to maintain and update?

[Eltragdors @eltragdors]

@docmilanfar @danivpv Referring to the equation for x tilde in tweet #3, the svd expansion, etc

[Eltragdors @eltragdors]

@docmilanfar Seems like a small typo -- the product with y is missing in the RHS (and same below)

[Eltragdors @eltragdors]

@nanjiang_cs @Thejakeyboy @ccanonne_ Pointwise --> matrix element wise

[Eltragdors @eltragdors]

@nanjiang_cs @Thejakeyboy @ccanonne_ Example I gave, (p, p, 1-2p) fails pointwise.

But pointwise I think can also be made true, but needs some simple additional assumption. For example z_i sampled independently across I and then x_i defined proportional to e^z_i.

[Eltragdors @eltragdors]

@nanjiang_cs @Thejakeyboy @ccanonne_ Suppose x = (p, p, 1-2p) where p is uniform on (0, 1). That's a random stochastic vector. Your claim seems obviously false.

[Eltragdors @eltragdors]

@francoisfleuret Meant without any *flips*

[Eltragdors @eltragdors]

@francoisfleuret Without that though, it seems like we could make at least one of U or V be a pure rotation without any but requiring this from both of them needs to relax the singular values to be negative as needed....?