post by Stuart Armstrong 286 days ago | discuss

A putative new idea for AI control; index here.

I’ve never really understood acausal trade. So in a short series of posts, I’ll attempt to analyse the concept sufficiently that I can grasp it - and hopefully so others can grasp it as well.

The simplest model

There are $$N$$ different rooms, with potential agents in them. The probability of the agents existing is a distribution $$Q$$, with marginal probabilities $$q_i$$, representing the probability that agent $$A_i$$ exists in room $$i$$. That agent has a utility $$u_i$$, which they are motivated to maximise.

The agents will never meet, never interact in any way, won’t even be sure of each other’s existence, may not known $$N$$, and may have uncertainty over the values of the other $$u_j$$’s. Each agent only acts in their own room. They may choose to diminish $$u_i$$ to increase one or more other $$u_j$$ with $$i\neq j$$; this is what allows the possibility of trade.

Infinities, utility weights, negotiations, trade before existence

There are a number of things I won’t be considering here. First of all, infinities. In reality, acausal trade would happen in the real universe, which is likely infinite. It’s not clear at all how to rank infinitely many causally disconnected world-pieces. So I’ll avoid that entirely, assuming $$N$$ is finite (though possibly large).

There’s also the thorny issue of how to weigh and compare different utility functions, and/or the process of negotiation about how to divide the gains from trade.

I’ll ignore all these issues, and see the $$u_i$$ as functions from states of the world to real numbers: individual representatives of utility functions, not equivalence classes of equivalence functions. And the bargaining will be a straight one for one increase and decrease: a fair deal is one where $$u_i$$ and $$u_j$$ get the same benefit – as measured by $$u_i$$ and $$u_j$$.

I’ll also ignore the possibility of trade before existence, or Rawlsian veils of ignorance. If you are a $$u_i$$ maximiser, but you could have been a $$u_j$$ maximiser if things had been different, then you have no responsibility to increase $$u_j$$. Similarly, if there are $$u_j$$ maximisers out there, then you have no responsibility to maximiser $$u_j$$ without getting any $$u_i$$ increases out of that. See this post for more on that.

Changing that last assumption could radically alter the nature of acausal trade - potentially reducing it to simply maximising a universal prior utility function. See this post for more on that behaviour.

NEW DISCUSSION POSTS

[Delegative Reinforcement
 by Vadim Kosoy on Stable Pointers to Value II: Environmental Goals | 1 like

Intermediate update: The
 by Alex Appel on Further Progress on a Bayesian Version of Logical ... | 0 likes

Since Briggs [1] shows that
 by 258 on In memoryless Cartesian environments, every UDT po... | 2 likes

This doesn't quite work. The
 by Nisan Stiennon on Logical counterfactuals and differential privacy | 0 likes

I at first didn't understand
 by Sam Eisenstat on An Untrollable Mathematician | 1 like

This is somewhat related to
 by Vadim Kosoy on The set of Logical Inductors is not Convex | 0 likes

This uses logical inductors
 by Abram Demski on The set of Logical Inductors is not Convex | 0 likes

Nice writeup. Is one-boxing
 by Tom Everitt on Smoking Lesion Steelman II | 0 likes

Hi Alex! The definition of
 by Vadim Kosoy on Delegative Inverse Reinforcement Learning | 0 likes

A summary that might be
 by Alex Appel on Delegative Inverse Reinforcement Learning | 1 like

I don't believe that
 by Alex Appel on Delegative Inverse Reinforcement Learning | 0 likes

This is exactly the sort of
 by Stuart Armstrong on Being legible to other agents by committing to usi... | 0 likes

When considering an embedder
 by Jack Gallagher on Where does ADT Go Wrong? | 0 likes

The differences between this
 by Abram Demski on Policy Selection Solves Most Problems | 1 like

Looking "at the very
 by Abram Demski on Policy Selection Solves Most Problems | 0 likes