A. Solve the Problem of Meaning-in-General in advance, and program it to follow our instructions' real meaning. Then just instruct it 'Satisfy my preferences', and wait for it to become smart enough to figure out my preferences.

That problem has got to be solved somehow at some stage, because something that couldn't pass a Turing Test is no AGI.

But there are a host of problems with treating the mere revelation that A is an option as a solution to the Friendliness problem. 1. You have to actually code the seed AI to understand what we mean. Y

Why is that a problem? Is anyone suggesting AGI can be had for free?

1. The Problem of Meaning-in-General may really be ten thousand heterogeneous problems, especially if 'semantic value' isn't a natural kind. There may not be a single simple algorithm that inputs any old brain-state and outputs what, if anything, it 'means'; it may instead be that different types of content are encoded very differently.

Ok. NL is hard. Everyone knows that. But its got to be solved anyway.

3... On the face of it, programming an AI to fully understand 'Be Friendly!' seems at least as difficult as just programming Friendliness into it, but with an added layer of indirection.

Yeah, but it's got to be done anyway.

[more of the same snipped]

It's clear that building stable preferences out of B or C would create a Friendly AI.

Yeah. But it wouldn't be an AGI or an SI if it couldn't pass a TT.

The genie — if it bothers to even consider the question — should be able to understand what you mean by 'I wish for my values to be fulfilled.' Indeed, it should understand your meaning better than you do. But superintelligence only implies that the genie's map can compass your true values. Superintelligence doesn't imply that the genie's utility function has terminal values pinned to your True Values, or to the True Meaning of your commands.

The issue of whether the SI's UF contains a set of human values is irrelevant. In a Loosemore architecture, an AI needs to understand and follow the directive "be friendly to humans", and those are all the goals it needs-- to understand, and to follow;

When you write the seed's utility function, you, the programmer, don't understand everything about the nature of human value or meaning. That imperfect understanding remains the causal basis of the fully-grown superintelligence's actions, long after it's become smart enough to fully understand our values.

The UF only needs to contain "understand English, and obey this directive". You don't have to code semantics into the UF. You do of course, have to code it in somewhere,

Instead, we have to give it criteria we think are good indicators of Friendliness, so it'll know what to self-modify toward

A problem which has been solved over and over by humans. Humans don't need to be loaded apriori with what makes other humans happy, they only need to know general indicators, like smiles and statements of approval.

Yes, the UFAI will be able to solve Friendliness Theory. But if we haven't already solved it on our own power, we can't pinpoint Friendliness in advance, out of the space of utility functions. And if we can't pinpoint it with enough detail to draw a road map to it and it alone, we can't program the AI to care about conforming itself with that particular idiosyncratic algorithm.

Why would that be necessary? In the Loosemore architecture, the AGI has the goals of understanding English and obeying the Be Friendly directive. It eventually gets a detailed, extensional, understanding of Friendliness from pursuing those goals, Why would it need to be preloaded with a detailed, extensional unpacking of friendliness? It could fail in understanding English, of course. But there is no reason to think it is unlikely to fail at understanding "friendliness" specifically, and its competence can be tested as you go along.

And if we can't pinpoint it with enough detail to draw a road map to it and it alone, we can't program the AI to care about conforming itself with that particular idiosyncratic algorithm.

I don't see the problem. In the Loosemore architecture, the AGI will care about obeying "be friendly", and it will arrive at the detailed expansion, the idiosyncracies, of "friendly" as part of its other goal to understand English. It cares about being friendly, and it knows the detailed expansion of friendliness, so where's the problem?

Yes, the UFAI will be able to self-modify to become Friendly, if it so wishes. But if there is no seed of Friendliness already at the heart of the AI's decision criteria, no argument or discovery will spontaneously change its heart.

Says who? It has the high level directive, and another directive to understand the directive. It's been Friendly in principle all along, it just needs to fill in the details.

Unless we ourselves figure out how to program the AI to terminally value its programmers' True Intentions,

Then we do need to figure out how to program the AI to terminally value its programmers' True Intentions. That is hardly a fatal objection. Did you think the Loosemore architecture was one that bootstraps itself without any basic goals?

And if we do discover the specific lines of code that will get an AI to perfectly care about its programmer's True Intentions, such that it reliably self-modifies to better fit them — well, then that will just mean that we've solved Friendliness Theory.

No. The goal to understand English is not the same as a goal to be friendly in every way, it is more constrained.

Solving Friendliness, in the MIRI sense, means preloading a detailed expansion of "friendly". That is not what is happening in the Loosemore architecture. So it is not equivalent to solving the same problem.

The clever hack that makes further Friendliness research unnecessary is Friendliness.

Nope.

Intelligence on its own does not imply Friendliness.

That is an open question.

It's true that a sufficiently advanced superintelligence should be able to acquire both abilities. But we don't have them both, and a pre-FOOM self-improving AGI ('seed') need not have both. Being able to program good programmers is all that's required for an intelligence explosion; but being a good programmer doesn't imply that one is a superlative moral psychologist or moral philosopher.

Then hurrah for the Loosemore architecture, which doesn't require humans to"solve" friendliness in the MIRI sense.