Attainable Utility Preservation: Concepts

Summary for the Alignment Newsletter:

In <@part 1@>(@Reframing Impact - Part 1@) of this sequence, we saw that an event is _impactful_ if it _changes our ability to get what we want_. This part takes this understanding and applies it to AI alignment.

In the real world, there are many events that cause _objective_ negative impacts: they reduce your ability to pursue nearly any goal. An asteroid impact that destroys the Earth is going to be pretty bad for you, whether you want to promote human flourishing or to make paperclips. Conversely, there are many plans that produce objective positive impacts: for many potential goals, it's probably a good idea to earn a bunch of money, or to learn a lot about the world, or to command a perfectly loyal army. This is particularly exacerbated when the environment contains multiple agents: for goals that benefit from having more resources, it is objectively bad for you if a different agent seizes your resources, and objectively good for you if you seize other agents' resources.

Based on this intuitive (but certainly not ironclad) argument, we get the **Catastrophic Convergence Conjecture (CCC)**: "Unaligned goals tend to have catastrophe-inducing optimal policies because of power-seeking incentives".

Let's now consider a _conceptual_ version of <@Attainable Utility Preservation (AUP)@>(@Towards a New Impact Measure@): the agent optimizes a primary (possibly unaligned) goal, but is penalized for changing its "power" (in the intuitive sense). Intuitively, such an agent no longer has power-seeking incentives, and so (by the contrapositive of the CCC) it will not have a catastrophe-inducing optimal policy -- exactly what we want! This conceptual version of AUP also avoids thorny problems such as ontology identification and butterfly effects, because the agent need only reason about its own beliefs, rather than having to reason directly about the external world.

Opinion:

This was my favorite part of the sequence, as it explains the conceptual case for AUP clearly and concisely. I especially liked the CCC: I believe that we should be primarily aiming to prevent an AI system "intentionally" causing catastrophe, while not attempting to guarantee an absence of "accidental" mistakes (<@1@>(@Clarifying "AI Alignment"@), <@2@>(@Techniques for optimizing worst-case performance@)), and the CCC is one way of cashing out this intuition. It's a more crisp version of the idea that convergent instrumental subgoals are in some sense the "source" of AI accident risk, and if we can avoid instrumental subgoals we will probably have solved AI safety.

Depending on the agent implementation you may find that it is demotivated to achieve any useful outcome if they are power limited. Half-assing things seems pointless and futile, they aren't sane actions in the world. E.g. trying to put out a fire when all you have is a squirt gun.

I liked this post, and look forward to the next one.

More specific, and critical commentary (It seems it is easier to notice surprise than agreement):

(With embedded footnotes)

1.

If the CCC is right, then if power gain is disincentivised, the agent isn't incentivised to overfit and disrupt our AU landscape.

(The CCC didn't make reference to overfitting.)

Premise:

If A is true then B will be true.

Conclusion:

If A is false B will be false.

The conclusion doesn't follow from the premise.

2.

Without even knowing who we are or what we want, the agent's actions preserve our attainable utilities.

Note that preserving our attainable utilities isn't a good thing, it's just not a bad thing.

Issues: Attainable utilities indefinitely 'preserved' are wasted.

Possible issues: If an AI just happened to discovered a cure for cancer, we'd probably want to know the cure. But if an AI didn't know what we wanted, and just focused on preserving utility*, then (perhaps as a side effect of considering both that we might want to know the cure, and might not want to know the cure) it might not tell us because that preserves utility. (The AI might operate on a framework that distinguishes between action and inaction, in a way that means it doesn't do thing that might be bad, at the cost of not doing things that might be good.)

*If we are going to calculate something and a reliable source (which has already done the calculation) tells us the result, we can save on energy (and preserve resources that can be converted into utility) by not doing the calculation. In theory this could include not only arithmetic, but simulations of different drugs or cancer treatments to come up with better options.

3.

We can tell it:

Is this a metaphor for making an 'agent' with that goal, or actually creating an agent that we can give different commands to and switch out/modify/add to its goals? (Why ask it to 'make paperclips' if that's dangerous, when we can ask it to 'make 100 paperclips'?)

4.

Narrowly improve paperclip production efficiency <- This is the kind of policy AUP_conceptual is designed to encourage and allow. We don't know if this is the optimal policy, but by CCC, the optimal policy won't be catastrophic.

Addressed in 1.

5.

Imagine I take over a bunch of forever inaccessible stars and jumble them up. This is a huge change in state, but it doesn't matter to us.

It does a little bit.

It means we can't observe them for astronomical purposes. But this isn't the same as losing a telescope looking at them - it's (probably) permanent, and maybe we learn something different from it. We learn that stars can be jumbled up. This may have physics/stellar engineering consequences, etc.

6.

AUP_conceptual solves this "locality" problem by regularizing the agent's impact on the nearby AU landscape.

Nearby from its perspective? (From a practical standpoint, if you're close to an airport you're close to a lot of places on earth, that you aren't from a 'space' perspective.)

7.

For past-impact measures, it's not clear that their conceptual thrusts are well-aimed, even if we could formalize everything correctly. Past approaches focus either on minimizing physical change to some aspect of the world or on maintaining ability to reach many world states.

If there's a limited amount of energy, then using energy limits ability to reach many world states - perhaps in a different sense than above. If there's a machine that can turn all pebbles into something else (obsidian, precious stones, etc.) but it takes a lot of energy, then using up energy limits the number of times it can be used. (This might seem quantifiable, moving the world* from containing 101 units of energy -> 99 units an effect on how many times the machine can be used if it requires 100, or 10 units to use. But this isn't robust against random factors decreasing energy (or decreasing it), or future improvements in energy efficiency of the machine - if the cost is brought down to 1 unit of energy, then using up 2 units prevents it from being used twice.

*Properly formalizing this should take a lot of other things into account, like 'distant' and notions of inaccessible regions of space, etc.

Also the agent might be concerned with flows rather than actions.* We have an intuitive notion that 'building factories increases power', but what about redirecting a river/stream/etc. with dams or digging new paths for water to flow? What does the agent do if it unexpectedly gains power by some means, or realizes its paperclip machines can be used to move strawberries/make a copy itself which is weaker but less constrained? Can the agent make a machine that makes paperclips/make making paperclips easier?

*As a consequence of this being a more effective approach - it makes certain improvements obvious. If you have a really long commute to work, you might wish you lived closer to your work. (You might also be aware that houses closer to your work are more expensive, but humans are good at picking up on this kind of low hanging fruit. A capable agent that thinks about process seeing 'opportunities to gain power' is of some general concern. In this case because an agent that tries to minimize reducing/affecting** other agents attainable utility, without knowing/needing to know about other agents is somewhat counterintuitive.

**It's not clear if increasing shows up on the AUP map, or how that's handled.

8.

Therefore, I consider AUP to conceptually be a solution to impact measurement.

Wait! Let's not get ahead of ourselves! I don't think we've fully bridged the concept/execution gap.

However for AUP, it seems possible - more on that later.

I appreciate this distinction being made. A post that explains the intuitions behind an approach is very useful, and my questions about the approach may largely relate to implementation details.

9.

AUP aims to prevent catastrophes by stopping bad agents from gaining power to do bad things, but it symmetrically impedes otherwise-good agents.

A number of my comments above were anticipated then.

The subagent problem remains: How do you prevent it from getting someone else to catastrophically maximize paperclips and leave it at its power level?

I was initially writing a comment about how AUP${}_{\text{conceptual}}$ doesn't seem to work in every case because there are actions that are catastrophic without raising its power (such as killing someone), but then I checked the post again and realized that it disincentivizes changes of power in both directions. This rules out the failure modes I had in mind. (It wouldn't press a button that blows up the earth...)

It does seem that AUP${}_{\text{conceptual}}$ will make it so an agent doesn't want to be shut off, though. If it's shut off, its power goes way down (to zero if it won't be turned on again). This might be fine, but it contradicts the utility indifference approach. And it feels dangerous – it seems like we would need an assurance like "AUP${}_{\text{conceptual}}$ will always prevent an agent from gaining enough power to resist being switched off"

For reference and ease of quoting, this comment is a text only version of the post above. (It starts at "Text:" below.) I am not the OP.

Formatting:

It's not clear how to duplicate the color effect* or cross words out**, so that hasn't been done. Instead crossed out words are followed by "? (No.)", and here's a list of some words by color to refresh the color/concept relation:

Blue words:

Power/impact/penalty/importance/respect/conservative/catastrophic/distance measure/impact measurement

Purple words:

incentives/actions/(reward)/expected utility/complicated human value/tasks

Text:

Last time on reframing impact:

(CCC)

Catastrophic Convergence Conjecture:

Unaligned goals tend to have catastrophe-inducing optimal policies because of power-seeking incentives

If the CCC is right, then if power gain is disincentivised, the agent isn't incentivised to overfit and disrupt our AU landscape.

Without even knowing who we are or what we want, the agent's actions preserve our attainable utilities.

We can tell it:

Make paperclips

or

Put that strawberry on the plate

or

Paint the car pink

...

but don't gain power.

This approach is called Attainable Utility preservation

We're focusing on concepts in this post. For now, imagine an agent receiving a reward for a primary task minus a scaled penalty for how much it's actions change its power (in the intuitive sense). This is AUP_conceptual, not any formalization you may be familiar with.

What might a paperclip-manufacturing AUP_conceptual agent do?

Build lots of factories? (No.)

Copy itself? (No.)

Nothing? (No.)

Narrowly improve paperclip production efficiency <- This is the kind of policy AUP_conceptual is designed to encourage and allow. We don't know if this is the optimal policy, but by CCC, the optimal policy won't be catastrophic.

AUP_conceptual dissolves thorny issues in impact measurement.

Is the agent's ontology reasonable?

Who cares.

Instead of regulating its complex physical effects on the outside world,

the agent is looking inwards at itself and its own abilities.

How do we ensure the impact penalty isn't dominated by distant state changes?

Imagine I take over a bunch of forever inaccessible stars and jumble them up. This is a huge change in state, but it doesn't matter to us.

AUP_conceptual solves this "locality" problem by regularizing the agent's impact on the nearby AU landscape.

What about butterfly effects?

How can the agent possibly determine which effects its responsible for?

Forget about it.

AUP_conceptual agents are respectful and conservative with respect to the local AUP landscape without needing to assume anything about its structure or the agents in it.

How can an idea go wrong?

There can be a gap between what we want and the concept, and then a gap between the concept and the execution.

For past-impact measures, it's not clear that their conceptual thrusts are well-aimed, even if we could formalize everything correctly. Past approaches focus either on minimizing physical change to some aspect of the world or on maintaining ability to reach many world states.

The hope is that in order for the agent to cause a large impact on us it has to snap a tripwire.

The problem is... well it's not clear how we could possibly know whether the agent can still find a catastrophic policy; in a sense the agent is still trying to sneak by the restrictions and gain power over us. An agent maximizing expected utility while actually minimally changing still probably leads to catastrophe.

That doesn't seem to be the case for AUP_conceptual.

Assuming CCC, an agent which doesn't gain much power, doesn't cause catastrophes. This has no dependency on complicated human value, and most realistic tasks should have reasonable, high-reward policies not gaining undue power.

So AUP_conceptual meets our desiderata:

The distance measure should:

1) Be easy to specify

2) Put catastrophes far away.

3) Put reasonable plans nearby

Therefore, I consider AUP to conceptually be a solution to impact measurement.

Wait! Let's not get ahead of ourselves! I don't think we've fully bridged the concept/execution gap.

However for AUP, it seems possible - more on that later.

Just trying to think this through ... at the risk of proving I haven't carefully read all your posts ... :-)

I program my AI to invent a better solar cell. So it starts by reading a materials science textbook. OK, now it knows materials science ... it didn't before ... Is that a disallowed AU increase? (As the saying goes, "knowledge is power"...?)