Seems to be missing old stuff by Stuart Armstrong (?)

Some further half-baked thoughts:

One thing that is still not clear (both in reality, and per this article) is the extent to which we should view a model as having a coherent persona/goal. 

This is a tiny bit related to the question of whether models are strictly simulators, or if some personas / optimization daemons "take on a life of their own", and e.g.:
1) bias the model towards simulating them and/or
2) influence the behavior of other personas

It seems like these things do in fact happen, and the implications are that the "simulator" viewpoint becomes less accurate over time.

Why?

• There needs to be some prior distribution over personas.
• Empirically, post-training seems to concentrate the prior over personas on some default persona (although it's unclear what to make of this).
• It seems like alignment faking, exploration/gradient hacking, and implicit meta-learning type effects are likely to be sensitive to goals of whichever personas are active and lead the model to preferentially update in a way that serves the goals of these personas.
• To the extent that different personas are represented in the prior (or conjured during post-training), the ones that more aggressively use such strategies to influence training updates would gain relatively more influence. 

This was an interesting article, however, taking a cynical/critical lens, it seems like "the void" is just... underspecification causing an inner alignment failure?  The post has this to say on the topic of inner alignment:

And one might notice, too, that the threat model – about inhuman, spontaneously generated, secret AI goals – predates Claude by a long shot. In 2016 there was an odd fad in the SF rationalist community about stuff kind of like this, under the name “optimization demons.” Then that discourse got sort of refurbished, and renamed to “inner alignment” vs. “outer alignment.”

This is in the context of mocking these concerns as delusional self-fulfilling prophecies.

I guess the devil is in the details, and the point of the post is more to dispute the framing and ontology of the safety community, which I found useful.  But it does seem weirdly uncharitable in how it does so.

First, RE the role of "solving alignment" in this discussion, I just want to note that:
1) I disagree that alignment solves gradual disempowerment problems.
2) Even if it would that does not imply that gradual disempowerment problems aren't important (since we can't assume alignment will be solved).
3) I'm not sure what you mean by "alignment is solved"; I'm taking it to mean "AI systems can be trivially intent aligned".  Such a system may still say things like "Well, I can build you a successor that I think has only a 90% chance of being aligned, but will make you win (e.g. survive) if it is aligned.  Is that what you want?" and people can respond with "yes" -- this is the sort of thing that probably still happens IMO.
4) Alternatively, you might say we're in the "alignment basin" -- I'm not sure what that means, precisely, but I would operationalize it as something like "the AI system is playing a roughly optimal CIRL game".  It's unclear how good of performance that can yield in practice (e.g. it can't actually be optimal due to compute limitations), but I suspect it still leaves significant room for fuck-ups.
5) I'm more interested in the case where alignment is not "perfectly" "solved", and so there are simply clear and obvious opportunities to trade-off safety and performance; I think this is much more realistic to consider.
6) I expect such trade-off opportunities to persist when it comes to assurance (even if alignment is solved), since I expect high-quality assurance to be extremely costly.  And it is irresponsible (because it's subjectively risky) to trust a perfectly aligned AI system absent strong assurances.  But of course, people who are willing to YOLO it and just say "seems aligned, let's ship" will win.  This is also part of the problem...


My main response, at a high level:
Consider a simple model:

• We have 2 human/AI teams in competition with each other, A and B.
• A and B both start out with the humans in charge, and then decide whether the humans should stay in charge for the next week.
• Whichever group has more power at the end of the week survives.
• The humans in A ask their AI to make A as powerful as possible at the end of the week.
• The humans in B ask their AI to make B as powerful as possible at the end of the week, subject to the constraint that the humans in B are sure to stay in charge.

I predict that group A survives, but the humans are no longer in power.  I think this illustrates the basic dynamic.  EtA: Do you understand what I'm getting at?  Can you explain what you think it wrong with thinking of it this way?

 

Responding to some particular points below:
 

Sure, but these things don't result in non-human entities obtaining power right? 

Yes, they do; they result in beaurocracies and automated decision-making systems obtaining power.  People were already having to implement and interact with stupid automated decision-making systems before AI came along.

Like usually these are somewhat negative sum, but mostly just involve inefficient transfer of power. I don't see why these mechanisms would on net transfer power from human control of resources to some other control of resources in the long run. To consider the most extreme case, why would these mechanisms result in humans or human appointed successors not having control of what compute is doing in the long run?

My main claim was not that these are mechanisms of human disempowerment (although I think they are), but rather that they are indicators of the overall low level of functionality of the world.  




 

I think we disagree about:
1) The level of "functionality" of the current world/institutions.
2) How strong and decisive competitive pressures are and will be in determining outcomes.

I view the world today as highly dysfunctional in many ways: corruption, coordination failures, preference falsification, coercion, inequality, etc. are rampant.  This state of affairs both causes many bad outcomes and many aspects are self-reinforcing.  I don't expect AI to fix these problems; I expect it to exacerbate them.

I do believe it has the potential to fix them, however, I think the use of AI for such pro-social ends is not going to be sufficiently incentivized, especially on short time-scales (e.g. a few years), and we will instead see a race-to-the-bottom that encourages highly reckless, negligent, short-sighted, selfish decisions around AI development, deployment, and use.  The current AI arms race is a great example -- Companies and nations all view it as more important that they be the ones to develop ASI than to do it carefully or put effort into cooperation/coordination.

Given these views:
1) Asking AI for advice instead of letting it take decisions directly seems unrealistically uncompetitive.  When we can plausibly simulate human meetings in seconds it will be organizational suicide to take hours-to-weeks to let the humans make an informed and thoughtful decision. 
2) The idea that decision-makers who "think a goverance structure will yield total human disempowerment" will "do something else" also seems quite implausible.  Such decision-makers will likely struggle to retain power.  Decision-makers who prioritize their own "power" (and feel empowered even as they hand off increasing decision-making to AI) and their immediate political survival above all else will be empowered.  

Another features of the future which seems likely and can already be witnessed beginning is the gradual emergence and ascendance of pro-AI-takeover and pro-arms-race ideologies, which endorse the more competitive moves of rapidly handing off power to AI systems in insufficiently cooperative ways.

This thought experiment is described in ARCHES FYI.  https://acritch.com/papers/arches.pdf

I think it's a bit sad that this comment is being so well-received -- it's just some opinions without arguments from someone who hasn't read the paper in detail.  

1. There are 2 senses in which I agree that we don't need full on "capital V value alignment":
   1. We can build things that aren't utility maximizers (e.g. consider the humble MNIST classifier)
   2. There are some utility functions that aren't quite right, but are still safe enough to optimize in practice (e.g. see "Value Alignment Verification", but see also, e.g. "Defining and Characterizing Reward Hacking" for negative results)
2. But also:
   1. Some amount of alignment is probably necessary in order to build safe agenty things (the more agenty, the higher the bar for alignment, since you start to increasingly encounter perverse instatiation-type concerns -- CAVEAT: agency is not a unidimensional quantity, cf: "Harms from Increasingly Agentic Algorithmic Systems").
   2. Note that my statement was about the relative requirements for alignment in text domains vs. real-world.  I don't really see how your arguments are relevant to this question.

Concretely, in domains with vision, we should probably be significantly more worried that an AI system learns something more like an adversarial "hack" on it's values leading to behavior that significantly diverges from things humans would endorse.
 

OTMH, I think my concern here is less:

• "The AI's values don't generalize well outside of the text domain (e.g. to a humanoid robot)"
  
  and more:
• "The AI's values must be much more aligned in order to be safe outside the text domain"
  
  I.e. if we model an AI and a human as having fixed utility functions over the same accurate world model, then the same AI might be safe as a chatbot, but not as a robot.
  
  This would be because the richer domain / interface of the robot creates many more opportunities to "exploit" whatever discrepancies exist between AI and human values in ways that actually lead to perverse instantiation.
  
   

This comment made me reflect on what fragility of values means.

To me this point was always most salient when thinking about embodied agents, which may need to reliably recognize something like "people" in its environment (in order to instantiate human values like "try not to hurt people") even as the world changes radically with the introduction of various forms of transhumanism.

I guess it's not clear to me how much progress we make towards that with a system that can do a very good job with human values when restricted to the text domain.  Plausibly we just translate everything into text and are good to go?  It makes me wonder where we're at with adversarial robustness of vision-language models, e.g.