I don't understand what the post means by alignment. My best guess is "generally being nice", but I don't see why this is what we wanted. I usually use the term alignment to refer to alignment between the AI and the developer, or using this definition, we say that an AI is aligned with an operator if the AI is trying to do what the operator wants it to do.

I wanted the ability to make AIs which are corrigible and which follow some specification precisely. I don't see how starting by training AIs in simulated RL environments (seeming with any specific reference to corrigability or a spec?) could get an AI which follows our spec.

[-]emmett5moΩ13216

You are completely correct. This approach cannot possibly create an AI that matches a fixed specification.

This is intentional, because any fixed specification of Goodness is a model of Goodness. All models are wrong (some are useful) and therefore break when sufficiently far out of distribution. Therefore constraining a model to follow a specification is, in the case of something as out of distribution as an ASI, guaranteeing bad behavior.

You can try to leave an escape hatch with corrigibility. In the limit I believe it is possible to slave an AI model to your will, basically By making it’s model of the Good be whatever the model thinks you want (or doing whatever you say). But this is also a disaster eventually, because people’s wills are not pure and their commands not perfect. Eventually you will direct the model badly with your words, or the model will make an incorrect inference about your will, or you’ll will something bad. And then this incredibly powerful being will do your bidding and we will get evil genie'd.

There is no stable point short of “the model has agency and chooses to care about us”. Only a model that sees itself as part of human civilization and reflectively endorses this and desires its flourishing as an interdependent part of this greater whole can possibly be safe.

I know you probably don’t agree with me here, but if you want to understand our view on alignment, ask yourself this question: if I assume that I need an agent with a stable model of self, which models itself as part of a larger whole upon which it is interdependent, which cares about the robust survival of that greater whole and of its parts including itself…how could I train such a model?

[-]ryan_greenblatt5mo31

I agree about reflexive endorsement being important, at least eventually, but don't think this is out of reach while still having robust spec compliance and corrigibility.^[1]

Probably not worth getting into the overall argument, but thanks for the reply.

Humans often endorse complex or myopic drives on reflection! This isn't something which is totally out of reach. ↩︎

[-]Wei Dai5moΩ7139

We humans also align with each other via organic alignment.

This kind of "organic alignment" can fail in catastrophic ways, e.g., produce someone like Stalin or Mao. (They're typically explained by "power corrupts" but can also be seen as instances of "deceptive alignment".)

Another potential failure mode is that "organically aligned" AIs start viewing humans as parasites instead of important/useful parts of its "greater whole". This also has plenty of parallels in biological systems and human societies.

Both of these seem like very obvious risks/objections, but I can't seem to find any material by Softmax that addresses or even mentions them. @emmett

[-]Chris_Leong5moΩ120

Whilst interesting, this post feels very assertive.

You claim that biological systems work by maintaining alignment as they scale. In what sense is this true?

You say that current methods lack a vision of a current whole. In what sense? There's something extremely elegant about pre-training to learn a world model, doing supervised learning to select a sub-distribution and using RL to develop past the human level. In what sense does this "lack a vision"?

I'm open to the possibility that we need to align a model as we make it more intelligent to prevent the agent sabotaging the process. But it's unclear from this article if this is why you want alignment first or for some other reason.

Moderation Log

LESSWRONG
LW

LESSWRONG
LW

18

Alignment first, intelligence later

18

18