Expected Utility Maximization is Not Enough
Consider a homomorphically encrypted computation running somewhere in the cloud. The computations correspond to running an AGI. Now from the outside, you can still model the AGI based on how it behaves, as an expected utility maximizer, if you have a lot of observational data about the AGI (or at least let's take this as a reasonable assumption).
No matter how closely you look at the computations, you will not be able to figure out how to change these computations in order to make the AGI aligned if it was not aligned already (Also, let's assume that you are some sort of Cartesian agent, otherwise you would probably already be dead if you were running these kinds of computations).
So, my claim is not that modeling a system as an expected utility maximizer can't be useful. Instead, I claim that this model is incomplete. At least with regard to the task of computing an update to the system, such that when we apply this update to the system, it would become aligned.
Of course, you can model any system, as an expected utility maximizer. But just because I can use the "high level" conceptual model of expected utility maximization, to model the behavior of a system very well. But behavior is not the only thing that we care about, we actually care about being able to understand the internal workings of the system, such that it becomes much easier to think about how to align the system.
So the following seems to be beside the point unless I am <missing/misunderstanding> something:
These two claims should probably not both be true! If any system can be modeled as maximizing a utility function, and it is possible to build a corrigible system, then naively the corrigible system can be modeled as maximizing a utility function.
Maybe I have missed the fact that the claim you listed says that expected utility maximization is not very useful. And I'm saying it can be useful, it might just not be sufficient at all to actually align a particular AGI system. Even if you can do it arbitrarily well.
Right now I am trying to better understand future AI systems, by first thinking about what sort of abilities I expect every system of high cognitive power will have, and second, trying to find a concrete practical implementation of this ability. One ability is building a model of the world, that has certain desiderata. For example, if we have multiple agents in the world, then we can factor the world, such that we can build just one model of the agent, and point to this model in our description of the world two times. This is something that Solomonoff induction can also do. I am interested in constraining the world model, such that we always get out a world model that has a similar structure, such that the world model becomes more interpretable. I.e. I try to find a way for building a world model, where we mainly need to understand the world model's content, as it is easy to understand how the content is organized.
Many people match "pivotal act" to "deploy AGI to take over the world", and ignore the underlying problem of preventing others from deploying misaligned AGI.
I have talked to two high-profile alignment/alignment-adjacent people who actively dislike pivotal acts.
I think both have contorted notions of what a pivotal act is about. They focused on how dangerous it would be to let a powerful AI system loose on the world.
However, a pivotal act is about this. So an act that ensures that misaligned AGI will not be built is a pivotal act. Many such acts might look like taking over the world. But this is not a core feature of a pivotal act. If I could prevent all people from deploying misaligned AGI, by eating 10 bananas in sixty seconds, then that would count as a pivotal act!
The two researchers were not talking about how to prevent misaligned AGI from being built at all. So I worry that they are ignoring this problem in their solution proposals. It seems "pivotal act" has become a term with bad connotations. When hearing "pivotal act", these people pattern match to "deploy AGI to take over the world", and ignore the underlying problem of preventing others from deploying misaligned AGI.
I expect there are a lot more people who fall into this trap. One of the people was giving a talk and this came up briefly. Other people seemed to be on board with what was said. At least nobody objected, except me.
See also Raemon's related post.
Solomonoff induction does not talk about how to make optimal tradeoffs in the programs that serve as the hypothesis.
Imagine you want to describe a part of the world that contains a gun. Solomonoff induction would converge on finding the program that perfectly predicts all the possible observations. So this program would be able to predict what sort of observations would I make after I stuff a banana into the muzzle and fire it. But knowing how the banana was splattered around is not the most useful fact about the gun. It is more useful to know that a gun can be used to kill humans and animals. So if you want to store your world model in only n bits of memory, you need to decide which information to put in. And this matters because some information is much more useful than others. So how can we find the world model that gives you the most power over the world, i.e. letting you reach the greatest number of states? Humans have the ability to judge the usefulness of information. You can ask yourself, what sort of knowledge would be most useful for you to learn? Or, What knowledge would be most bad to forget?
It seems potentially important to compare this to GPT4o. In my experience when asking GPT4 for research papers on particular subjects it seemed to make up non-existent research papers (at least I didn't find them after multiple minutes of searching the web). I don't have any precise statistics on this.