But, the gist of your post seems to be: "Since coming up with UDT, we ran into these problems, made no progress, and are apparently at a dead end. Therefore, UDT might have been the wrong turn entirely."

This is a bit stronger than how I would phrase it, but basically yes.

On the other hand, my view is: Since coming up with those problems, we made a lot of progress on agent theory within the LTA

I tend to be pretty skeptical of new ideas. (This backfired spectacularly once, when I didn't pay much attention to Satoshi when he contacted me about Bitcoin, but I think in general has served me well.) My experience with philosophical questions is that even when some approach looks a stone's throw away from a final solution to some problem, a bunch of new problems pop up and show that we're still quite far away. With an approach that is still as early as yours, I just think there's quite a good chance it doesn't work out in the end, or gets stuck somewhere on a hard problem. (Also some people who have digged into the details don't seem as optimistic that it is the right approach.) So I'm reluctant to decrease my probability of "UDT was a wrong turn" too much based on it.

The rest of your discussion about 2TDT-1CDT seems plausible to me, although of course depends on whether the math works out, doing something about monotonicity, and also a solution to the problem of how to choose one's IBH prior. (If the solution was something like "it's subjective/arbitrary" that would be pretty unsatisfying from my perspective.)

Do you think part of it might be that even people with graduate philosophy educations are too prone to being wedded to their own ideas, or don't like to poke holes at them as much as they should? Because part of what contributes to my wanting to go more meta is being dissatisfied with my own object-level solutions and finding more and more open problems that I don't know how to solve. I haven't read much academic philosophy literature, but did read some anthropic reasoning and decision theory literature earlier, and the impression I got is that most of the authors weren't trying that hard to poke holes in their own ideas.

I don't understand your ideas in detail (am interested but don't have the time/ability/inclination to dig into the mathematical details), but from the informal writeups/reviews/critiques I've seen of your overall approach, as well as my sense from reading this comment of how far away you are from a full solution to the problems I listed in the OP, I'm still comfortable sticking with "most are wide open". :)

On the object level, maybe we can just focus on Problem 4 for now. What do you think actually happens in a 2IBH-1CDT game? Presumably CDT still plays D, and what do the IBH agents do? And how does that imply that the puzzle is resolved?

As a reminder, the puzzle I see is that this problem shows that a CDT agent doesn't necessarily want to become more UDT-like, and for seemingly good reason, so on what basis can we say that UDT is a clear advancement in decision theory? If CDT agents similarly don't want to become more IBH-like, isn't there the same puzzle? (Or do they?) This seems different from the playing chicken with a rock example, because a rock is not a decision theory so that example doesn't seem to offer the same puzzle.

ETA: Oh, I think you're saying that the CDT agent could turn into a IBH agent but with a different prior from the other IBH agents, that ends up allowing it to still play D while the other two still play C, so it's not made worse off by switching to IBH. Can you walk this through in more detail? How does the CDT agent choose what prior to use when switching to IBH, and how do the different priors actual imply a CCD outcome in the end?

Even items 1, 3, 4, and 6 are covered by your research agenda? If so, can you quickly sketch what you expect the solutions to look like?

@jessicata @Connor Leahy @Domenic @Daniel Kokotajlo @romeostevensit @Vanessa Kosoy @cousin_it @ShardPhoenix @Mitchell_Porter @Lukas_Gloor (and others, apparently I can only notify 10 people by mentioning them in a comment)

Sorry if I'm late in responding to your comments. This post has gotten more attention and replies than I expected, in many different directions, and it will probably take a while for me to process and reply to them all. (In the meantime, I'd love to see more people discuss each other's ideas here.)

Do you have any examples that could illustrate your theory?

It doesn't seem to fit my own experience. I became interested in Bayesian probability, universal prior, Tegmark multiverse, and anthropic reasoning during college, and started thinking about decision theory and ideas that ultimately led to UDT, but what heuristics could I have been applying, learned from what "domains with feedback"?

Maybe I used a heuristic like "computer science is cool, lets try to apply it to philosophical problems" but if the heuristics are this coarse grained, it doesn't seem like the idea can explain how detailed philosophical reasoning happens, or be used to ensure AI philosophical competence?

I expect at this moment in time me building a company is going to help me deconfuse a lot of things about philosophy more than me thinking about it really hard in isolation would

Hard for me to make sense of this. What philosophical questions do you think you'll get clarity on by doing this? What are some examples of people successfully doing this in the past?

It seems plausible that there is no such thing as “correct” metaphilosophy, and humans are just making up random stuff based on our priors and environment and that’s it and there is no “right way” to do philosophy, similar to how there are no “right preferences”.

Definitely a possibility (I've entertained it myself and maybe wrote some past comments along these lines). I wish there was more people studying this possibility.

I have short timelines and think we will be dead if we don’t make very rapid progress on extremely urgent practical problems like government regulation and AI safety. Metaphilosophy falls into the unfortunate bucket of “important, but not (as) urgent” in my view.

Everyone dying isn't the worst thing that could happen. I think from a selfish perspective, I'm personally a bit more scared of surviving into a dystopia powered by ASI that is aligned in some narrow technical sense. Less sure from an altruistic/impartial perspective, but it seems at least plausible that building an aligned AI without making sure that the future human-AI civilization is "safe" is a not good thing to do.

I would say that better philosophy/arguments around questions like this is a bottleneck. One reason for my interest in metaphilosophy that I didn't mention in the OP is that studying it seems least likely to cause harm or make things worse, compared to any other AI related topics I can work on. (I started thinking this as early as 2012.) Given how much harm people have done in the name of good, maybe we should all take "first do no harm" much more seriously?

There are no good institutions, norms, groups, funding etc to do this kind of work.

Which also represents an opportunity...

It’s weird. I happen to have a very deep interest in the topic, but it costs you weirdness points to push an idea like this when you could instead be advocating more efficiently for more pragmatic work.

Is it actually that weird? Do you have any stories of trying to talk about it with someone and having that backfire on you?

Philosophy is a social/intellectual process taking place in the world. If you understand the world, you understand how philosophy proceeds.

What if I'm mainly interested in how philosophical reasoning ideally ought to work? (Similar to how decision theory studies how decision making normatively should work, not how it actually works in people.) Of course if we have little idea how real-world philosophical reasoning works, understanding that first would probably help a lot, but that's not the ultimate goal, at least not for me, for both intellectual and AI reasons.

The latter because humans do a lot of bad philosophy and often can’t recognize good philosophy. (See popularity of two-boxing among professional philosophers.) I want a theory of ideal/normative philosophical reasoning so we can build AI that improves upon human philosophy, and in a way that convinces many people (because they believe the theory is right) to trust the AI's philosophical reasoning.

This leads to a view where philosophy is one of many types of discourse/understanding that each shape each other (a non-foundationalist view). This is perhaps disappointing if you wanted ultimate foundations in some simple framework.

Sure ultimate foundations in some simple framework would be nice but I'll take whatever I can get. How would you flesh out the non-foundationalist view?

Most thought is currently not foundationalist, but perhaps a foundational re-orientation could be found by understanding the current state of non-foundational thought.

I don't understand this sentence at all. Please explain more?

If we keep stumbling into LLM type things which are competent at a surprisingly wide range of tasks, do you expect that they’ll be worse at philosophy than at other tasks?

I'm not sure but I do think it's very risky to depend on LLMs to be good at philosophy by default. Some of my thoughts on this:

  • Humans do a lot of bad philosophy and often can't recognize good philosophy. (See popularity of two-boxing among professional philosophers.) Even if a LLM has learned how to do good philosophy, how will users or AI developers know how to prompt it to elicit that capability (e.g., which philosophers to emulate)? (It's possible that even solving metaphilosophy doesn't help enough with this, if many people can't recognize the solution as correct, but there's at least a chance that the solution does look obviously correct to many people, especially if there's not already wrong solutions to compete with).
  • What if it learns how to do good philosophy during pre-training, but RLHF trains that away in favor of optimizing arguments to look good to the user.
  • What if philosophy is just intrinsically hard for ML in general (I gave an argument for why ML might have trouble learning philosophy from humans in the section Replicate the trajectory with ML? of Some Thoughts on Metaphilosophy, but I'm not sure how strong it is) or maybe it's just some specific LLM architecture that has trouble with this, and we never figure this out because the AI is good at finding arguments that look good to humans?
  • Or maybe we do figure out that AI is worse at philosophy than other tasks, after it has been built, but it's too late to do anything with that knowledge (because who is going to tell the investors that they've lost their money because we don't want to differentially decelerate philosophical progress by deploying the AI).

If it’s not misuse, the provisions in 5.1.4-5 will steer the search process away from policies that attempt to propagandize to humans.

Ok I'll quote 5.1.4-5 to make it easier for others to follow this discussion:

5.1.4. It may be that the easiest plan to find involves an unacceptable degree of power-seeking and control over irrelevant variables. Therefore, the score function should penalize divergence of the trajectory of the world state from the trajectory of the status quo (in which no powerful AI systems take any actions).

5.1.5. The incentives under 5.1.4 by default are to take control over irrelevant variables so as to ensure that they proceed as in the anticipated “status quo”. Infrabayesian uncertainty about the dynamics is the final component that removes this incentive. In particular, the infrabayesian prior can (and should) have a high degree of Knightian uncertainty about human decisions and behaviour. This makes the most effective way to limit the maximum divergence (of human trajectories from the status quo) actually not interfering.

I'm not sure how these are intended to work. How do you intend to define/implement "divergence"? How does that definition/implementation combined with "high degree of Knightian uncertainty about human decisions and behaviour" actually cause the AI to "not interfere" but also still accomplish the goals that we give it?

In order to accomplish its goals, the AI has to do lots of things that will have butterfly effects on the future, so the system has to allow it to do those things, but also not allow it to "propagandize to humans". It's just unclear to me how you intend to achieve this.

Load More