(Thanks! Haven’t forgotten about this, will try to respond soon.)
JesseClifton
Karma: 560
Reasons-based choice and cluelessness
Leadership change at the Center on Long-Term Risk
Thanks for this! IMO thinking about what it even means to do good under extreme uncertainty is still underrated.
I don’t see how this post addresses the concern about cluelessness, though.
My problem with the construction analogy is: Our situation is more like, whenever we place a brick we might also be knocking bricks out of other parts of the house. Or placing them in ways that preclude good building later. So we don’t know if we’re actually contributing to the construction of the house on net.
On your takeaway at the bottom, it seems to be: “if someone doing A is a necessary condition for a particular good outcome X, that’s a reason for you to do A”. Granted. But the whole problem is that I don’t know how to weigh this reason against the reasons favoring me doing not-A. Why do you think we ought to privilege the particular reason that you point to?
Center on Long-Term Risk: Annual review and fundraiser 2023
[linkpost] When does technical work to reduce AGI conflict make a difference?: Introduction
We at CLR are now using a different definition of s-risks.
New definition:
S-risks are risks of events that bring about suffering in cosmically significant amounts. By “significant”, we mean significant relative to expected future suffering.
Note that it may turn out that the amount of suffering that we can influence is dwarfed by suffering that we can’t influence. By “expectation of suffering in the future” we mean “expectation of action-relevant suffering in the future”.
I found it surprising that you wrote: …
Because to me this is exactly the heart of the asymmetry. It’s uncontroversial that creating a person with a bad life inflicts on them a serious moral wrong. Those of us who endorse the asymmetry don’t see such a moral wrong involved in not creating a happy life.
+1. I think many who have asymmetric sympathies might say that there is a strong aesthetic pull to bringing about a life like Michael’s, but that there is an overriding moral responsibility not to create intense suffering.
Very late here, but a brainstormy thought: maybe one way one could start to make a rigorous case for RDM is to suppose that there is a “true” model and prior that you would write down if you had as much time as you needed to integrate all of the relevant considerations you have access to. You would like to make decisions in a fully Bayesian way with respect to this model, but you’re computationally limited so you can’t. You can only write down a much simpler model and use that to make a decision.
We want to pick a policy which, in some sense, has low regret with respect to the Bayes-optimal policy under the true model. If we regard our simpler model as a random draw from a space of possible simplified models that we could’ve written down, then we can ask about the frequentist properties of the regret incurred by different decision rules applied to the simple models. And it may be that non-optimizing decision rules like RDM have a favorable bias-variance tradeoff, because they don’t overfit to the oversimplified model. Basically they help mitigate a certain kind of optimizer’s curse.
nil already kind of addressed this in their reply, but it seems important to keep in mind the distinction between the intensity of a stimulus and the moral value of the experience caused by the stimulus. Statements like “experiencing pain just slightly stronger than that threshold” risk conflating the two. And, indeed, if by “pain” you mean “moral disvalue” then to discuss pain as a scalar quantity begs the question against lexical views.
Sorry if this is pedantic, but in my experience this conflation often muddles discussions about lexical views.
Some Bayesian statisticians put together prior choice recommendations. I guess what they call a “weakly informative prior” is similar to your “low-information prior”.
Nice comment; I’d also like to see a top-level post.
One quibble: Several of your points risk conflating “far-future” with “existential risk reduction” and/or “AI”. But there is far-future work that is non-x-risk focused (e.g. Sentience Institute and Foundational Research Institute) and non-AI-focused (e.g. Sentience Institute) which might appeal to someone who shares some of the concerns you listed.
Distribution P is your credence. So you are saying “I am worried that my credences don’t have to do with my credence.” That doesn’t make sense. And sure we’re uncertain of whether our beliefs are accurate, but I don’t see what the problem with that is.
I’m having difficulty parsing the statement you’ve attributed to me, or mapping it what I’ve said. In any case, I think many people share the intuition that “frequentist” properties of one’s credences matter. People care about calibration training and Brier scores, for instance. It’s not immediately clear to me why it’s nonsensical to say “P is my credence, but should I trust it?”
It sounds to me like this scenario is about a difference in the variances of the respective subjective probability distributions over future stock values. The variance of a distribution of credences does not measure how “well or poorly supported by evidence” that distribution is.
My worry about statements of the form “My credences over the total future utility given intervention A are characterized by distribution P” does not have to do with the variance of the distribution P. It has to do with the fact that I do not know whether I should trust the procedures that generated P to track reality.
whether you are Bayesian or not, it means that the estimate is robust to unknown information
I’m having difficulty understanding what it means for a subjective probability to be robust to unknown information. Could you clarify?
subjective expected utility theory is perfectly capable of encompassing whether your beliefs are grounded in good models.
Could you give an example where two Bayesians have the same subjective probabilities, but SEUT tells us that one subjective probability is better than the other due to better robustness / resulting from a better model / etc.?
For a Bayesian, there is no sense in which subjective probabilities are well or poorly supported by the evidence, unless you just mean that they result from calculating the Bayesian update correctly or incorrectly.
Likewise there is no true expected utility to estimate. It is a measure of an epistemic state, not a feature of the external world.
I am saying that I would like this epistemic state to be grounded in empirical reality via good models of the world. This goes beyond subjective expected utility theory. As does what you have said about robustness and being well or poorly supported by evidence.
But that just means that people are making estimates that are insufficiently robust to unknown information and are therefore vulnerable to the optimizer’s curse.
I’m not sure what you mean. There is nothing being estimated and no concept of robustness when it comes to the notion of subjective probability in question.
I can’t speak for the author, but I don’t think the problem is the difficulty of “approximating” expected value. Indeed, in the context of subjective expected utility theory there is no “true” expected value that we are trying to approximate. There is just whatever falls out of your subjective probabilities and utilities.
I think the worry comes more from wanting subjective probabilities to come from somewhere — for instance, models of the world that have a track-record of predictive success. If your subjective probabilities are not grounded in such a model, as is arguably often the case with EAs trying to optimize complex systems or the long-run future, then it is reasonable to ask why they should carry much epistemic / decision-theoretic weight.
(People who hold this view might not find the usual Dutch book or representation theorem arguments compelling.)
Thanks for writing this. I think the problem of cluelessness has not received as much attention as it should.
I’d add that, in addition to the brute good and x-risks approaches, there are approaches which attempt to reduce the likelihood of dystopian long-run scenarios. These include suffering-focused AI safety and values-spreading. Cluelessness may still plague these approaches, but one might argue that they are more robust to both empirical and moral uncertainty.
Some reasons why animal welfare work seems better:
I put some weight on a view which says: “When doing consequentialist decision-making, we should set the net weight of the reasons we have no idea how to weigh up (e.g., long-run flowthrough effects) to zero.” This probably implies restricting attention to near-term consequences, and animal welfare interventions seem best for that. (I just made a post that discusses this approach to decision-making.)
I think this kind of view is hard to make theoretically satisfying, but it does a good enough job of capturing intuitions relative to alternatives that I currently want to give it some weight.
Non-consequentialist considerations might push towards fighting the worst ongoing atrocities / injustices, which also suggests animal-related work.