For what it's worth, I do specifically have power-law shaped intuitions about the value of pleasure as you arrange matter to optimize for it more and more. But I agree with you both, I didn't argue for this and it's not important to my core point. 

Sure, there are various ways to do this. Scale up ems, for example, or build superintelligence from symbolic systems with strong verifiability guarantees, for starters. 

Appreciate this comment, and very much agree. I generally think that humanity's descendents are going to saturate the stars with Dyson swarms making stuff (there's good incentives to achieve explosive growth) but I think we're (1) too quick to assume that, (2) too quick to assume we will stop being attached to inefficient earth stuff, and (3) too quick to assume the Dyson swarms will be implementing great stuff rather than, say, insentient digital slaves used to amass power or solve scientific problems.

Let's say there are three threat models here: (a) Weird Stuff Matters A Lot, (b) Attachment to Biological Organisms, (c) Disneyland With No Children (the machines aren't conscious).

I focused mainly on Weird Stuff Matters A Lot. The main reason I focused on this rather than Attachment to Biological Organisms is that I still think that computers are going to be so much more economically efficient than biology that in expectation ~75% of everything is computer. Computers are just much more useful than animals for most purposes, and it would be super crazy from most perspectives not to turn most of the stars into computers. (I wouldn't totally rule out us failing to do that, but incentives push towards it strongly.) If, in expectation, ~75% of everything is computer, then maximizing computer only makes the world better by 1/3.

I think the Disneyland With No Children threat model is much scarier. I focused on it less here because I wanted to shore up broadly appealing theoretical reasons for trajectory change, and this argument feels much more partisan. But on my partisan worldview:

• Consciousness is an indeterminate folk notion that we will reduce to computational properties. (Or something very close to this.)
• These computational properties are going to be much, much more precise and gradable than our folk notion and they won't wear folk psychological properties on their sleeves.
• As a result we're just going to have to make some choices about what stuff we think is conscious and what isn't. There's going to be a sharp borderline we're going to have to pick arbitrarily, probably based on nothing more than whimsical values.
• People will disagree about where the borderline is.
• Even if people don't disagree about the borderline, they'll disagree substantially about cardinality, i.e. how much to value different computational properties relative to others.
• Given the Power Laws of Value point, some people's choices will be a mere shadow of value from the perspective of other people's choices. 

If this "irrealist" view is right, it's extremely easy to lose out on almost all value. 

Separately, I just don't think our descendents are going to care very much about whether the computers are actually conscious, and so AI design choices are going to be orthogonal to moral value. On this different sort of orthogonality thesis, we'll lose out on most value just because our descendents will use AI for practical reasons other than moral reasons, and so their intrinsic value will be unoptimized.

So Disneyland With No Children type threat models look very credible to me. 

(I do think humans will make a lot of copies of themselves, which is decently valuable, but not if you're comparing it to the most valuable world or if you value diversity.)

You could have a more realist view where we just make a big breakthrough in cognitive science and realize that a very glowy, distinctive set of computational properties was what we were talking about all along when we talked about consciousness, and everyone would agree to that. I don't really think that's how science works, but even if you did have that view it's hard to see how the computational properties would just wear their cardinality on their sleeves. Whatever computational properties you find you can always value them differently. If you find some really natural measure of hedons in computational space you can always map hedons to moral value with different functions. (E.g. map 1 hedon to 1 value, 2 hedons to 10 value, 3 hedons to 100 value...)

So I didn't focus on it here, but I think it's definitely good to think about the Disneyland concern and it's closely related to what I was thinking about when writing the OP.

I really liked @Joe_Carlsmith articulation of your 23-word summary: what if all people are paperclippers relative to one another? Though it does make stronger assumptions than we are here. 

I don't know! It's possible that you can just solve a bargain and then align AI to that, like you can align AI to citizens assemblies. I want to be pitched.

I don't understand why that matters. Whatever discount rate you have, if you're prioritizing between extinction risk and trajectory change you will have some parameters that tell you something about what is going to happen over N years. It doesn't matter how long this time horizon is. I think you're not thinking about whether your claims have bearing on the actual matter at hand. 

It would probably be most useful for you to try to articulate a view that avoids the dilemma I mentioned in the first comment of this thread. 

You're not going to be prioritizing between extinction risk and long term trajectory changes based on tractability if you don't care about the far future. And for any moral theory you can ask "why do you think this will be a good outcome?" and as long as you don't value life intrinsically you'll have to state some empirical hypotheses about the far future 

Any disagreement about longtermist prioritization should presuppose longtermism 

I want to see a bargain solver for AI alignment to groups: a technical solution that would allow AI systems to solve the pie cutting problem for groups and get them the most of what they want, for AI alignment. The best solutions I've seen for maximizing long run value involve using a bargain solver to decide what ASI does, which preserves the richness and cardinality of people's value functions and gives everyone as much of what they want as possible, weighted by importance. (See WWOTF Afterwards, the small literature on bargaining-theoretic approaches to moral uncertainty.) But existing democratic approaches to AI alignment seem to not be fully leveraging AI tools, and instead aligning AI systems to democratic processes that aren't empowered with AI tools (e.g. CIPs and CAIS'S alignment to the written output of citizens' assemblies.) Moreover, in my experience the best way to make something happen is just to build the solution. If you might be interested in building this tool and have the background, I would love to try to connect you to funding for it. 

For deeper motivation see here.

Here's a shower thought: 

• If you think extinction risk reduction is highly valuable, then you need some kind of a model of what Earth-originating life will do with its cosmic endowment
• Some of the parameters in you model must be related to things other than mere survival, like what this life is motivated by or will attempt to do
• Plausibly, there are things you can do to change the values of those parameters and not just the extinction parameter 

It won't work for every model (maybe the other parameters just won't budge), but for some of them it should. 

(Low effort comment as I run out the door, but hope it adds value) To me the most compelling argument in favour of tractability is:

• We could make powerful AI agents whose goals are well understood and do not change or update in ex ante predictable ways.
• These agents are effectively immortal and the most powerful thing in the affectable universe, with no natural competition. They would be able to overcome potentially all natural obstacles, so they would determine what happens in the lightcone.
• So, we can make powerful AI agents that determine what happens in the lightcone, whose goals are well understood and update in ex ante predictable ways.
• So, we can take actions that determine what happens in the lightcone in an ex ante predictable way.