Somewhat related:
Last Thursday on the Discord we had people any% speedrunning and racing the Lean tutorial project . This fits very well into my general worldview: I think that doing mathematics in Lean is like solving levels in a computer puzzle game, the exciting thing being that mathematics is so rich that there are many many kinds of puzzles which you can solve.
https://xenaproject.wordpress.com/2020/05/23/the-complex-number-game/
I’ve had this same question and wrote the Wikiquote page on Vassar while doing research on him.
See also this comment thread. The Harper’s piece from that post also talks a lot about Vassar.
I’m curious how this has turned out. Could you give an update (or point me to an existing one, in case I missed it)?
I’m confused about the tradeoff you’re describing. Why is the first bullet point “Generating better ground truth data”? It would make more sense to me if it said instead something like “Generating large amounts of non-ground-truth data”. In other words, the thing that amplification seems to be providing is access to more data (even if that data isn’t the ground truth that is provided by the original human).
Also in the second bullet point, by “increasing the amount of data that you train on” I think you mean increasing the amount of data from the original human (rather than data coming from the amplified system), but I want to confirm.
Aside from that, I think my main confusion now is pedagogical (rather than technical). I don’t understand why the IDA post and paper don’t emphasize the efficiency of training. The post even says “Resource and time cost during training is a more open question; I haven’t explored the assumptions that would have to hold for the IDA training process to be practically feasible or resource-competitive with other AI projects” which makes it sound like the efficiency of training isn’t important.
And I’ve seen Eliezer make the claim a few times. But I can’t find an article describing the idea. Does anyone have a link?
Eliezer talks about this in Do Earths with slower economic growth have a better chance at FAI? e.g.
Relative to UFAI, FAI work seems like it would be mathier and more insight-based, where UFAI can more easily cobble together lots of pieces. This means that UFAI parallelizes better than FAI.
The addition of the distillation step is an extra confounder, but we hope that it doesn’t distort anything too much—its purpose is to improve speed without affecting anything else (though in practice it will reduce capabilities somewhat).
I think this is the crux of my confusion, so I would appreciate if you could elaborate on this. (Everything else in your answer makes sense to me.) In Evans et al., during the distillation step, the model learns to solve the difficult tasks directly by using example solutions from the amplification step. But if can do that, then why can’t it also learn directly from examples provided by the human?
To use your analogy, I have no doubt that a team of Rohins or a single Rohin thinking for days can answer any question that I can (given a single day). But with distillation you’re saying there’s a robot that can learn to answer any question I can (given a single day) by first observing the team of Rohins for long enough. If the robot can do that, why can’t the robot also learn to do the same thing by observing me for long enough?
I want to highlight a potential ambiguity, which is that “Newton’s approximation” is sometimes used to mean Newton’s method for finding roots, but the “Newton’s approximation” I had in mind is the one given in Tao’s Analysis I, Proposition 10.1.7, which is a way of restating the definition of the derivative. (Here is the statement in Tao’s notes in case you don’t have access to the book.)
I had a similar idea which was also based on an analogy with video games (where the analogy came from let’s play videos rather than speedruns), and called it a live math video.
What is the plan going forward for interviews? Are you planning to interview people who are more pessimistic?
In the first categorization scheme, I’m also not exactly sure what nihilism is referring to. Do you know? Is it just referring to Error Theory (and maybe incoherentism)?
Yes, Huemer writes: “Nihilism (a.k.a. ‘the error theory’) holds that evaluative statements are generally false.”
Usually non-cognitivism would fall within nihilism, no?
I’m not sure how the term “nihilism” is typically used in philosophical writing, but if we take nihilism=error theory then it looks like non-cognitivism wouldn’t fall within nihilism (just like non-cognitivism doesn’t fall within error theory in your flowchart).
I actually don’t think either of these diagrams place Nihilism correctly.
For the first diagram, Huemer writes “if we say ‘good’ purports to refer to a property, some things have that property, and the property does not depend on observers, then we have moral realism.” So for Huemer, nihilism fails the middle condition, so is classified as anti-realist. For the second diagram, see the quote below about dualism vs monism.
I’m not super well acquainted with the monism/dualism distinction, but in the common conception don’t they both generally assume that morality is real, at least in some semi-robust sense?
Huemer writes:
Here, dualism is the idea that there are two fundamentally different kinds of facts (or properties) in the world: evaluative facts (properties) and non-evaluative facts (properties). Only the intuitionists embrace this.
Everyone else is a monist: they say there is only one fundamental kind of fact in the world, and it is the non-evaluative kind; there aren’t any value facts over and above the other facts. This implies that either there are no value facts at all (eliminativism), or value facts are entirely explicable in terms of non-evaluative facts (reductionism).
It seems like “agricultural revolution” is used to mean both the beginning of agriculture (“First Agricultural Revolution”) and the 18th century agricultural revolution (“Second Agricultural Revolution”).
Michael Huemer gives two taxonomies of metaethical views in section 1.4 of his book Ethical Intuitionism:
As the preceding section suggests, metaethical theories are traditionally divided first into realist and anti-realist views, and then into two forms of realism and three forms of anti-realism:
Naturalism / Realism / \ / Intuitionism / \ \ Subjectivism \ / Anti-Realism—Non-Cognitivism \ NihilismThis is not the most illuminating way of classifying positions. It implies that the most fundamental division in metaethics is between realists and anti-realists over the question of objectivity. The dispute between naturalism and intuitionism is then seen as relatively minor, with the naturalists being much closer to the intuitionists than they are, say, to the subjectivists. That isn’t how I see things. As I see it, the most fundamental division in metaethics is between the intuitionists, on the one hand, and everyone else, on the other. I would classify the positions as follows:
Dualism—Intuitionism / / Subjectivism / / \ Reductionism \ / \ \ / Naturalism Monism \ Non-Cognitivism \ / Eliminativism \ Nihilism
Do you have prior positions on relationships that you don’t want to get corrupted through the dating process, or something else?
I think that’s one way of putting it. I’m fine with my prior positions on relationships changing because of better introspection (aided by dating), but not fine with my prior positions changing because they are getting corrupted.
Intelligence beyond your cone of tolerance is usually a trait that people pursue because they think it’s “ethical”
I’m not sure I understand what you mean. Could you try re-stating this in different words?
A question about romantic relationships: Let’s say currently I think that a girl needs to have a certain level of smartness in order for me to date her long-term/marry her. Suppose I then start dating a girl and decide that actually, being smart isn’t as important as I thought because the girl makes up for it in other ways (e.g. being very pretty/pleasant/submissive). I think this kind of change of mind is legitimate in some cases (e.g. because I got better at figuring out what I value in a woman) and illegitimate in other cases (e.g. because the girl I’m dating managed to seduce me and mess up my introspection). My question is, is this distinction real, and if so, is there any way for me to tell which situation I am in (legitimate vs illegitimate change of mind) once I’ve already begun dating the girl?
This problem arises because I think dating is important for introspecting about what I want, i.e. there is a point after which I can no longer obtain new information about my preferences via thinking alone. The problem is that dating is also potentially a values-corrupting process, i.e. dating someone who doesn’t meet certain criteria I think I might have means that I can get trapped in a relationship.
I’m also curious to hear if people think this isn’t a big problem (and if so, why).
I have only a very vague idea of what you mean. Could you give an example of how one would do this?
I think that makes sense, thanks.
Just to make sure I understand, the first few expansions of the second one are:
f(n)
f(n+1)
f((n+1) + 1)
f(((n+1) + 1) + 1)
f((((n+1) + 1) + 1) + 1)
Is that right? If so, wouldn’t the infinite expansion look like f((((...) + 1) + 1) + 1) instead of what you wrote?
I read the post and parts of the paper. Here is my understanding: conditions similar to those in Theorem 2 above don’t exist, because Alex’s paper doesn’t take an arbitrary utility function and prove instrumental convergence; instead, the idea is to set the rewards for the MDP randomly (by sampling i.i.d. from some distribution) and then show that in most cases, the agent seeks “power” (states which allow the agent to obtain high rewards in the future). So it avoids the twitching robot not by saying that it can’t make use of additional resources, but by saying that the twitching robot has an atypical reward function. So even though there aren’t conditions similar to those in Theorem 2, there are still conditions analogous to them (in the structure of the argument “expected utility/reward maximization + X implies catastrophe”), namely X = “the reward function is typical”. Does that sound right?
Writing this comment reminded me of Oliver’s comment where X = “agent wasn’t specifically optimized away from goal-directedness”.
Can you say more about Alex Turner’s formalism? For example, are there conditions in his paper or post similar to the conditions I named for Theorem 2 above? If so, what do they say and where can I find them in the paper or post? If not, how does the paper avoid the twitching robot from seeking convergent instrumental goals?
I’m not entirely sure what I’m trying to learn here (which is part of what I was trying to express with the final paragraph of my question); this just seemed like a natural question to ask as I started thinking more about AI takeoff.
In “I Heart CYC”, Robin Hanson writes: “So we need to explicitly code knowledge by hand until we have enough to build systems effective at asking questions, reading, and learning for themselves. Prior AI researchers were too comfortable starting every project over from scratch; they needed to join to create larger integrated knowledge bases.”
It sounds like he expects early AGI systems to have lots of hand-coded knowledge, i.e. the minimum number of bits needed to specify a seed AI is large compared to what Eliezer Yudkowsky expects. (I wish people gave numbers for this so it’s clear whether there really is a disagreement.) It also sounds like Robin Hanson expects progress in AI capabilities to come from piling on more hand-coded content.
If ML source code is small and isn’t growing in size, that seems like evidence against Hanson’s view.
If ML source code is much smaller than the human genome, I can do a better job of visualizing the kind of AI development trajectory that Robin Hanson expects, where we stick in a bunch of content and share content among AI systems. If ML source code is already quite large, then it’s harder for me to visualize this (in this case, it seems like we don’t know what we’re doing, and progress will come from better understanding).
If the human genome is small, I think that makes a discontinuity in capabilities more likely. When I try to visualize where progress comes from in this case, it seems like it would come from a small number of insights. We can take some extreme cases: if we knew that the code for a seed AGI could fit in a 500-line Python program (I don’t know if anybody expects this), a FOOM seems more likely (there’s just less surface area for making lots of small improvements). Whereas if I knew that the smallest program for a seed AGI required gigabytes of source code, I feel like progress would come in smaller pieces.
I’m not sure. The content/architecture split doesn’t seem clean to me, and I haven’t seen anyone give a clear definition. Specialized data structures seems like a good example of something that’s in between.