I am a very strong satisficer, in direct conflict to my moral system which would rather I maximise, so I live under the general understanding that I’m very far from my ideal.
Veedrac
I would be interested in how you would falsify it regardless. I am confused about what I am meant to be confused about (what does it mean for it to not be okay?) and I suspect the excersise would remedy that.
The exercise in falsification refers to Conor’s last sentence, only no longer applied specific to him.
I’m wondering how you would falsify the claim (that I predict you will make and be justified in making) that I don’t get it.
When I say I am confused about what I am meant to be confused about, I mean that I’m failing to identify as Alex. He at least has a command he knows he cannot do (Look above that! / That’s the top.), whereas I am stuck in the realm of unknown unknowns.
Your paragraph on the “it” from your kenshō is a much closer description of how I currently feel than the inverse is; I don’t understand what it would mean for this claim to be untrue except in the sense that it “not being okay” accurately describes external reality. But that feels like it falls into the same trap that your bullet points are said to, only in the opposite direction.
Your later post about the benefits does this more clearly; with absolute exception of the point about energy, and potential exception of the last, the other points seem oddly accurate representations of the difference between me and the average person. But I don’t think I am enlightened.
So, on a concrete level, this comes through as the question of how would you differentiate someone who was born enlightened from someone who was not, but is perhaps mistakenly labelling a shallow surface immitation?
I believe I understood this metaphor. However, it seems to me this isn’t a good place to be, since I predict the metaphor is only useful to ground discussion about the thing that’s actually taking place. It is that second step that hasn’t worked.
Let’s flip this around. How do you know when someone is Looking? Is there a way to do so based on external behaviours? What is your equivalent of the following?
“I’m watching you stare at your phone. If your Looking, your head would be up and your eyes would be pointed at me.”
You give a good example with the hair clipper, but I don’t know how much, if at all, that relates to Looking. If it is closely related I have a few follow-up questions that probably get to the crux of the issue I specifically am stuck on.
Fair warning, the following is pretty sketchy and I wouldn’t bet I’d stick with it if I thought a bit longer.
---
Imagine a simple computer running a simple chess playing program. The program uses purely integer computation, except to calculate its reward function and to run minimax over them, which is in floating point. The search looks for the move that maximizes the outcome, which corresponds to a win.
This, if I understand your parlance, is ‘rational’ behaviour.
Now consider that the reward is negated, and the planner instead looks for the move that minimizes the outcome.
This, if I understand your parlance, is ‘anti-rational’ behaviour.
Now consider that this anti-rational program is run on a machine where floating point values encoded with a sign bit ‘1’ represent a positive number and those with a ‘0’ sign bit a negative number—the opposite to the standard encoding.
It’s the same ‘anti-rational’ program, but exactly the same wires are lit up in the same pattern on this hardware as with the ‘rational’ program on the original hardware.
In what sense can you say the difference between rationality and anti-rationality at all exists in the program (or in humans), rather than in the model of them, when the same wires are both rational and anti-rational? I believe the same dilemma holds for indifferent planners. It doesn’t seem like reward functions of the type your paper talks about are a real thing, at least in a sense independent of interpretation, so it makes sense that you struggle to distinguish them when they aren’t there to distinguish.
---
I am tempted to base an argument off the claim that misery is avoided because it’s bad rather than being bad because it’s avoided. If true, this shortcuts a lot of your concern: reward functions exist only in the map, where numbers and abstract symbols can be flipped arbitrarily, but in the physical world these good and bad states have intrinsic quality to them and can be distinguished meaningfully. Thus the question is not how to distinguish indistinguishable reward functions, but how to understand this aspect of qualitative experience. Then, presumably, if a computer could understand what the experience of unhappiness is like, it would not have to assume our preferences.
This doesn’t help solve the mystery. Why couldn’t a species evolve to maximise its negative internal emotional states? We can’t reasonably have gotten preference and optimization lined up by pure coincidence, so there must be a reason. But it seems like a more reasonable stance to shove the question off into the ineffable mysteries of qualia than to conflate it with a formalism that seems necessarily independent of the thing we’re trying to measure.
Pushing q towards 1 might be a disaster
If I consider satisfaction of my preferences to be a disaster, in what sense can I realistically call them my preferences? It feels like you’re more caught up on the difficulty of extrapolating these preferences outside of their standard operation, but that seems like a rather different issue.
Obviously misery would be avoided because it’s bad, not the other way around.
As mentioned, this isn’t obvious to me, so I’d be interested in your reasoning. Why should evolution build systems that want to avoid intrinsically bad mental states?
We are trying to figure out what is bad by seeing what we avoid. And the problem remains whether we might be accidentally avoiding misery, while trying to avoid its opposite.
Yes, my point here was twofold. One, the formalism used in the paper does not seem to be deeply meaningful, so it would be best to look for some other angle of attack. Two, given the claim about intrinsic badness, the programmer is embedding domain knowledge (about conscious states), not unlearnable assumptions. A computer system would fail to learn this because qualia is a hard problem, not because it’s unlearnable. This makes it asymmetric and circumventable in a way that the no free lunch theorem is not.
But that doesn’t detract from the main point: that simplicity, on its own, is not sufficient to resolve the issue.
It kind of does. You have shown that simplicity cannot distinguish (p, R) from (-p, -R), but you have not shown that simplicity cannot distinguish a physical person optimizing competently for a good outcome from a physical person optimizing nega-competently for a bad outcome.
If it seems unreasonable for there to be a difference, consider a similar map-territory distinction of a height map to a mountain. An optimization function that gradient descents on a height map is the same complexity, or nearabouts, as one that gradient ascents on the height map’s inverse. However, a system that physically gradient descents on the actual mountains can be much simpler than one that gradient ascents on the mountain’s inverse. Since negative mental experiences are somehow qualitatively different to positive ones, it would not surprise me much if they did in fact effect a similar asymmetry here.
One of us is missing what the other is saying. I’m honestly not sure what argument you are putting forth here.
I agree that preference/reward is an interpretation (the terms I used were map and territory). I agree that (p,R) and (-p,-R) are approximately equally complex. I do not agree that complexity is necessarily isomorphic between the map and the territory. This means although the model might be a strong analogy when talking about behaviour, it is sketchy to use it as a model for complexity of behaviour.
“Agent A has preferences R” is not a fact about the world. It is a stance about A, or an interpretation of A. A stance or an interpretation that we choose to take, for some purpose or reason.
I find it hard to imagine that you’re actually denying that you or I have things that, colloquially, one would describe as preferences, and exist in an objective sense. I do have a preference for a happy and meaningful life over a life of pure agony. Anyone who thinks I do not is factually wrong about the state of the world.
Then there is a sense in which the interpretations of these systems we build are fully interpretative. If “preferences R” refers to a function returning a real number, for sure this is not some facet of the real world, and there are many such seemingly-different models for any agent. Here again I believe we agree.
But we seem not to be agreeing at the next step, with the preference stance. Here I claim your goal should not be to maximize the function “preferences R”, whose precise values are irrelevant and independent, but to maximise the actual human preferences.
Consider measuring a simpler system, temperature, and projecting this onto some number. Clearly, depending on how you do this projection, you can end up at any number for a given temperature. Even with a simplicity prior, higher temperatures can correspond to larger numbers or smaller numbers in the projection, with pretty much equal plausibility. So even in this simplified situation, where we can agree that some temperatures are objectively higher than others, you cannot reliably maximize temperature by maximizing its projection.
Your preference function is a projection. The arbitrary choices you have to make to build this function are not assumptions about the world, they are choices about the model. When you prove that you have many models of human preference, you are not proving that preference is entirely subjective.
That’s why, when you use empathy to figure out someone’s goals and rationality, this also allows you to better predict them. But this is a fact about you (and me), not about the world. Just as “Thor is angry” is actually much more complex than electromagnetism, our prediction of other people via our empathy machine is simpler for us to do—but is actually more complex for an agent that doesn’t already have this empathy machinery to draw on.
This Thor analogy is… enlightening of the differences in our perspectives. Imagining an angry Thor is a much more complex hypothesis up until the point you see an actual Thor in the sky hurling spears of lightning. Then it becomes the only reasonable conclusion, because although brains seem like they involve a lot of assumptions, a brain is ultimately many fewer assumptions (to the pre-industrial Norse people) than that same amount of coincidence.
This is the point I am making with people. If your computer models people as arbitrary, randomly sampled programs, of course you struggle to distinguish human behaviour from their contrapositives. However, people are not fully independent, nor arbitrary computing systems. Arguing that a physical person optimizing competently for a good outcome and a physical person optimizing nega-competently for a bad outcome are similarly simple has to overcome at least two hurdles:
1. We seem to know things about which mental states are good and which mental states are bad. This implies there is objective knowledge that can be learnt about it.
2. You would need to extend your arguments about mathematical functions into the real world. I don’t know how this could be approached.
I have a hard time believing that in another world people think that the qualia corresponding to our suffering is good and the qualia corresponding to our happiness is bad, and if it is, this strikes me as a much bigger deal than anything else you are saying.
I deny that a generic outside observer would describe us as having any specific set of preferences, in an objective sense.
It’s possible that we’ve been struggling with this conversation because I’ve been failing to grasp just how radically different your opinions are to mine.
Imagine your generic outside observer was superintelligent, and understood (through pure analysis) qualia and all the corresponding mysteries of the mind. Would you then still say this outside observer would not consider us to have any specific set of preferences, in an objective sense, where “preferences” takes on its colloquial meaning?
If not, why? I think my stance is obvious; where preferences colloquially means approximately “a greater liking for one alternative over another or others”, all I have to claim is that there is an objective sense in which I like things, which is simple because there’s an objective sense in which I have that emotional state and internal stance.
Here’s a rather out-there hypothesis.
I’m sure many LessWrong members have had the experience of arguing some point piecemeal, where they’ve managed to get weak agreement on every piece of the argument, but as soon as they step back and point from start to end their conversation partner ends up less than convinced. In this sense, in humans even implication isn’t transitive. Mathematics is an example with some fun tales I’m struggling to find sources for, where pre-mathematical societies might have people unwilling to trade two of A for two of B, but happy to trade A for B twice, or other such oddities.
It’s plausible to me that the need for consistent models of the world only comes about as intelligence grows and allows people to arbitrage value between these different parts of their thoughts. Early humans and their lineage before that weren’t all that smart, so it makes sense that evolution didn’t force their beliefs to be consistent all that much—as long as it was locally valid, it worked. As intelligence evolved, occasionally certain issues might crop up, but rather than fixing the issue in a fundamental way, which would be hard, minor kludges were put in place.
For example, I don’t like being exploited. If someone leads me around a pump, I’m going to value the end state less than its ‘intrinsic’ value. You can see this behaviour a lot in discussions of trolley problem scenarios: people take objection to having these thoughts traded off against each other to the degree it often overshadows the underlying dilema. Similarly, I find gambling around opinions intrinsically uncomfortable, and notice that fairly frequently people take objection to me asking them to more precisely quantify their claims, even in cases where I’m not staking an opposing claim. Finally, since some people are better at sounding convincing than I am, it’s completely reasonable to reject some things more broadly because of the possibility the argument is an exploit—this is epistemic learned helplessness, sans ‘learned’.
There are other explanations for all the above, so this is hardly bulletproof, but I think there is merit to considering evolved defenses to exploitation that don’t involve being exploit-free, as well as whether there is any benefit to something of this form. Behaviours that avoid and back away from these exploits seem fairly obvious places to look into. One could imagine (sketchily, non-endorsingly) an FAI built on these principles, so that even without a bulletproof utility function, the AI would still avoid self-exploit.
Moore’s Law is not dead. I could rant about the market dynamics that made people think otherwise, but it’s easier just to point to the data.
https://docs.google.com/spreadsheets/d/1NNOqbJfcISFyMd0EsSrhppW7PT6GCfnrVGhxhLA5PVw
Moore’s Law might die in the short future, but I’ve yet to hear a convincing argument for when or why. Even if it does die, Cerebras presumably has at least 4 node shrinks left in the short term (16nm→10nm→7nm→5nm→3nm) for a >10x density scaling, and many sister technologies (3D stacking, silicon photonics, new non-volatile memories, cheaper fab tech) are far from exhausted. One can easily imagine a 3nm Cerebras waffle coated with a few layers of Nantero’s NRAM, with a few hundred of these connected together using low-latency silicon photonics. That would easily train quadrillion parameter models, using only technology already on our roadmap.
Alas, the nature of technology is that while there are many potential avenues for revolutionary improvement, only some small fraction of them win. So it’s probably wrong to look at any specific unproven technology as a given path to 10,000x scaling. But there are a lot of similarly revolutionary technologies, and so it’s much harder to say they will all fail.
Density is important because it affects both price and communication speed. These are the fundamental roadblocks to building larger models. If you scale to too large clusters of computers, or primarily use high-density off-chip memory, you spend most of your time waiting for data to arrive in the right place.
If the issue is the size of having a fine-tuned model for each individual task you care about, why not just fine-tune on all your tasks simultaneously, on one model? GPT-3 has plenty of capacity.
I think the results in that paper argue that it’s not really a big deal as long as you don’t make some basic errors like trying to fine-tune on tasks sequentially. MT-A outperforms Full in Table 1. GPT-3 is already a multi-task learner (as is BERT), so it would be very surprising if training on fewer tasks was too difficult for it.
Apple’s launch events get pretty big crowds, a lot of talk, and a lot of celebration.
Legalizing blackmail gives people with otherwise no motivation to harm someone through the sharing of information the motive to do so. I’m going to take that as the dividing line between blackmail and other forms of trade or coercion. I believe this much is generally agreed on in this debate.
If you’re going to legalize forced negative-sum trades, I think you need a much stronger argument that assuming that, on net, the positive externalities will make it worthwhile. It’s a bit like legalizing violence from shopkeepers because most of the time they’re punching thieves. Maybe that’s true now, when shopkeepers punching people is illegal, but one, I think there’s a large onus on anyone suggesting this to justify that it’s the case, and two, is it really going to stay the case, once you’ve let the system run with this newfound form of legalized coercion?
Before I read these excerpts, I was pretty much in the ‘blackmail bad, duh’ category. After I read them, I was undecided; maybe it is in fact true that many harms from information sharing comes with sufficient positive externalities, and those that do not are sufficiently clearly delimited to be separately legislated. Having thought about it longer, I now see a lot of counterexamples. Consider some person, who:
had a traumatic childhood,
has a crush on another person, and is embarrassed about it,
has plans for a surprise party or gift for a close friend,
or the opposite; someone else is planning a surprise for them,
has an injury or disfiguration on a covered part of their body,
had a recent break-up, that they want to hold out on sharing with their friends for a while,
left an unkind partner, and doesn’t want that person to know they failed a recent exam,
posts anonymously for professional reasons, or to have a better work-life balance,
doesn’t like a coworker, but tries not to show it on the job.
I’m sure I could go on for quite a while. Legalizing blackmail means that people are de-facto incentivized to exploit information when it would harm people, because their payout stops being derived from the public interest, through mechanisms like public reception, appreciation from those directly helped by the reveal of information, or payment from a news agency, and becomes proportional almost purely to the damage you can do.
It’s true that in some cases these are things which should be generally disincentivized or made illegal, nonconsensual pornography being a prime example. In general I don’t think this approach scales, because the public interest is so context dependent. Sometimes it is in the public interest to share someone’s traumatic childhood, spoil a surprise or tell their coworker they are disliked. But the reward should be derived from the public interest, not the harm! If we want to monetarily incentivize people to share information they have on sexual abuse, pay them for sharing information that led to a conviction. And if you’re not wanting to do that because it causes the bad incentive to lie… surely blackmail gives more incentive to lie, and the accuser being paid requires the case never to have gone to trial, so is worse on all accounts.
Robin Hanson argued that negative gossip is probably net positive for society.
Yes, this is what my post was addressing and the analogy was about. I consider it an interesting hypothesis, but not one that holds up to scrutiny.
Lying about someone in a damaging way is already covered by libel/slander laws.
I know, but this only further emphasizes how much better paying those who helped a conviction is. Blackmail is private, threat-based, and necessarily unpoliced, whereas the courts have oversight and are an at least somewhat impartial test for truth.
I formed a similar argument around vegetarianism; I predicted that it is easier for me to draw a hard line than it is to reconsider that line on a case by case basis. Rational me is more than capable of distinguishing between lobster and cow, but there is a lot of power in being able to tell myself to just eat the things with the label.
This is an extreme overapproximation but, given the moral stakes and my general unreliability, the successful results seem sufficient justification.