= 27d567bc2d48d9f40e9ccc20ea370b15)
cousin_it comments on How Not to be Stupid: Adorable Maybes - Less Wrong
You are viewing a comment permalink. View the original post to see all comments and the full post content.
= 783df68a0f980790206b9ea87794c5b6)
Loading…= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Powered by Reddit
= f037147d6e6c911a85753b9abdedda8d)
You are viewing a comment permalink. View the original post to see all comments and the full post content.
Comments (54)
I downvoted this post because it's not a question of buying; your ranking argument is logically invalid, as indicated by gjm, Vladimir_Nesov and me in the comments to your previous post in the series.
"buying" in the sense of "assuming you consider the argument valid" but actually, I've rethought about it several times and I think you're right about that. I think I'm going to edit that bit somewhat in light of that.
Do you accept that IF for some agent it can be said that for any two states, they prefer one to the other or are indifferent (ie, have just as much preference for one as for the other) THEN, that combined by the "don't be stupid" rule, prohibits cycles in the preference rankings?
Yes for idealized agents. Not yet convinced about humans.
See, if your theory eventually runs counter to common sense on Pascal's Mugging (Eliezer says he has no good solution, common sense says decline the offer) or Dust Specks (Eliezer chooses torture, common sense chooses dust specks), we will have to reexamine the assumptions again. It could easily be that the utility function assumption is faulty, or well-orderedness is faulty, or something else.
Actually, IIRC, Eliezer said that he thinks Robin Hanson's (I think it was his) solution to the mugging seems to be in the right direction. But that gets into computational power issues. Actually, my original intent was to name this sequence "How not to Be Stupid (given unbounded computational power)"
Obviously we can't do the full decision theory computations in full exact correctness. And I did give the warning against hastily giving an oversimplified human preference generator. What I'm going for here is more "why assume that Bayesian decision theory is the thing we should be building approximations to, rather than some other entirely different blob of math?"
(Oh, incidentally. I originally chose SPECKS, then later one of the comments in that sequence of posts (the comment that stepped through it, incrementally reducing, etc) ended up convincing me to switch to TORTURE.)
Also, finished editing the offending argument.
Over the last couple years I went from believing that statement to deeply doubting it. If you want a chess player that will win games by holding the opponents' kids hostage, sure, build a Bayesian optimizer. My personal feeling is that even an ordinary human modified to be deeply and genuinely driven by an explicit utility function would pose a substantial danger to this world. No need for AIs.
That's where the whole "don't assume an overly simplistic preference ranking for yourself" warnings come in.
ie, nothing wrong with the utility function being composed of terms for all the things we value, and simply happening to include for that player a component that translates to "win at chess by actually playing chess", and other components giving stuff that lowers utility for "kids have been kidnapped" situations, etc etc etc.
The hard part is, of course, actually translating the algorithms we're running (including the bits that respond to arguments that lead us to become convinced to change our minds about a moral question, etc etc) into a more explicit algorithm. Any simple one is going to get it WRONG.
But that's not a hit against decision theory. That's a hit against bad utility functions.
Or did I utterly misunderstand your point?
We know from Eliezer's writings that almost any strong goal-directed chessplayer AI will destroy the world. Well guess what, if a non-world-destroying utility function appears almost impossibly hard to formulate, in my book it counts as a hit against the concept of utility functions. Especially seeing as machines based on e.g. control theory (RichardKennaway) behave much more sensibly - they almost never display any urge to screw up the whole world, instead being content to sit there and tweak their needle.
Well, a recursively self modifying chess playing AI is a very different beast than a human who, AMONG OTHER THINGS, cares about doing well at chess. The sum total of those other things and chess together is a very different goal system than "chess and nothing else".
As far as control theory, well... that's because control theory based systems are currently too stupid to pose such a threat to us, no?
Your judgment against decision theory seems to be "an agent based on it will act in accordance with its utility function... which may not resemble meaningfully my preferences. It may not be moral, etc etc etc. It will be good at what it's trying to do... but it isn't exactly trying to do the stuff I care about."
Do you consider this a fair summary of your position?
If so, then the response is, well... So, it's good at doing the stuff it's trying to do. It's not trying to do what we'd prefer it to be doing. This is a serious problem. But that problem isn't a flaw with decision theory itself. I mean, if decision theory is leading it to be good optimizing reality in accordance with its preference rankings, then decision theory is acting as promised. The problem is "it's trying to do stuff we don't want it to do!"
The things we care abut are complicated. To actually specifically accurately fully and explicitly specify that is REALLY HARD. That doesn't mean decision theory is inherently flawed. It means, well, fully specifying what we actually want is a highly nontrivial problem.
I agree with you that the math is right. Given assumptions, it acts as promised. But the assumptions just aren't a good model of reality. Like naive game theory: you can go with the mathematically justified option of Always Defect, or you can go with common sense. Reality doesn't contain preference rankings over all possible situations; shoehorning reality into preference rankings might hurt you. Hasn't this point clicked yet? I'll try again.
Human beings aren't goal systems. We DON'T SUM, anymore than a car "sums" the value of its speedometer with the value of the fuel gauge. If we actually summed, you'd get the outcome Eliezer once advocated: every one of us "picking one charity and donating as much to it as he can". Your superintelligent chess player with the "correct" utility function won't ever play chess while there are other util-rich tasks anywhere in the world, like hunger in Africa.
We shouldn't need to fully specify what we actually want, if we're building a specialized machine to e.g. cure world hunger or design better integrated circuits. It would be better to build such machines based on a theory that typically results in localized screw-ups... rather than a theory that destroys the world by default, unless you tell it everything about you.
What if we're building a specialized machine to prevent a superintelligence from annihilating us?
Where's the "I super-agree" button?
I agree with you that maximizing utility is dangerous and wrong even just in ordinary humans. That's not what we're for and that's not what the good life is about.
We don't need a clean-cut, provable decision theory that will drive the universe into a hole of 'utility'. We need more of a wibbly-wobbly, humany-ethicy ball of... stuff.
That seems an obviously fallacious argument to me. Many posts on OB have talked about other motivations behind charitable giving - whether it's 'buying fuzzies' or signalling. You seem to be arguing that because one possible (but naive and inaccurate) model of a person's utility function would predict different behaviour than what we actually observere, that the observed behaviour is evidence against any utility function being maximized. There are pretty clearly at least two possibilities here: either humans don't maximize a utility function or they maximize a different utility function from the one you have in mind.
Personally I think humans are imperfect maximizers of utility functions that are sufficiently complex that the 'function' terminology is as misleading as it is enlightening but your argument really doesn't support your conclusion.
This is a rather bad example -- machines based on control theory can easily display an "urge" to screw up as much of the world as they can touch. Short version: slapping a PID controller onto a system gives it second order dynamics, and those can have a resonant frequency. If the random disturbance has power at the resonant frequency, the system goes into a positive feedback loop and blows up.
I agree. But what do you do with this situation? To give up, you have to be certain that there is no way out, and we are much too confused to say anything like that yet. Someone is bound to build a doom machine someday if you don't do something about it.
Normative decision theory – the structure our final, stable preferences if we knew more, thought faster, were more the people we wished we were, had grown up further together – needn't be good engineering design; agreed that utility functions often aren't the latter, but that doesn't count against them as the former.
Maybe Psy-Kosh should say "becoming" instead of "building"?
That is a right sentiment about strength: there are no simple rules, only goals, which makes a creative mind extremely dangerous. And we shouldn't build things like this without understanding what the outcome will be. This is one of the reasons it's important to understand human values in this light, to guard them from this destructive potential.
Whatever you want accomplished, whatever you want averted, instrumental rationality defines an optimal way of doing that (without necessarily giving the real-world means, that's a next step). If you really want life to continue as before, the correctly implemented explicit utility function for doing that won't lead a Bayesian optimizer to do something horrible. (Although inaction may be considered horrible in itself, where so much more could've been done.)
You don't get to assume that till tomorrow.
Your statements about application of decision-making to humans still fail to make any sense to me. I fail to form a coherent model of how you understand this issue. Could you try to write up a short step-by-step introduction to your position, maybe basic terms only, just to establish a better vocabulary to build on? Open thread seems like a right place for such post.
Short version: beyond a certain (very coarse) precision you can't usefully model humans as logical, goal-directed, decision-making agents contaminated by pesky "biases". Goals, decisions and agency are very leaky abstractions, illusions that arise from the mechanical interplay of our many ad-hoc features. Rather than heading off for the sunset, the 99% typical behavior of humans is going around in circles day after day; if this is goal-directed, the goal must be weird indeed. If you want to make predictions about actual human beings, don't talk about their goals, talk about their tendencies.
Far from distressing me, this situation makes me happy. It's great we have so few optimizers around. Real-world strong optimizers, from natural selection to public corporations to paperclippers, look psychopathic and monstrous when viewed through the lens of our tendency-based morality.
For more details see thread above. Or should I compile this stuff into a toplevel post?
Okay, I've probably captured the gist of your position now. Correct me if I'm speaking something out of its character below.
Humans are descriptively not utility maximizers, they can only be modeled this way under coarse approximation and with a fair number of exceptions. There seems to be no reason to normatively model them with some ideal utility maximizer, to apply the concepts like should in more rigorous sense of decision theory.
Humans do what they do, not what they "should" according to some rigorous external model. This is an argument and intuition similar to not listening to philosopher-constructed rules of morality, non-intuitive conclusions reached from considering a thought experiment, or God-declared moral rules, since you first have to accept each moral rule yourself, according to your own criteria, which might even be circular.
I thought this was the point of the Overcoming Bias project and the endeavor not to be named until tomorrow (cf. "Thou Art Godshatter" and "Value is Fragile"): that we want to put the fearsome power of optimization in the service of humane values, instead just of leaving things to nature, which is monstrous.
I would love to see a top-level post on this issue.
Is that addressed to cousin_it or Psy-Kosh?
To cousin_it, obviously...
Thanks. (It wasn't obvious to me, because I've seen similar comments from you to Psy-Kosh recently, and don't remember seeing any such to cousin_it. And it's not entirely outside the bounds of possibility for someone to make a comment a sibling rather than a child of what it's responding to.)