Decision Theory is and can be applied to a variety of problems here. It’s just that AI may face Newcomb-like problems and in particular we want to ensure a 1-boxing-like behavior on the part of AI.
The rationale for TDT-like decision theories is even more general, I think. There’s no guarantee that our world contains only one copy of something. We want a decision theory that would let the AI cooperate with its copies or logical correlates, rather than wage pointless wars.
We want a decision theory that would let the AI cooperate with its copies or logical correlates, rather than wage pointless wars.
Constructing rigorous mathematical foundation of decision theory to explain what a decision problem or a decision or a goal are, is potentially more useful than resolving any given informally specified class of decision problems.
Given the week+ delay in this response, it’s probably not going to see much traffic, but I’m not convinced “reading” source code is all that helpful. Omega is posited to have nearly god-like abilities in this regard, but since this is a rationalist discussion, we probably have to rule out actual omnipotence.
If Omega intends to simply run the AI on spare hardware it has, then it has to be prepared to validate (in finite time and memory) that the AI hasn’t so obfuscated its source as to be unintelligible to rational minds. It’s also possible that the source to an AI is rather simple but it is dependent a large amount of input data in the form of a vast sea of numbers. I.e., the AI in question could be encoded as an ODE system integrator that’s reliant on a massive array of parameters to get from one state to the next. I don’t see why we should expect Omega to be better at picking out the relevant, predictive parts of these numbers than we are.
If the AI can hide things in its code or data, then it can hide functionality that tests to determine if it is being run by Omega or on its own protected hardware. In such a case it can lie to Omega just as easily as Omega can lie to the “simulated” version of the AI.
I think it’s time we stopped positing an omniscient Omega in these complications to Newcomb’s problem. They’re like epicycles on Ptolemaic orbital theory in that they continue a dead end line of reasoning. It’s better to recognize that Newcomb’s problem is a red herring. Newcomb’s problem doesn’t demonstrate problems that we should expect AI’s to solve in the real world. It doesn’t tease out meaningful differences between decision theories.
That is, what decisions on real-world problems do we expect to be different between two AIs that come to different conclusions about Newcomb-like problems?
You should note that every problem you list is a special case. Obviously, there are ways of cheating at Newcomb’s problem if you’re aware of salient details beforehand. You could simply allow a piece of plutonium to decay, and do whatever the resulting Geiger counter noise tells you to. That does not, however, support your thesis that Newcomb’s problem is a totally artificial problem with no logical intrusions into reality.
As a real-world example, imagine an off-the-shelf stock market optimizing AI. Not sapient, to make things simpler, but smart. When any given copy begins running, there are already hundreds or thousands of near-identical copies running elsewhere in the market. If it fails to predict their actions from its own, it will do objectively worse than it might otherwise do.
i don’t see how your example is apt or salient. My thesis is that Newcomb-like problems are the wrong place to be testing decision theories because they do not represent realistic or relevant problems. We should focus on formalizing and implementing decision theories and throw real-world problems at them rather than testing them on arcane logic puzzles.
Well… no, actually. A good decision theory ought to be universal. It ought to be correct, and it ought to work. Newcomb’s problem is important, not because it’s ever likely to happen, but because it shows a case in which the normal, commonly accepted approach to decision theory (CDT) failed miserably. This ‘arcane logic puzzle’ is illustrative of a deeper underlying flaw in the model, which needs to be addressed. It’s also a flaw that’d be much harder to pick out by throwing ‘real world’ problems at it over and over again.
Seems unlikely to work out to me. Humans evolved intelligence without Newcomb-like problems. As the only example of intelligence that we know of, it’s clearly possible to develop intelligence without Newcomb-like problems. Furthermore, the general theory seems to be that AIs will start dumber than humans and iteratively improve until they’re smarter. Given that, why are we so interested in problems like these (which humans don’t universally agree about the answers to)?
I’d rather AIs be able to help us with problems like “what should we do about the economy?” or even “what should I have for dinner?” instead of worrying about what we should do in the face of something godlike.
Additionally, human minds aren’t universal (assuming that universal means that they give the “right” solutions to all problems), so why should we expect AIs to be? We certainly shouldn’t expect this if we plan on iteratively improving our AIs.
It might be nice to be able to see the voting history (not the voters’ names, but the number of up and down votes) on a comment. I can’t tell if my comments are controversial or just down-voted by two people. Perhaps even just the number of votes would be sufficient (e.g. −2/100 vs. −2/2).
If it helps: it’s a fairly common belief in this community that a general-purpose optimization tool is both far superior to, and more interesting to talk about, than a variety of special-purpose tools.
Of course, that doesn’t mean you have to be interested in general-purpose optimization tools; if you’re more interested in decision theory for dinner-menu or economic planners, by all means post about that if you have something to say.
But I suspect there are relatively few communities in which “why are you all so interested in such a stupid and uninteresting topic?” will get you much community approval, and this isn’t one of them.
I’m interested in general purpose optimizers, but I bet that they will be evolved from AIs that were more special purpose to begin with. E.g., IBM Watson moving from Jeopardy!-playing machine to medical diagnostic assistant with a lot of the upfront work being on rapid NLP for the J! “questions”.
Also, there’s no reason that I’ve seen here to believe that Newcomb-like problems give insights into how to develop to decision theories that allow us to solve real-world problems. It seems like arguing about corner cases. Can anyone establish a practical problem that TDT fails to solve because it fails to solve these other problems?
Beyond this, my belief is that without formalization and programming of these decision frameworks, we learn very little. Asking what does xDT do in some abstract situation, so far, seems very handy-wavy. Furthermore, it seems to me that the community is drawn to these problems because they are deceptively easy to state and talk about online, but minds are inherently complex, opaque, and hard to reason about.
I’m having a hard time understanding how correctly solving Newcomb-like problems is expected to advance the field of general optimizers. It seems out of proportion to the problems at hand to expect a decision theory to solve problems of this level of sophistication when the current theories don’t seem to obviously “solve” questions like “what should we have for lunch?”. I get the feeling that supporters of research on these theories assume that, of course, xDT can solve the easy problems so let’s do the hard ones. And, I think evidence for this assumption is very lacking.
Again, if you are interested in more discussion about automated optimization on the level of “what should we have for lunch?” I encourage you to post about it; I suspect a lot of other people are interested as well.
(Deleting comments seems not to be working. Consider this a manual delete.)
Decision Theory is and can be applied to a variety of problems here. It’s just that AI may face Newcomb-like problems and in particular we want to ensure a 1-boxing-like behavior on the part of AI.
The rationale for TDT-like decision theories is even more general, I think. There’s no guarantee that our world contains only one copy of something. We want a decision theory that would let the AI cooperate with its copies or logical correlates, rather than wage pointless wars.
Constructing rigorous mathematical foundation of decision theory to explain what a decision problem or a decision or a goal are, is potentially more useful than resolving any given informally specified class of decision problems.
What is an example of such a real-world problem?
Negotiations with entities who can read the AI’s source code.
Given the week+ delay in this response, it’s probably not going to see much traffic, but I’m not convinced “reading” source code is all that helpful. Omega is posited to have nearly god-like abilities in this regard, but since this is a rationalist discussion, we probably have to rule out actual omnipotence.
If Omega intends to simply run the AI on spare hardware it has, then it has to be prepared to validate (in finite time and memory) that the AI hasn’t so obfuscated its source as to be unintelligible to rational minds. It’s also possible that the source to an AI is rather simple but it is dependent a large amount of input data in the form of a vast sea of numbers. I.e., the AI in question could be encoded as an ODE system integrator that’s reliant on a massive array of parameters to get from one state to the next. I don’t see why we should expect Omega to be better at picking out the relevant, predictive parts of these numbers than we are.
If the AI can hide things in its code or data, then it can hide functionality that tests to determine if it is being run by Omega or on its own protected hardware. In such a case it can lie to Omega just as easily as Omega can lie to the “simulated” version of the AI.
I think it’s time we stopped positing an omniscient Omega in these complications to Newcomb’s problem. They’re like epicycles on Ptolemaic orbital theory in that they continue a dead end line of reasoning. It’s better to recognize that Newcomb’s problem is a red herring. Newcomb’s problem doesn’t demonstrate problems that we should expect AI’s to solve in the real world. It doesn’t tease out meaningful differences between decision theories.
That is, what decisions on real-world problems do we expect to be different between two AIs that come to different conclusions about Newcomb-like problems?
You should note that every problem you list is a special case. Obviously, there are ways of cheating at Newcomb’s problem if you’re aware of salient details beforehand. You could simply allow a piece of plutonium to decay, and do whatever the resulting Geiger counter noise tells you to. That does not, however, support your thesis that Newcomb’s problem is a totally artificial problem with no logical intrusions into reality.
As a real-world example, imagine an off-the-shelf stock market optimizing AI. Not sapient, to make things simpler, but smart. When any given copy begins running, there are already hundreds or thousands of near-identical copies running elsewhere in the market. If it fails to predict their actions from its own, it will do objectively worse than it might otherwise do.
i don’t see how your example is apt or salient. My thesis is that Newcomb-like problems are the wrong place to be testing decision theories because they do not represent realistic or relevant problems. We should focus on formalizing and implementing decision theories and throw real-world problems at them rather than testing them on arcane logic puzzles.
Well… no, actually. A good decision theory ought to be universal. It ought to be correct, and it ought to work. Newcomb’s problem is important, not because it’s ever likely to happen, but because it shows a case in which the normal, commonly accepted approach to decision theory (CDT) failed miserably. This ‘arcane logic puzzle’ is illustrative of a deeper underlying flaw in the model, which needs to be addressed. It’s also a flaw that’d be much harder to pick out by throwing ‘real world’ problems at it over and over again.
Seems unlikely to work out to me. Humans evolved intelligence without Newcomb-like problems. As the only example of intelligence that we know of, it’s clearly possible to develop intelligence without Newcomb-like problems. Furthermore, the general theory seems to be that AIs will start dumber than humans and iteratively improve until they’re smarter. Given that, why are we so interested in problems like these (which humans don’t universally agree about the answers to)?
I’d rather AIs be able to help us with problems like “what should we do about the economy?” or even “what should I have for dinner?” instead of worrying about what we should do in the face of something godlike.
Additionally, human minds aren’t universal (assuming that universal means that they give the “right” solutions to all problems), so why should we expect AIs to be? We certainly shouldn’t expect this if we plan on iteratively improving our AIs.
Harsh crowd.
It might be nice to be able to see the voting history (not the voters’ names, but the number of up and down votes) on a comment. I can’t tell if my comments are controversial or just down-voted by two people. Perhaps even just the number of votes would be sufficient (e.g. −2/100 vs. −2/2).
If it helps: it’s a fairly common belief in this community that a general-purpose optimization tool is both far superior to, and more interesting to talk about, than a variety of special-purpose tools.
Of course, that doesn’t mean you have to be interested in general-purpose optimization tools; if you’re more interested in decision theory for dinner-menu or economic planners, by all means post about that if you have something to say.
But I suspect there are relatively few communities in which “why are you all so interested in such a stupid and uninteresting topic?” will get you much community approval, and this isn’t one of them.
I’m interested in general purpose optimizers, but I bet that they will be evolved from AIs that were more special purpose to begin with. E.g., IBM Watson moving from Jeopardy!-playing machine to medical diagnostic assistant with a lot of the upfront work being on rapid NLP for the J! “questions”.
Also, there’s no reason that I’ve seen here to believe that Newcomb-like problems give insights into how to develop to decision theories that allow us to solve real-world problems. It seems like arguing about corner cases. Can anyone establish a practical problem that TDT fails to solve because it fails to solve these other problems?
Beyond this, my belief is that without formalization and programming of these decision frameworks, we learn very little. Asking what does xDT do in some abstract situation, so far, seems very handy-wavy. Furthermore, it seems to me that the community is drawn to these problems because they are deceptively easy to state and talk about online, but minds are inherently complex, opaque, and hard to reason about.
I’m having a hard time understanding how correctly solving Newcomb-like problems is expected to advance the field of general optimizers. It seems out of proportion to the problems at hand to expect a decision theory to solve problems of this level of sophistication when the current theories don’t seem to obviously “solve” questions like “what should we have for lunch?”. I get the feeling that supporters of research on these theories assume that, of course, xDT can solve the easy problems so let’s do the hard ones. And, I think evidence for this assumption is very lacking.
That’s fair.
Again, if you are interested in more discussion about automated optimization on the level of “what should we have for lunch?” I encourage you to post about it; I suspect a lot of other people are interested as well.
Yeah, I might, but here I was just surprised by the down-voting for contrary opinion. It seems like the thing we ought to foster not hide.
As I tried to express in the first place, I suspect what elicited the disapproval was not the contrary opinion, but the rudeness.
Sorry. It didn’t seem rude to me. I’m just frustrated with where I see folks spending their time.
My apologies to anyone who was offended.