Huh? I’m not sure I understand what your objection.
An NP oracle would let you do stuff like “given this sensory data, find a model of size N or less that within K computational steps or less will reproduce the data to within error x, given such a model exists”
Then one can run “which sequence of actions, given this model, will, within S steps, produce outcome A with probability P?”
Whether or not we can give a rigorous definition of intelligence, seems like the above is sufficient to act like an intelligence, right? Yeah, there’re a few tricky parts re reflective decision theory, but even without that. Even if we let the thing be non-self-modifying… giving a nice chunk of computing power to the above, given an efficient NP oracle, would be enough to potentially cause trouble. Or so I’d imagine.
Just some property of the outcome that you’re interested in. ie, all of the above, with the question being “blah blah blah, with f(outcome) = blah with probability blah blah blah blah blah”
Huh? I don’t follow. (Note, the whole point is that I was claiming that an NP oracle would make, say, a UFAI potentially easy, while to achieve the rather more specific FAI would still be difficult.)
It seems that we may be talking past each other. Could you give an explicit example of what sort of question you would ask the NP oracle to help get a UFAI?
Oooh, sorry, I was unclear. I meant the NP oracle itself would be a component of the AI.
ie, give me an algorithm for efficiently solving NP complete problems, and one could then use that to perform the sort of computations I mentioned earlier.
Hmm, I’m confused. Why do you think that such an object would be helpful as part of the AI? I see how it would be useful to an AI once one had one, but I don’t see why you would want it as a component that makes it easier to make an AGI.
It would make AI easy specifically because it would allow the sorts of computations I described above. If I had a fast way to solve NP complete problems, then I could turn that into a way of performing the previously mentioned computations. Those previously mentioned computations amount to “efficient learning” + “efficient planning”.
The oracle itself is what gives one the ability to perform this computation: Find compact model that efficiently predicts/explains with bounded error the observed sensory data. (This is rough description of the more precise version stated above)
Also gives one the ability to efficiently perform this computation: Given a generated model, determine actions that will lead to desired outcome in bounded number of steps, with reasonably good probability.
The ability to perform the former computation would amount to the ability to efficiently learn. The ability to perform the latter computation would amount to the ability to efficiently plan.
ie, if one has an algorithm for efficiently solving NP complete problems, one can be really good at doing the above two things. The above two things amount to the ability to learn and the ability to plan.
clearer version: the first type of computation would allow it, from observation and such, to determine stuff like the laws of physics, human psychology, etc...
The second computation would allow it to do stuff like… figure out what actions it needs to take to increase the rate of paperclip production or whatever.
(incidentally A slight amount of additional reflectivity, without needing to solve the really hard problems of reflective decision theory or such, would probably be sufficient to allow it to figure out what experiments it needs to do to gain data it needs to form better models.)
Huh? I’m not sure I understand what your objection.
An NP oracle would let you do stuff like “given this sensory data, find a model of size N or less that within K computational steps or less will reproduce the data to within error x, given such a model exists”
Then one can run “which sequence of actions, given this model, will, within S steps, produce outcome A with probability P?”
Whether or not we can give a rigorous definition of intelligence, seems like the above is sufficient to act like an intelligence, right? Yeah, there’re a few tricky parts re reflective decision theory, but even without that. Even if we let the thing be non-self-modifying… giving a nice chunk of computing power to the above, given an efficient NP oracle, would be enough to potentially cause trouble. Or so I’d imagine.
Ok, but how are you specifying the outcome? And are you going to specify each input and output?
Just some property of the outcome that you’re interested in. ie, all of the above, with the question being “blah blah blah, with f(outcome) = blah with probability blah blah blah blah blah”
Then you will need to specify your AI for every single output-input pair you are interested in.
Huh? I don’t follow. (Note, the whole point is that I was claiming that an NP oracle would make, say, a UFAI potentially easy, while to achieve the rather more specific FAI would still be difficult.)
It seems that we may be talking past each other. Could you give an explicit example of what sort of question you would ask the NP oracle to help get a UFAI?
Oooh, sorry, I was unclear. I meant the NP oracle itself would be a component of the AI.
ie, give me an algorithm for efficiently solving NP complete problems, and one could then use that to perform the sort of computations I mentioned earlier.
Hmm, I’m confused. Why do you think that such an object would be helpful as part of the AI? I see how it would be useful to an AI once one had one, but I don’t see why you would want it as a component that makes it easier to make an AGI.
It would make AI easy specifically because it would allow the sorts of computations I described above. If I had a fast way to solve NP complete problems, then I could turn that into a way of performing the previously mentioned computations. Those previously mentioned computations amount to “efficient learning” + “efficient planning”.
I’m sorry but I’m still not following what learning and planning you would do. Are you attaching the oracle to some sort of reward mechanism?
The oracle itself is what gives one the ability to perform this computation: Find compact model that efficiently predicts/explains with bounded error the observed sensory data. (This is rough description of the more precise version stated above)
Also gives one the ability to efficiently perform this computation: Given a generated model, determine actions that will lead to desired outcome in bounded number of steps, with reasonably good probability.
The ability to perform the former computation would amount to the ability to efficiently learn. The ability to perform the latter computation would amount to the ability to efficiently plan.
ie, if one has an algorithm for efficiently solving NP complete problems, one can be really good at doing the above two things. The above two things amount to the ability to learn and the ability to plan.
clearer version: the first type of computation would allow it, from observation and such, to determine stuff like the laws of physics, human psychology, etc...
The second computation would allow it to do stuff like… figure out what actions it needs to take to increase the rate of paperclip production or whatever.
(incidentally A slight amount of additional reflectivity, without needing to solve the really hard problems of reflective decision theory or such, would probably be sufficient to allow it to figure out what experiments it needs to do to gain data it needs to form better models.)