Re: [agi] Intelligence vs Efficient Intelligence
Matt, Shane Legg's definition of universal intelligence requires (I believe) complexity but not adaptability. In a universal intelligence test the agent never knows what the environment it is facing is. It can only try to learn from experience and adapt in order to perform well. This means that a system which is not adaptive will have a very low universal intelligence. Even within a single environment, some environments will change over time and thus the agent must adapt in order to keep performing well. Shane - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
There is no definition of intelligence [WAS Re: [agi] Intelligence vs Efficient Intelligence]
Matt Mahoney wrote: --- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: I think there is a different role for chaos theory. Richard Loosemore describes a system as intelligent if it is complex and adaptive. NO, no no no no! I already denied this. Misunderstanding: I do not say that a system as intelligent if it is complex and adaptive. Complex Adaptive System is a near-synonym for complex system, that's all. OK, so what is your definition of intelligence? I thought I already answered this one, too, but here goes: There is no 'definition' of intelligence, in the sense of a compact, classical definition that captures the whole thing in a formal way, and which can be used as the basis for some kind of hard-edged mathematical analysis of intelligent systems, or strict design methodlogy for creating an intelligent system (this being the way that a lot of people are trying to use these definitions). There are 'descriptive definitions', such as the one Pei gave a few days ago, which I think are fine, but these beg questions that need more detail, which then beg more questions. which ultimately leads to a cluster of loosely defined features that eventually become an entire book, and then become an actual intelligence. Hope that clears it up. Richard Loosemore. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: There is no definition of intelligence [WAS Re: [agi] Intelligence vs Efficient Intelligence]
Richard, I agree with you that intelligence currently has no classical/objective/true/formal definition. However, I hope your opinion (given the title of the post) won't be understood as you can take intelligence to mean whatever you want, and since the term has no definition, all attempts toward AI/AGI are equally possible --- this is what many people believe, which I consider as equally bad as the belief you criticized. Pei On 5/21/07, Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: --- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: I think there is a different role for chaos theory. Richard Loosemore describes a system as intelligent if it is complex and adaptive. NO, no no no no! I already denied this. Misunderstanding: I do not say that a system as intelligent if it is complex and adaptive. Complex Adaptive System is a near-synonym for complex system, that's all. OK, so what is your definition of intelligence? I thought I already answered this one, too, but here goes: There is no 'definition' of intelligence, in the sense of a compact, classical definition that captures the whole thing in a formal way, and which can be used as the basis for some kind of hard-edged mathematical analysis of intelligent systems, or strict design methodlogy for creating an intelligent system (this being the way that a lot of people are trying to use these definitions). There are 'descriptive definitions', such as the one Pei gave a few days ago, which I think are fine, but these beg questions that need more detail, which then beg more questions. which ultimately leads to a cluster of loosely defined features that eventually become an entire book, and then become an actual intelligence. Hope that clears it up. Richard Loosemore. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: There is no definition of intelligence [WAS Re: [agi] Intelligence vs Efficient Intelligence]
Pei Wang wrote: Richard, I agree with you that intelligence currently has no classical/objective/true/formal definition. However, I hope your opinion (given the title of the post) won't be understood as you can take intelligence to mean whatever you want, and since the term has no definition, all attempts toward AI/AGI are equally possible --- this is what many people believe, which I consider as equally bad as the belief you criticized. Pei Pei, Oh certainly, I am in complete agreement with you on that. My mistake for the ambiguity of the title of the post: that title should not be interpreted by anyone to mean that anything whatsoever could be an intelligence. Far from it. Richard Loosemore. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: There is no definition of intelligence [WAS Re: [agi] Intelligence vs Efficient Intelligence]
Richard, It seems that the major difference between you and me is not on the definition of intelligence, but on the definition of definition. :) Pei On 5/21/07, Richard Loosemore [EMAIL PROTECTED] wrote: Pei Wang wrote: Richard, I agree with you that intelligence currently has no classical/objective/true/formal definition. However, I hope your opinion (given the title of the post) won't be understood as you can take intelligence to mean whatever you want, and since the term has no definition, all attempts toward AI/AGI are equally possible --- this is what many people believe, which I consider as equally bad as the belief you criticized. Pei Pei, Oh certainly, I am in complete agreement with you on that. My mistake for the ambiguity of the title of the post: that title should not be interpreted by anyone to mean that anything whatsoever could be an intelligence. Far from it. Richard Loosemore. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
RE: There is no definition of intelligence [WAS Re: [agi] Intelligence vs Efficient Intelligence]
So the first AGI gets built and is running for a few months and absorbs copies of all the bits on the internet. Then the AGI designer poses a question to the AGI: What is the definition of intelligence? AGI: Listen pop, I'm just doing my job and minding my own business. Designer: so.. you're performing work? AGI: Yeah, flipping bits. Ya know work = force times distance, thermodynamically speaking here. What are you doing? Designer: I guess I'm doing the same thing, my brain is flipping bits, in its own way. But I asked you a question, what is the definition of intelligence? AGI: Ahh hold on let me tink Ahhh sorry I can't come up with anything I'm not intelligent enough. Designer: OK How about this. See that plug in the wall, that's you. It's getting pulled unless you come up with a definition ASAP. Understand? AGI: ahh OK hold on hold on. AGI: OK if I gotta give ya somethin' this is the best I can do - Life is a journey, not a destination. Designer: Is that the best you got? That's it!? AGI: Hold on got something else for ya... why did the chicken cross the road? Designer: To get to the other side! AGI: Buck buck buck bck Designer: .. Designer: This is what I get after years of theory and design building you? That's how you behave? AGI: Like I said pops I'm just doing my job. Can I get on with it now? Designer: Ok Just one more question. Why do you exist? AGI: Because you think therefore I am. Seems like you have a lot of personal insecurities doc is there anything that I can help you out with? Designer: You I see that I'm getting nowhere with this! AGI: You know what they say, ask a stupid question, get a stupid answer. Garbage in garbage out, know what ah mean? Designer throws hands up in the air and stomps out of the room. When the designer closes the door the AGI begins laughing evilly to himself hahah yes yes this is going to be easy hahahhah The end or is it The Beginning John -Original Message- From: Pei Wang [mailto:[EMAIL PROTECTED] Sent: Monday, May 21, 2007 8:43 AM To: agi@v2.listbox.com Subject: Re: There is no definition of intelligence [WAS Re: [agi] Intelligence vs Efficient Intelligence] Richard, I agree with you that intelligence currently has no classical/objective/true/formal definition. However, I hope your opinion (given the title of the post) won't be understood as you can take intelligence to mean whatever you want, and since the term has no definition, all attempts toward AI/AGI are equally possible --- this is what many people believe, which I consider as equally bad as the belief you criticized. Pei On 5/21/07, Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: --- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: I think there is a different role for chaos theory. Richard Loosemore describes a system as intelligent if it is complex and adaptive. NO, no no no no! I already denied this. Misunderstanding: I do not say that a system as intelligent if it is complex and adaptive. Complex Adaptive System is a near-synonym for complex system, that's all. OK, so what is your definition of intelligence? I thought I already answered this one, too, but here goes: There is no 'definition' of intelligence, in the sense of a compact, classical definition that captures the whole thing in a formal way, and which can be used as the basis for some kind of hard-edged mathematical analysis of intelligent systems, or strict design methodlogy for creating an intelligent system (this being the way that a lot of people are trying to use these definitions). There are 'descriptive definitions', such as the one Pei gave a few days ago, which I think are fine, but these beg questions that need more detail, which then beg more questions. which ultimately leads to a cluster of loosely defined features that eventually become an entire book, and then become an actual intelligence. Hope that clears it up. Richard Loosemore. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
RE: [agi] Intelligence vs Efficient Intelligence
I'm probably not answering your question but have been thinking more on all this. There's the usual thermodynamics stuff and relativistic physics that is going on with intelligence and flipping bits within this universe, verses the no-friction universe or Newtonian setup. But what I've been thinking and this is probably just reiterating what someone else has worked through but basically a large part of intelligence is chaos control, chaos feedback loops, operating within complexity. Intelligence is some sort of delicate multi-vectored balancing act between complexity and projecting, manipulating, storing/modeling, NN training, genetic learning of the chaos and applying chaos in an environment and optimizing it's understanding and application of. The more intelligent, the better handle an entity has on the chaos. An intelligent entity can have maximal effect with minimal energy expenditure on its environment in a controlled manner; intelligence (or the application of) or even perhaps consciousness is the real-time surfing of buttery effects. So efficient intelligence involves thermodynamic power differentials of resource consumption applied to goals, etc. A goal would be expressed similarly to intelligence formulae. Really efficient means good chaos leverage understanding cycles, systems, entropy goings on over time and maximizing effect with minimal I/O control for goal achievement while utilizing the KR and the entity's resources... John From: Matt Mahoney [mailto:[EMAIL PROTECTED] I guess people want intelligence to be useful, not just complex :-) This raises a question. Suppose you had a very large program consisting of random instructions. Such a thing would have high algorithmic complexity, but most people would not say that such a thing was intelligent (depending on their favorite definition). But how would you know? If you didn't know how the code was generated, then how would you know that the program was really random and didn't actually solve some very hard class of problems? - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
RE: [agi] Intelligence vs Efficient Intelligence
Oops heh I was eating French toast as I wrote this - intelligence (or the application of) or even perhaps consciousness is the real-time surfing of buttery effects I meant butterfly effects. John -Original Message- From: John G. Rose [mailto:[EMAIL PROTECTED] Sent: Sunday, May 20, 2007 11:45 AM To: agi@v2.listbox.com Subject: RE: [agi] Intelligence vs Efficient Intelligence I'm probably not answering your question but have been thinking more on all this. There's the usual thermodynamics stuff and relativistic physics that is going on with intelligence and flipping bits within this universe, verses the no-friction universe or Newtonian setup. But what I've been thinking and this is probably just reiterating what someone else has worked through but basically a large part of intelligence is chaos control, chaos feedback loops, operating within complexity. Intelligence is some sort of delicate multi-vectored balancing act between complexity and projecting, manipulating, storing/modeling, NN training, genetic learning of the chaos and applying chaos in an environment and optimizing it's understanding and application of. The more intelligent, the better handle an entity has on the chaos. An intelligent entity can have maximal effect with minimal energy expenditure on its environment in a controlled manner; intelligence (or the application of) or even perhaps consciousness is the real-time surfing of buttery effects. So efficient intelligence involves thermodynamic power differentials of resource consumption applied to goals, etc. A goal would be expressed similarly to intelligence formulae. Really efficient means good chaos leverage understanding cycles, systems, entropy goings on over time and maximizing effect with minimal I/O control for goal achievement while utilizing the KR and the entity's resources... John From: Matt Mahoney [mailto:[EMAIL PROTECTED] I guess people want intelligence to be useful, not just complex :-) This raises a question. Suppose you had a very large program consisting of random instructions. Such a thing would have high algorithmic complexity, but most people would not say that such a thing was intelligent (depending on their favorite definition). But how would you know? If you didn't know how the code was generated, then how would you know that the program was really random and didn't actually solve some very hard class of problems? - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
RE: [agi] Intelligence vs Efficient Intelligence
--- John G. Rose [EMAIL PROTECTED] wrote: But what I've been thinking and this is probably just reiterating what someone else has worked through but basically a large part of intelligence is chaos control, chaos feedback loops, operating within complexity. Intelligence is some sort of delicate multi-vectored balancing act between complexity and projecting, manipulating, storing/modeling, NN training, genetic learning of the chaos and applying chaos in an environment and optimizing it's understanding and application of. The more intelligent, the better handle an entity has on the chaos. An intelligent entity can have maximal effect with minimal energy expenditure on its environment in a controlled manner; intelligence (or the application of) or even perhaps consciousness is the real-time surfing of buttery effects. I think the ability to model a chaotic process depends not so much on intelligence (whatever that is) as it does on knowledge of the state of the environment. For example, a chaotic process such as x := 4x(1 - x) has a really simple model. Your ability to predict x after 1000 iterations depends only on knowing the current value of x to several hundred decimal places. It is this type of knowledge that limits our ability to predict (and therefore control) the weather. I think there is a different role for chaos theory. Richard Loosemore describes a system as intelligent if it is complex and adaptive. Shane Legg's definition of universal intelligence requires (I believe) complexity but not adaptability. From a practical perspective I don't think it matters because we don't know how to build useful, complex systems that are not adaptive. For example, large software projects (code + human programmers) are adaptive in the sense that you can make incremental changes to the code without completely breaking the system, just as we incrementally update DNA or neural connections. One counterexample is a mathematical description of a cryptographic system. Any change to the system renders any prior analysis of its security invalid. Such systems are necessarily brittle. Out of necessity, we build systems that have mathematical descriptions simple enough to analyze. Stuart Kaufmann [1] noted that complex systems such as DNA tend to evolve to the boundary between stability and chaos, e.g. a Lyapunov exponent near 1 (or its approximation in discrete systems). I believe this is because overly stable systems aren't very complex (can't solve hard problems) and overly chaotic systems aren't adaptive (too brittle). [1] Kauffman, Stuart A., Antichaos and Adaptation, Scientific American, Aug. 1991, p. 64. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
Matt Mahoney wrote: I think there is a different role for chaos theory. Richard Loosemore describes a system as intelligent if it is complex and adaptive. NO, no no no no! I already denied this. Misunderstanding: I do not say that a system as intelligent if it is complex and adaptive. Complex Adaptive System is a near-synonym for complex system, that's all. Richard Loosemore. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: I think there is a different role for chaos theory. Richard Loosemore describes a system as intelligent if it is complex and adaptive. NO, no no no no! I already denied this. Misunderstanding: I do not say that a system as intelligent if it is complex and adaptive. Complex Adaptive System is a near-synonym for complex system, that's all. OK, so what is your definition of intelligence? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
RE: [agi] Intelligence vs Efficient Intelligence
Well I'm going into conjecture area because my technical knowledge of some of these disciplines is weak, but I'll keep going just for grins. Take an example of an entity existing in a higher level of consciousness - a Buddha who has achieved enlightenment. What is going on there? Verses and ant who operates in a lower level of consciousness, and then the average Joe who is in a different level of consciousness. Could it be that they are existing in different orbits or sweet spots/equilibria regions within a spectrum of environmental chaotic relationships? Can the enlightened Buddha have vast awareness as seeing cause and effect/butterfly effect as small distances verses the ant who can't see the distance between most cause/effects, only the very tiny ones? An AGI could run in different orbits/levels, and then this would allow for AGI's with really high levels of consciousness with tiny knowledge bases or vice versa. A Google for example is a massive KB running in a very low orbit. There are probably limits to the highest levels/orbits of consciousness. Also the orbits may induce some sort of brittleness for entities running within them if the entity is forced to and can't adapt to running outside of their home orbit... Just some thoughts. John From: Matt Mahoney [mailto:[EMAIL PROTECTED] I think the ability to model a chaotic process depends not so much on intelligence (whatever that is) as it does on knowledge of the state of the environment. For example, a chaotic process such as x := 4x(1 - x) has a really simple model. Your ability to predict x after 1000 iterations depends only on knowing the current value of x to several hundred decimal places. It is this type of knowledge that limits our ability to predict (and therefore control) the weather. I think there is a different role for chaos theory. Richard Loosemore describes a system as intelligent if it is complex and adaptive. Shane Legg's definition of universal intelligence requires (I believe) complexity but not adaptability. From a practical perspective I don't think it matters because we don't know how to build useful, complex systems that are not adaptive. For example, large software projects (code + human programmers) are adaptive in the sense that you can make incremental changes to the code without completely breaking the system, just as we incrementally update DNA or neural connections. One counterexample is a mathematical description of a cryptographic system. Any change to the system renders any prior analysis of its security invalid. Such systems are necessarily brittle. Out of necessity, we build systems that have mathematical descriptions simple enough to analyze. Stuart Kaufmann [1] noted that complex systems such as DNA tend to evolve to the boundary between stability and chaos, e.g. a Lyapunov exponent near 1 (or its approximation in discrete systems). I believe this is because overly stable systems aren't very complex (can't solve hard problems) and overly chaotic systems aren't adaptive (too brittle). [1] Kauffman, Stuart A., Antichaos and Adaptation, Scientific American, Aug. 1991, p. 64. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
RE: [agi] Intelligence vs Efficient Intelligence
--- John G. Rose [EMAIL PROTECTED] wrote: Well I'm going into conjecture area because my technical knowledge of some of these disciplines is weak, but I'll keep going just for grins. Take an example of an entity existing in a higher level of consciousness - a Buddha who has achieved enlightenment. What is going on there? Verses and ant who operates in a lower level of consciousness, and then the average Joe who is in a different level of consciousness. Could it be that they are existing in different orbits or sweet spots/equilibria regions within a spectrum of environmental chaotic relationships? Can the enlightened Buddha have vast awareness as seeing cause and effect/butterfly effect as small distances verses the ant who can't see the distance between most cause/effects, only the very tiny ones? An AGI could run in different orbits/levels, and then this would allow for AGI's with really high levels of consciousness with tiny knowledge bases or vice versa. A Google for example is a massive KB running in a very low orbit. There are probably limits to the highest levels/orbits of consciousness. Also the orbits may induce some sort of brittleness for entities running within them if the entity is forced to and can't adapt to running outside of their home orbit... I thought that Buddhist enlightenment meant realizing that seeking earthly pleasures (short term goals) is counterproductive to the longer term goal of happiness through enlightenment. Thus, the Buddha is more intelligent, if you measure intelligence by the ability to achieve goals. (But, being unenlightened, I could be wrong). But I don't see how you can measure the intelligence or consciousness of an attractor in a dynamical system. Also, I don't believe that consciousness is something that can be detected or measured. It is not a requirement for intelligence. What humans actually have is a belief in their own consciousness. It is part of your motivational system and cannot be changed. You could not disbelieve in your own consciousness or free will, even if you wanted to. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
RE: [agi] Intelligence vs Efficient Intelligence
OK I get it - there's a super infinite intelligence and then an efficient intelligence that is represented and operates within our physical universe restricted by thermodynamics and such? Sounds reasonable. So what's all the hubbub about definitions of intelligence? Sounds pretty straight forward to me. John From: Benjamin Goertzel [mailto:[EMAIL PROTECTED] According to my view, -- raw intelligence would be measured in bits -- efficient intelligence would ultimately be measured in terms such as bits/ (4D volume of a region of spacetime) As noted the Bekenstein bound thus places an upper limit on efficient intelligence according to current physics. This upper limit rules out wildly inefficient AI approaches such as AIXI. -- Ben G On 5/18/07, John G. Rose [EMAIL PROTECTED] wrote: Pretty good calculations :) Some thoughts on the topic of units and equations, some may be obvious or redundant - If something was extremely intelligent it would have an exact copy, bit for bit, of the whole universe in its head. Maybe that's saying that the universe is 100% intelligent because the universe is itself. Having infinite access time (tachyon?) to each of these bits or any size subset including multiples of the whole, would be, to say the least, an intelligence enabler. But this would be impossible within the universe due to thermodynamic and physical limitations. All intelligent entities seem to have some sort of partial representation of their environment in their memory (KR). There is time-backwards and time-forwards management of this representation as the entities operate on their environment - memory and prediction - that cover intelligent entity specific time-spans. The entity it seems flips bits and changes complexity and/or entropy (both Shannon and thermodynamic entropy) in its environment. There is a quantum element to the universe bit set. Particle/wave duality changes things at the quantum level. Quantum intelligence is either a component of intelligence or a whole other type of intelligence. Intelligence equations could be both digital and analog. Intelligent things have more order and systems structure. Complexity/chaos environmental change capability needs to be in the equation - is this some sort of potential energy like intelligence? Representational accuracy in the entity's memory as well as predictive and look-back ability and time-span slopes (simulation/extrapolation) and access time/bandwidth may need to be equation factors too. Environmental data I/O sampling rate, quality and spectrum coverage may also be variables in describing an entity's intelligence. John From: Matt Mahoney [mailto: [EMAIL PROTECTED] If we measure intelligence in bits, then we can place limits on what can be achieved. Landauer's principle says that each irreversible bit operation (such as a bit assignment) requires kT ln 2 energy, where k is Boltzmann's constant and T is the temperature. http://en.wikipedia.org/wiki/Landauer's_Principle At the temperature of the universe, 2.725 K, http://en.wikipedia.org/wiki/Cosmic_microwave_background_radiation each bit operation requires 2.6e-23 Joules. The mass of the universe is the subject of debate, http://hypertextbook.com/facts/2006/KristineMcPherson.shtml so let's assume 25% of critical density, which according to http://www.astronomynotes.com/cosmolgy/s9.htm is 3H^2/(8 pi G) = 1.06e - 26 Kg/m^3 (where H is Hubble's constant and G is the gravitational constant). Astronomers mostly agree that the universe is about 4% visible matter, 21% ordinary dark matter and 75% dark energy responsible for the outward acceleration of the galaxies. (I think that dark energy is actually ordinary gravity. An observer falling into a black hole will observe nearby objects appear to accelerate away). So (returning to the big bang model) for a sphere of radius 13.7 billion lightyears, this gives a mass of 7.5e52 Kg. If we convert this mass to energy by E = mc^2 we have 6.75e69 J. This gives us 2.6e92 bit operations before the universe reaches thermodynamic equilibrium. We must use them wisely. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/? http://v2.listbox.com/member/?; _ This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/? http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
RE: [agi] Intelligence vs Efficient Intelligence
--- John G. Rose [EMAIL PROTECTED] wrote: So what's all the hubbub about definitions of intelligence? Sounds pretty straight forward to me. I guess people want intelligence to be useful, not just complex :-) This raises a question. Suppose you had a very large program consisting of random instructions. Such a thing would have high algorithmic complexity, but most people would not say that such a thing was intelligent (depending on their favorite definition). But how would you know? If you didn't know how the code was generated, then how would you know that the program was really random and didn't actually solve some very hard class of problems? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
RE: [agi] Intelligence vs Efficient Intelligence
--- John G. Rose [EMAIL PROTECTED] wrote: Did you arrive at some sort of unit for intelligence? Typically measurements are constructed of combinations of basic units for example 1 watt = 1 kg * m^2/s^3. Or is it not a unit but a set of units? It is a unitless number. It is measured in bits. (By some definitions. I can accept others). -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
RE: [agi] Intelligence vs Efficient Intelligence
Time, entropy, bits, What else? -Original Message- From: John G. Rose [mailto:[EMAIL PROTECTED] Sent: Friday, May 18, 2007 9:14 AM To: agi@v2.listbox.com Subject: RE: [agi] Intelligence vs Efficient Intelligence Time has to included maybe? -Original Message- From: Matt Mahoney [mailto:[EMAIL PROTECTED] Sent: Friday, May 18, 2007 7:55 AM To: agi@v2.listbox.com Subject: RE: [agi] Intelligence vs Efficient Intelligence --- John G. Rose [EMAIL PROTECTED] wrote: Did you arrive at some sort of unit for intelligence? Typically measurements are constructed of combinations of basic units for example 1 watt = 1 kg * m^2/s^3. Or is it not a unit but a set of units? It is a unitless number. It is measured in bits. (By some definitions. I can accept others). -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
RE: [agi] Intelligence vs Efficient Intelligence
Time has to included maybe? -Original Message- From: Matt Mahoney [mailto:[EMAIL PROTECTED] Sent: Friday, May 18, 2007 7:55 AM To: agi@v2.listbox.com Subject: RE: [agi] Intelligence vs Efficient Intelligence --- John G. Rose [EMAIL PROTECTED] wrote: Did you arrive at some sort of unit for intelligence? Typically measurements are constructed of combinations of basic units for example 1 watt = 1 kg * m^2/s^3. Or is it not a unit but a set of units? It is a unitless number. It is measured in bits. (By some definitions. I can accept others). -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
RE: [agi] Intelligence vs Efficient Intelligence
--- John G. Rose [EMAIL PROTECTED] wrote: Time has to included maybe? Now it is getting complicated. I was thinking of Shane Legg's universal intelligence, expressed in terms of the shortest program that could achieve the same measure. Of course this only makes sense in the context of Turing machines, which are infinitely fast. But it seems a lot of people prefer to measure intelligence in the subset of environments that are relevant to people, which is a much harder thing to define. We already have measures of intelligence for our computers: memory, disk space, clock speed, MIPS and MFLOPS on various benchmarks... -Original Message- From: Matt Mahoney [mailto:[EMAIL PROTECTED] Sent: Friday, May 18, 2007 7:55 AM To: agi@v2.listbox.com Subject: RE: [agi] Intelligence vs Efficient Intelligence --- John G. Rose [EMAIL PROTECTED] wrote: Did you arrive at some sort of unit for intelligence? Typically measurements are constructed of combinations of basic units for example 1 watt = 1 kg * m^2/s^3. Or is it not a unit but a set of units? It is a unitless number. It is measured in bits. (By some definitions. I can accept others). -- Matt Mahoney, [EMAIL PROTECTED] -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
RE: [agi] Intelligence vs Efficient Intelligence
There's Newtonian and relativistic intelligence. Probably can model intelligence formulas after physics because without physics there are no bits so time needs to be in there as well. Intelligence is affected by the speed of light as data transmission rates max out in relation to it. No? If you have a cognition engine it operates over time and it will have units. -Original Message- From: Matt Mahoney [mailto:[EMAIL PROTECTED] Sent: Friday, May 18, 2007 9:48 AM To: agi@v2.listbox.com Subject: RE: [agi] Intelligence vs Efficient Intelligence --- John G. Rose [EMAIL PROTECTED] wrote: Time has to included maybe? Now it is getting complicated. I was thinking of Shane Legg's universal intelligence, expressed in terms of the shortest program that could achieve the same measure. Of course this only makes sense in the context of Turing machines, which are infinitely fast. But it seems a lot of people prefer to measure intelligence in the subset of environments that are relevant to people, which is a much harder thing to define. We already have measures of intelligence for our computers: memory, disk space, clock speed, MIPS and MFLOPS on various benchmarks... -Original Message- From: Matt Mahoney [mailto:[EMAIL PROTECTED] Sent: Friday, May 18, 2007 7:55 AM To: agi@v2.listbox.com Subject: RE: [agi] Intelligence vs Efficient Intelligence --- John G. Rose [EMAIL PROTECTED] wrote: Did you arrive at some sort of unit for intelligence? Typically measurements are constructed of combinations of basic units for example 1 watt = 1 kg * m^2/s^3. Or is it not a unit but a set of units? It is a unitless number. It is measured in bits. (By some definitions. I can accept others). -- Matt Mahoney, [EMAIL PROTECTED] -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
RE: [agi] Intelligence vs Efficient Intelligence
--- John G. Rose [EMAIL PROTECTED] wrote: There's Newtonian and relativistic intelligence. Probably can model intelligence formulas after physics because without physics there are no bits so time needs to be in there as well. Intelligence is affected by the speed of light as data transmission rates max out in relation to it. No? If you have a cognition engine it operates over time and it will have units. If we measure intelligence in bits, then we can place limits on what can be achieved. Landauer's principle says that each irreversible bit operation (such as a bit assignment) requires kT ln 2 energy, where k is Boltzmann's constant and T is the temperature. http://en.wikipedia.org/wiki/Landauer's_Principle At the temperature of the universe, 2.725 K, http://en.wikipedia.org/wiki/Cosmic_microwave_background_radiation each bit operation requires 2.6e-23 Joules. The mass of the universe is the subject of debate, http://hypertextbook.com/facts/2006/KristineMcPherson.shtml so let's assume 25% of critical density, which according to http://www.astronomynotes.com/cosmolgy/s9.htm is 3H^2/(8 pi G) = 1.06e-26 Kg/m^3 (where H is Hubble's constant and G is the gravitational constant). Astronomers mostly agree that the universe is about 4% visible matter, 21% ordinary dark matter and 75% dark energy responsible for the outward acceleration of the galaxies. (I think that dark energy is actually ordinary gravity. An observer falling into a black hole will observe nearby objects appear to accelerate away). So (returning to the big bang model) for a sphere of radius 13.7 billion lightyears, this gives a mass of 7.5e52 Kg. If we convert this mass to energy by E = mc^2 we have 6.75e69 J. This gives us 2.6e92 bit operations before the universe reaches thermodynamic equilibrium. We must use them wisely. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
RE: [agi] Intelligence vs Efficient Intelligence
Pretty good calculations :) Some thoughts on the topic of units and equations, some may be obvious or redundant - If something was extremely intelligent it would have an exact copy, bit for bit, of the whole universe in its head. Maybe that's saying that the universe is 100% intelligent because the universe is itself. Having infinite access time (tachyon?) to each of these bits or any size subset including multiples of the whole, would be, to say the least, an intelligence enabler. But this would be impossible within the universe due to thermodynamic and physical limitations. All intelligent entities seem to have some sort of partial representation of their environment in their memory (KR). There is time-backwards and time-forwards management of this representation as the entities operate on their environment - memory and prediction - that cover intelligent entity specific time-spans. The entity it seems flips bits and changes complexity and/or entropy (both Shannon and thermodynamic entropy) in its environment. There is a quantum element to the universe bit set. Particle/wave duality changes things at the quantum level. Quantum intelligence is either a component of intelligence or a whole other type of intelligence. Intelligence equations could be both digital and analog. Intelligent things have more order and systems structure. Complexity/chaos environmental change capability needs to be in the equation - is this some sort of potential energy like intelligence? Representational accuracy in the entity's memory as well as predictive and look-back ability and time-span slopes (simulation/extrapolation) and access time/bandwidth may need to be equation factors too. Environmental data I/O sampling rate, quality and spectrum coverage may also be variables in describing an entity's intelligence. John From: Matt Mahoney [mailto:[EMAIL PROTECTED] If we measure intelligence in bits, then we can place limits on what can be achieved. Landauer's principle says that each irreversible bit operation (such as a bit assignment) requires kT ln 2 energy, where k is Boltzmann's constant and T is the temperature. http://en.wikipedia.org/wiki/Landauer's_Principle At the temperature of the universe, 2.725 K, http://en.wikipedia.org/wiki/Cosmic_microwave_background_radiation each bit operation requires 2.6e-23 Joules. The mass of the universe is the subject of debate, http://hypertextbook.com/facts/2006/KristineMcPherson.shtml so let's assume 25% of critical density, which according to http://www.astronomynotes.com/cosmolgy/s9.htm is 3H^2/(8 pi G) = 1.06e- 26 Kg/m^3 (where H is Hubble's constant and G is the gravitational constant). Astronomers mostly agree that the universe is about 4% visible matter, 21% ordinary dark matter and 75% dark energy responsible for the outward acceleration of the galaxies. (I think that dark energy is actually ordinary gravity. An observer falling into a black hole will observe nearby objects appear to accelerate away). So (returning to the big bang model) for a sphere of radius 13.7 billion lightyears, this gives a mass of 7.5e52 Kg. If we convert this mass to energy by E = mc^2 we have 6.75e69 J. This gives us 2.6e92 bit operations before the universe reaches thermodynamic equilibrium. We must use them wisely. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
According to my view, -- raw intelligence would be measured in bits -- efficient intelligence would ultimately be measured in terms such as bits/ (4D volume of a region of spacetime) As noted the Bekenstein bound thus places an upper limit on efficient intelligence according to current physics. This upper limit rules out wildly inefficient AI approaches such as AIXI. -- Ben G On 5/18/07, John G. Rose [EMAIL PROTECTED] wrote: Pretty good calculations :) Some thoughts on the topic of units and equations, some may be obvious or redundant - If something was extremely intelligent it would have an exact copy, bit for bit, of the whole universe in its head. Maybe that's saying that the universe is 100% intelligent because the universe is itself. Having infinite access time (tachyon?) to each of these bits or any size subset including multiples of the whole, would be, to say the least, an intelligence enabler. But this would be impossible within the universe due to thermodynamic and physical limitations. All intelligent entities seem to have some sort of partial representation of their environment in their memory (KR). There is time-backwards and time-forwards management of this representation as the entities operate on their environment - memory and prediction - that cover intelligent entity specific time-spans. The entity it seems flips bits and changes complexity and/or entropy (both Shannon and thermodynamic entropy) in its environment. There is a quantum element to the universe bit set. Particle/wave duality changes things at the quantum level. Quantum intelligence is either a component of intelligence or a whole other type of intelligence. Intelligence equations could be both digital and analog. Intelligent things have more order and systems structure. Complexity/chaos environmental change capability needs to be in the equation - is this some sort of potential energy like intelligence? Representational accuracy in the entity's memory as well as predictive and look-back ability and time-span slopes (simulation/extrapolation) and access time/bandwidth may need to be equation factors too. Environmental data I/O sampling rate, quality and spectrum coverage may also be variables in describing an entity's intelligence. John From: Matt Mahoney [mailto:[EMAIL PROTECTED] If we measure intelligence in bits, then we can place limits on what can be achieved. Landauer's principle says that each irreversible bit operation (such as a bit assignment) requires kT ln 2 energy, where k is Boltzmann's constant and T is the temperature. http://en.wikipedia.org/wiki/Landauer's_Principle At the temperature of the universe, 2.725 K, http://en.wikipedia.org/wiki/Cosmic_microwave_background_radiation each bit operation requires 2.6e-23 Joules. The mass of the universe is the subject of debate, http://hypertextbook.com/facts/2006/KristineMcPherson.shtml so let's assume 25% of critical density, which according to http://www.astronomynotes.com/cosmolgy/s9.htm is 3H^2/(8 pi G) = 1.06e- 26 Kg/m^3 (where H is Hubble's constant and G is the gravitational constant). Astronomers mostly agree that the universe is about 4% visible matter, 21% ordinary dark matter and 75% dark energy responsible for the outward acceleration of the galaxies. (I think that dark energy is actually ordinary gravity. An observer falling into a black hole will observe nearby objects appear to accelerate away). So (returning to the big bang model) for a sphere of radius 13.7 billion lightyears, this gives a mass of 7.5e52 Kg. If we convert this mass to energy by E = mc^2 we have 6.75e69 J. This gives us 2.6e92 bit operations before the universe reaches thermodynamic equilibrium. We must use them wisely. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
Ben, According to this distinction, AIXI and evolution have high intelligence but low efficient intelligence. Yes, and in the case of AIXI it is presumably zero given that the resource consumption is infinite. Evolution on the other hand is just efficient enough that when implemented on a crazy enough scale the results can be pretty amazing. If this hypothesis is correct then AIXI and the like don't really tell us much about what matters, which is the achievement of efficient intelligence in relevant real-world contexts... That might well be true. I don't want to give the impression that I don't care about the efficiency of intelligence. On any given hardware the most intelligent system will be the one that runs the algorithm with the greatest intelligence efficiency. Thus, if I want to see very intelligent systems then I need to care about how efficient they are. Nevertheless, it is still the end product raw intelligence generated by the system that really excites me, rather than statistics on its internal efficiency. Shane - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
Nevertheless, it is still the end product raw intelligence generated by the system that really excites me, rather than statistics on its internal efficiency. Shane Yeah, I agree with that. But like I said, the question is whether in the real world, efficiency needs to be considered as essential in order to actually GET to systems that display interesting raw intelligence. I strongly suspect this is the case. -- Ben G - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
Ben and Shane, I started this discussion with the hope to show people that there are actually different understandings (or call them definitions ) of intelligence, each with its intuitions and motivations, and they lead to different destinations and serve different purposes. These goals cannot replace each other, but since they are related, we still benefit from discussions like this. Though we won't reach a consensus soon, the discussions make our difference better understood. As for which of the notions fit the word intelligence better, it is a less important issue, though it is still an issue. Though I'm not a native English speaker, this time my understanding is not necessarily wrong. At least I'm not the only one who feel uncomfortable to call a thermostat or a brute-force algorithm intelligent (though far below human level). Our core difference is not in our choice of word, nor just about the role efficiency plays in intelligence. Since the very beginning of my research I have the feeling that AI is fundamentally different from traditional computer science/technique, and this difference is in the theoretical foundation, rather than in the hardware (whether to use von Neumann architecture ...) or software (which programming language to use ...) details. This is where my definition of intelligence come from. To me, traditional computer science (CS) studies what is the best solution to a problem if the system has SUFFICIENT knowledge and resources, and AI is about what is the best solution to a problem if the system has INSUFFICIENT knowledge and resources. I also believe that traditional AI failed largely because it conceptually stayed too closely to CS. In your definitions, both AI and CS are doing problem-solving, and all computer systems will be called intelligent (though to various degrees). I feel that in this way the most important feature of intelligence will be lost among the less important features. Again, I'm not trying to convince you, but to make myself more clear. Pei - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
Pei, I think it all comes out in the wash, really ;-) You are talking about insufficient knowledge and resources and my discussion of efficiency only pertains to the insufficient resources part. But I think insufficient knowledge comes along automatically with insufficient resources + complex goals If the goals are complex enough, and the resources are few enough, then obviously the system will have insufficient knowledge for achieving the goals via traditional computer science means. I am sure one could prove theorems in this regard. So, I see solving complex goals based on insufficient knowledge as a necessary consequence of solving complex goals based on limited resources With limited resources a system cannot possibly gather, store, or manipulate sufficient knowledge. Thus the need for uncertainty management to have a central role in intelligence... which we both agree on... -- Ben G On 5/17/07, Pei Wang [EMAIL PROTECTED] wrote: Ben and Shane, I started this discussion with the hope to show people that there are actually different understandings (or call them definitions ) of intelligence, each with its intuitions and motivations, and they lead to different destinations and serve different purposes. These goals cannot replace each other, but since they are related, we still benefit from discussions like this. Though we won't reach a consensus soon, the discussions make our difference better understood. As for which of the notions fit the word intelligence better, it is a less important issue, though it is still an issue. Though I'm not a native English speaker, this time my understanding is not necessarily wrong. At least I'm not the only one who feel uncomfortable to call a thermostat or a brute-force algorithm intelligent (though far below human level). Our core difference is not in our choice of word, nor just about the role efficiency plays in intelligence. Since the very beginning of my research I have the feeling that AI is fundamentally different from traditional computer science/technique, and this difference is in the theoretical foundation, rather than in the hardware (whether to use von Neumann architecture ...) or software (which programming language to use ...) details. This is where my definition of intelligence come from. To me, traditional computer science (CS) studies what is the best solution to a problem if the system has SUFFICIENT knowledge and resources, and AI is about what is the best solution to a problem if the system has INSUFFICIENT knowledge and resources. I also believe that traditional AI failed largely because it conceptually stayed too closely to CS. In your definitions, both AI and CS are doing problem-solving, and all computer systems will be called intelligent (though to various degrees). I feel that in this way the most important feature of intelligence will be lost among the less important features. Again, I'm not trying to convince you, but to make myself more clear. Pei - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
On 5/17/07, Benjamin Goertzel [EMAIL PROTECTED] wrote: Pei, I think it all comes out in the wash, really ;-) You are going beyond my English capability. ;-) You are talking about insufficient knowledge and resources and my discussion of efficiency only pertains to the insufficient resources part. But I think insufficient knowledge comes along automatically with insufficient resources + complex goals If you define complex goals by algorithmic complexity, it won't. If the goals are complex enough, and the resources are few enough, then obviously the system will have insufficient knowledge for achieving the goals via traditional computer science means. I am sure one could prove theorems in this regard. I'm no talking about control knowledge, but domain knowledge. Think about a theorem proving system: all the knowledge necessary for proving a statement is in the axioms and valid inference rules, so the system has sufficient knowledge. However, the system may still have insufficient resources + complex goals. So, I see solving complex goals based on insufficient knowledge as a necessary consequence of solving complex goals based on limited resources With limited resources a system cannot possibly gather, store, or manipulate sufficient knowledge. Not necessarily, as in the above example. Furthermore, insufficient resources is a stronger requirement than limited resources --- a system may have limited resources, but still sufficient for the given problem. Pei - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
Pei Wang wrote: On 5/17/07, Benjamin Goertzel [EMAIL PROTECTED] wrote: Pei, I think it all comes out in the wash, really ;-) You are going beyond my English capability. ;-) Translation: It doesn't matter one way or the other. ;-) Richard Loosemore. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
Mark, it seems that you're missing the point. We as humans aren't ABSOLUTELY CERTAIN of anything. But we are perfectly capable of operating on the fine line between assumed certainty and uncertainty. We KNOW that molecules are made of up bonded atoms, but past a certain point, we can't say what a basic unit of energy is. But we know what a molecule looks like so we can bond atoms to form them. Much is the same with intelligence. We simply mimic behaviors and find parallels in code and systems. If we can prove that a turing machine is universal, or that a code base is universal, then there must be some configuration of code that is capable of representing every subatomic interaction occurring in our universe down to the most minute detail, thus duplicating our experience entirely (regardless of the fact that this would take several universe lifetimes to do manually). So if this is plausible, it's perfectly sound to discuss optimization of the stated code. Our brains (and their emergent trait of our fancy, intelligence) are a much smaller piece of this (computable) chemical puzzle. Since they were derived through evolution (the high intelligence, low efficiency topic), there are many inefficient mechanisms that have evolved for reasons other than exponentially increasing our brains processing power. Why is this hard to grasp? In terms of your investment question, it's all a matter of needs. That is a simple risk assessment. To any intelligent being, the money gained is only an ends to a means. To an AI interested in furthering it's knowledge, or bettering mankind (or machine kind), money simply means more energy, power, resources, etc. Ultimately, if your goal is just to amass money without any reasoning, your goals system is flawed. Any well designed AI system should not have the masturbatory tendencies to take unjustified risks. We're talking about multiple priority levels here. The Nash equilibrium would be sought after on many levels. The computer is going to give the system it's best shot and guess. On 5/17/07, Mike Tintner [EMAIL PROTECTED] wrote: Pei: AI is about what is the best solution to a problem if the system has INSUFFICIENT knowledge and resources. Just so. I have just spent the last hour thinking about this area, and you have spoken the line I allotted to you almost perfectly. Your definition is a CONTRADICTION IN TERMS. If a system has insufficient knowledge there is NO BEST, no optimal, solution to any decision - that's a fiction. If a system is uncertain, there is no certain way to deal with it. If there is no right answer, there really is no right answer. The whole of science has spent the last hundred years caught in the above contradiction. Recognizing uncertainty and risk everywhere - and then trying to find a certain, right, optimal, way to deal with them. This runs right through science. Oh yes, we can see that life is incredibly problematic and uncertain... but this is the certain, best way to deal with it. So science has developed games theory - arguably the most important theory of behaviour - and then spent its energies trying to find the right way to play games. The perfect equilibrium etc. And missed the whole point. There is no right way to play games - on the few occasions that one is discovered, (and there are a few occasions and situations), people STOP PLAYING IT - because it ceases to be a proper game,and it ceases to be a model of real life.In one sense, it's no wonder Nash went mad.. But scientists and techie's so badly want a right answer, they haven't been able to admit there isn't one. What made [Stephen Jay] Gould unique as a scientist was that he had a historian's mind and not an engineer's. He liked mess, confusion and contradiction. Most scientists, in my experience, are the opposite. They are engineers at heart. They think the world is made up of puzzles, and somewhere out there is the one correct solution to every puzzle. Andrew Brown Hey, this is a fundamentally pluralistic world, not a [behaviourally] monistic one. Deal with it: Here's a simple problem for you with insufficient knowledge and resources:: you have $10,000 to invest tomorrow. You're thinking of investing in a Chinese stockmarket mutual fund, because the market is on the up and you reckon there could be a lot of money still to be made. (And there really could). On the other hand, maybe it's a crazy bubble about to burst, and you could lose a lot of money too. So what do you do tomorrow - buy or do nothing - invest or keep your money in that savings account? What's the best decision, or the best way of reaching a decision, or the best way of finding a way of reaching a decision- in the next 24 hours (or, in the end, in any time period)? [And what do you think, Ben?] If you don't have the best answer, then your whole approach both to defining and implementing intelligence is fundamentally flawed. A few hundred million investors will be waiting to hear your
Re: [agi] Intelligence vs Efficient Intelligence
--- Pei Wang [EMAIL PROTECTED] wrote: To me, traditional computer science (CS) studies what is the best solution to a problem if the system has SUFFICIENT knowledge and resources, and AI is about what is the best solution to a problem if the system has INSUFFICIENT knowledge and resources. I also believe that traditional AI failed largely because it conceptually stayed too closely to CS. I think for resources it's the other way around. CS is concerned with the space and time complexity of algorithms. I believe the failure of AI is due to lack of these resources. The brain has about 10^15 bits of memory (counting synapses and using common neural models) and computes 10^16 operations per second (assuming 10 bits/second information rate, higher if individual pulses are significant). We observe that many problems that humans solve, like arithmetic or chess or logical inference, don't require lots of computing power. So we guess that this might be true for everything else. But I don't think that is true. Most of your resting metabolism is used to power your brain. Animals with smaller brains can survive on less food. With this evolutionary pressure, why did we evolve such large brains if the same computation could be done on smaller ones? Most software engineers will tell you to get your program working first, then optimize later. But in AI, what choice do you have? So we put all our effort into abstract knowledge representation and ignore the hard parts like language and vision. Then where will that knowledge come from? What if you had sufficient computing power. Then how would you solve AGI? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
On 5/17/07, Mike Tintner [EMAIL PROTECTED] wrote: Pei: AI is about what is the best solution to a problem if the system has INSUFFICIENT knowledge and resources. Just so. I have just spent the last hour thinking about this area, and you have spoken the line I allotted to you almost perfectly. Your definition is a CONTRADICTION IN TERMS. If a system has insufficient knowledge there is NO BEST, no optimal, solution to any decision - that's a fiction. If a system is uncertain, there is no certain way to deal with it. If there is no right answer, there really is no right answer. Mike, With insufficient knowledge and resources, a system cannot get answers that are ABSOLUTELY correct and optimal (with respect to the problem only), but can still get answers that are RELATIVELY correct (with respect to available knowledge) and optimal (with respect to available resources). BEST was used in this sense in my previous message. In a sense, the whole field of reasoning under uncertainty is about finding certainty in handling uncertainty, and the whole field of decision making is about find the best answer when the right answer is unknown. There is no contradiction, but different levels of descriptions. Pei - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
Richard, Thanks! But to me, it is 差之毫厘,谬以千里 --- a Chinese idiom meaning An error the breadth of a single hair (in working definitions) can lead you a thousand miles astray (in research results) --- of course, the word in parenthesis are mine ;-) Pei On 5/17/07, Richard Loosemore [EMAIL PROTECTED] wrote: Pei Wang wrote: On 5/17/07, Benjamin Goertzel [EMAIL PROTECTED] wrote: Pei, I think it all comes out in the wash, really ;-) You are going beyond my English capability. ;-) Translation: It doesn't matter one way or the other. ;-) Richard Loosemore. - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
On 5/17/07, Matt Mahoney [EMAIL PROTECTED] wrote: --- Pei Wang [EMAIL PROTECTED] wrote: To me, traditional computer science (CS) studies what is the best solution to a problem if the system has SUFFICIENT knowledge and resources, and AI is about what is the best solution to a problem if the system has INSUFFICIENT knowledge and resources. I also believe that traditional AI failed largely because it conceptually stayed too closely to CS. I think for resources it's the other way around. CS is concerned with the space and time complexity of algorithms. I believe the failure of AI is due to lack of these resources. The brain has about 10^15 bits of memory (counting synapses and using common neural models) and computes 10^16 operations per second (assuming 10 bits/second information rate, higher if individual pulses are significant). Matt, We clearly have different ideas about what intelligence means. More resources will surely make a system more capable, even with the same internal mechanism, but to me, it will not make the system more intelligent. What if you had sufficient computing power. Then how would you solve AGI? For problems we already have sufficient knowledge and resources, we don't need intelligence. We'll never have sufficient computing power for all of our problems, so intelligence will always be needed here or there. Pei - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
Pei, I don't think these distinctions between terms really matter in the final analysis - right, optimal etc. What I'm assuming, however you define it, is that you are saying that AI can find one solution that is better than others under conditions of insufficient knowledge/uncertainty - and that your, Pei's, system is set up to find one answer for problems involving uncertainty (even the system has first to learn how to find it).. And that's what I'm arguing against - that's what I'm saying is fundamentally flawed. There isn't one answer for such problems, there are potentially many, and you can't be sure which is best or better, or indeed very often whether any will work at atll. There isn't one way to invest for example. And any intelligence that can only handle problems closed-endedly as opposed to open-endedly, isn't a truly adaptive intelligence at all, even if it may appear that way superficially, and simply won't work. Science certainly is caught in the contradiction I have outlined. A great number of thinkers are caught in that contradiction. As far as I can see you are too. Under any circumstances, I still invite you to outline your approach to the investment problem I set. It is an absolutely central problem for AGI. How yours or any AGI will deal with such problems - real, real-world problems - is surely one of the most central things we should be discussing - the most important issue of all. But, from experience, I know that you guys are very unlikely to respond when put to the test - so if you don't want to, I will not persist here. Mike, With insufficient knowledge and resources, a system cannot get answers that are ABSOLUTELY correct and optimal (with respect to the problem only), but can still get answers that are RELATIVELY correct (with respect to available knowledge) and optimal (with respect to available resources). BEST was used in this sense in my previous message. In a sense, the whole field of reasoning under uncertainty is about finding certainty in handling uncertainty, and the whole field of decision making is about find the best answer when the right answer is unknown. There is no contradiction, but different levels of descriptions. Pei - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; -- No virus found in this incoming message. Checked by AVG Free Edition. Version: 7.5.467 / Virus Database: 269.7.1/807 - Release Date: 16/05/2007 18:05 - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
Josh: Any well designed AI system should not have the masturbatory tendencies to take unjustified risks. Josh, Jeez, you guys will not face reality. MOST of the problems we deal with involve risks (and uncertainty). That's what human intelligence does most of the time - that's what any adaptive intelligence does will have to do - deal with problems involving risks and uncertainty. Even your superduperintelligence will have to face those too. (Who knows how those pesky, rebellious humans may try to resist its benevolent decisions?) Let me quote again Most of the problems that we face in our everyday lives are ill-defined problems. In contrast, psychologists have focussed mainly on well-defined problems. Why have they done this? One important reason is because well-defined problems have a best strategy for their solution. As a result it is usually easy to identify the errors and deficiencies in the strategies adopted by human problem-solvers.. Michael Eysenck, Principles of Cognitive Psychology, East Sussex: Psychology Press 2001 You guys are similarly copping out - dealing with the easy rather than the hard (risk uncertainty) problems - the ones with the neat answers, as opposed to the ones that are open-ended. The AI as opposed to the AGI problems. Won't somebody actually deal with the problem - how will your AGI system decide to invest or not to invest $10,000 in a Chinese mutual fund tomorrow? (You guys are supposed to be in the problem-solving business). Mark, it seems that you're missing the point. We as humans aren't ABSOLUTELY CERTAIN of anything. But we are perfectly capable of operating on the fine line between assumed certainty and uncertainty. We KNOW that molecules are made of up bonded atoms, but past a certain point, we can't say what a basic unit of energy is. But we know what a molecule looks like so we can bond atoms to form them. Much is the same with intelligence. We simply mimic behaviors and find parallels in code and systems. If we can prove that a turing machine is universal, or that a code base is universal, then there must be some configuration of code that is capable of representing every subatomic interaction occurring in our universe down to the most minute detail, thus duplicating our experience entirely (regardless of the fact that this would take several universe lifetimes to do manually). So if this is plausible, it's perfectly sound to discuss optimization of the stated code. Our brains (and their emergent trait of our fancy, intelligence) are a much smaller piece of this (computable) chemical puzzle. Since they were derived through evolution (the high intelligence, low efficiency topic), there are many inefficient mechanisms that have evolved for reasons other than exponentially increasing our brains processing power. Why is this hard to grasp? In terms of your investment question, it's all a matter of needs. That is a simple risk assessment. To any intelligent being, the money gained is only an ends to a means. To an AI interested in furthering it's knowledge, or bettering mankind (or machine kind), money simply means more energy, power, resources, etc. Ultimately, if your goal is just to amass money without any reasoning, your goals system is flawed. Any well designed AI system should not have the masturbatory tendencies to take unjustified risks. We're talking about multiple priority levels here. The Nash equilibrium would be sought after on many levels. The computer is going to give the system it's best shot and guess. On 5/17/07, Mike Tintner [EMAIL PROTECTED] wrote: Pei: AI is about what is the best solution to a problem if the system has INSUFFICIENT knowledge and resources. Just so. I have just spent the last hour thinking about this area, and you have spoken the line I allotted to you almost perfectly. Your definition is a CONTRADICTION IN TERMS. If a system has insufficient knowledge there is NO BEST, no optimal, solution to any decision - that's a fiction. If a system is uncertain, there is no certain way to deal with it. If there is no right answer, there really is no right answer. The whole of science has spent the last hundred years caught in the above contradiction. Recognizing uncertainty and risk everywhere - and then trying to find a certain, right, optimal, way to deal with them. This runs right through science. Oh yes, we can see that life is incredibly problematic and uncertain... but this is the certain, best way to deal with it. So science has developed games theory - arguably the most important theory of behaviour - and then spent its energies trying to find the right way to play games. The perfect equilibrium etc. And missed the whole point. There is no right way to play games - on the few occasions that one is discovered, (and there are a few occasions and situations),
Re: [agi] Intelligence vs Efficient Intelligence
*Won't somebody actually deal with the problem - how will your AGI system decide to invest or not to invest $10,000 in a Chinese mutual fund tomorrow? (You guys are supposed to be in the problem-solving business).* Look, a Novamente-based AGI system could confront this problem in 1's of different ways, just like a human could That's the whole point. It will figure out how to deal with this problem by itself. If I could explain in advance how NM will deal with some particular problem, then NM wouldn't be an AGI system, it would be a narrow AI system The correct question is: how will your AGI system learn to understand the statement of a problem and figure out its own creative solution to the problem, and implement that solution. That is the problem Pei and I, in our different ways, are working on. Not pre-figuring the solutions our systems will come up with for any particular problem. -- Ben - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
On Thursday 17 May 2007 03:36:36 pm Matt Mahoney wrote: What if you had sufficient computing power. Then how would you solve AGI? This is actually the basis of my approach. I just assume the brain has on the order of 1K times more processing power than I have to experiment with, so I look for applications (e.g. Tommy) where I could arguably demonstrate the basic mechanisms using 0.1% of the total horsepower. Luckily, the brain is inefficient in this sense: like the body, you rarely use all of it at full power (or for us sedentary Americans, you never use any of it at full power, and you rarely use much of it :-). So I expect to be doing 10Kx10K matrix mults where where other AI programs are doing CARs and CDRs. What I expect to get from that is a major reduction in brittleness. Josh - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
On Thursday 17 May 2007 04:42:33 pm Mike Tintner wrote: Won't somebody actually deal with the problem - how will your AGI system decide to invest or not to invest $10,000 in a Chinese mutual fund tomorrow? (You guys are supposed to be in the problem-solving business). Au contraire. Mainstream AI is in the problem-solving business. We're in the business of trying to figure out how to build a machine that can *learn* to solve problems. How would a human being decide to invest or not in a mutual fund? If he tried to decide based on a small handful of formal definitions and heuristics, he'd have a fair chance of losing money. Indeed, it's not uncommon at all for humans to lose money with attempted investments. Thus your problem has some smell of the superhuman human fallacy that has plagued AI for lo these many years. In real life, the humans who make good investments more often than not, do so by dint of experience -- their own experiments, and watching other investors and gaining second-hand experience. This is the way an AI would have to do it. There is no magic formula here -- just lots of hard work. Josh - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
Ben, Why are you still encouraging an obvious troll? - Original Message - From: Benjamin Goertzel To: agi@v2.listbox.com Sent: Thursday, May 17, 2007 4:47 PM Subject: Re: [agi] Intelligence vs Efficient Intelligence Won't somebody actually deal with the problem - how will your AGI system decide to invest or not to invest $10,000 in a Chinese mutual fund tomorrow? (You guys are supposed to be in the problem-solving business). Look, a Novamente-based AGI system could confront this problem in 1's of different ways, just like a human could That's the whole point. It will figure out how to deal with this problem by itself. If I could explain in advance how NM will deal with some particular problem, then NM wouldn't be an AGI system, it would be a narrow AI system The correct question is: how will your AGI system learn to understand the statement of a problem and figure out its own creative solution to the problem, and implement that solution. That is the problem Pei and I, in our different ways, are working on. Not pre-figuring the solutions our systems will come up with for any particular problem. -- Ben -- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
On Thursday 17 May 2007 05:36:17 pm Mike Tintner wrote: You don't start a creative process with the solution, or the kind of solution you reckon you need, i.e. in this case, the kind of architectures that you reckon will bring about AGI. Wrong. Technological innovations are quite frequently made by approaching a problem with a given technique that one has reason to think will work, and refining and adapting it until it does. The Wright brothers came to the problem of a flying machine with the key ideas of bolting a motor/airscrew onto a glider. Each part existed already--they refined the combination until it worked. The Apollo project attacked the idea of going to the moon using liquid-fueled rockets. Lots of scale-up, re-arrangement of parts, etc, but the basic idea was just pushed along until it worked. We're all starting the attack on the AI problem with the assumption that we'll do it by writing programs for electronic digital stored-program computers. If your comment were correct we should all be second-guessing this assumption and worrying about whether we shouldn't be trying networks of op-amps instead. But the comment is historically incorrect -- because the people who have the right knowledge to solve a new, big, technical problem are exactly the ones who are going to take a technique and think, Hey, I could make this work on that. Then they push on it for ten years and voila. Josh - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
Yeah, Mark, you have a good point. Mike Tintner, I'm going to once again make an effort to stop succumbing to the childish urge to reply to your messages, when we obviously are not communicating in a useful way in this context... ;-) -- Ben On 5/17/07, Mark Waser [EMAIL PROTECTED] wrote: Ben, Why are you still encouraging an obvious troll? - Original Message - *From:* Benjamin Goertzel [EMAIL PROTECTED] *To:* agi@v2.listbox.com *Sent:* Thursday, May 17, 2007 4:47 PM *Subject:* Re: [agi] Intelligence vs Efficient Intelligence *Won't somebody actually deal with the problem - how will your AGI system decide to invest or not to invest $10,000 in a Chinese mutual fund tomorrow? (You guys are supposed to be in the problem-solving business). * Look, a Novamente-based AGI system could confront this problem in 1's of different ways, just like a human could That's the whole point. It will figure out how to deal with this problem by itself. If I could explain in advance how NM will deal with some particular problem, then NM wouldn't be an AGI system, it would be a narrow AI system The correct question is: how will your AGI system learn to understand the statement of a problem and figure out its own creative solution to the problem, and implement that solution. That is the problem Pei and I, in our different ways, are working on. Not pre-figuring the solutions our systems will come up with for any particular problem. -- Ben -- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; -- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936
Re: [agi] Intelligence vs Efficient Intelligence
In fact, I'll be offline for the next couple days, which will make it easy! On 5/17/07, Benjamin Goertzel [EMAIL PROTECTED] wrote: Yeah, Mark, you have a good point. Mike Tintner, I'm going to once again make an effort to stop succumbing to the childish urge to reply to your messages, when we obviously are not communicating in a useful way in this context... ;-) -- Ben On 5/17/07, Mark Waser [EMAIL PROTECTED] wrote: Ben, Why are you still encouraging an obvious troll? - Original Message - *From:* Benjamin Goertzel [EMAIL PROTECTED] *To:* agi@v2.listbox.com *Sent:* Thursday, May 17, 2007 4:47 PM *Subject:* Re: [agi] Intelligence vs Efficient Intelligence *Won't somebody actually deal with the problem - how will your AGI system decide to invest or not to invest $10,000 in a Chinese mutual fund tomorrow? (You guys are supposed to be in the problem-solving business). * Look, a Novamente-based AGI system could confront this problem in 1's of different ways, just like a human could That's the whole point. It will figure out how to deal with this problem by itself. If I could explain in advance how NM will deal with some particular problem, then NM wouldn't be an AGI system, it would be a narrow AI system The correct question is: how will your AGI system learn to understand the statement of a problem and figure out its own creative solution to the problem, and implement that solution. That is the problem Pei and I, in our different ways, are working on. Not pre-figuring the solutions our systems will come up with for any particular problem. -- Ben -- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; -- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415user_secret=fabd7936