Re: RSI without input (was Re: [agi] Updated AGI proposal (CMR v2.1))
Matt Mahoney wrote: --- On Tue, 10/14/08, Charles Hixson [EMAIL PROTECTED] wrote: It seems clear that without external inputs the amount of improvement possible is stringently limited. That is evident from inspection. But why the without input? The only evident reason is to ensure the truth of the proposition, as it doesn't match any intended real-world scenario that I can imagine. (I've never considered the Oracle AI scenario [an AI kept within a black box that will answer all your questions without inputs] to be plausible.) If input is allowed, then we can't clearly distinguish between self improvement and learning. Clearly, learning is a legitimate form of improvement, but it is not *self* improvement. What I am trying to debunk is the perceived risk of a fast takeoff singularity launched by the first AI to achieve superhuman intelligence. In this scenario, a scientist with an IQ of 180 produces an artificial scientist with an IQ of 200, which produces an artificial scientist with an IQ of 250, and so on. I argue it can't happen because human level intelligence is the wrong threshold. There is currently a global brain (the world economy) with an IQ of around 10^10, and approaching 10^12. Oh man. It is so tempting in today's economic morass to point out the obvious stupidity of this purported super-super-genius. Why would you assign such an astronomical intelligence to the economy? Even from the POV of the best of Austrian micro-economic optimism it is not at all clear that billions of minds of human level IQ interacting with one another can be said to produce some such large exponential of the average human IQ.How much of the advancement of humanity is the result of a relatively few exceptionally bright minds rather than the billions of lesser intelligences? Are you thinking more of the entire cultural environment rather than specifically the economy? - samantha --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: RSI without input (was Re: [agi] Updated AGI proposal (CMR v2.1))
--- On Sun, 10/19/08, Samantha Atkins [EMAIL PROTECTED] wrote: Matt Mahoney wrote: There is currently a global brain (the world economy) with an IQ of around 10^10, and approaching 10^12. Oh man. It is so tempting in today's economic morass to point out the obvious stupidity of this purported super-super-genius. Why would you assign such an astronomical intelligence to the economy? Without the economy, or the language and culture needed to support it, you would be foraging for food and sleeping in the woods. You would not know that you could grow crops by planting seeds, or that you could make a spear out of sticks and rocks and use it for hunting. There is a 99.9% chance that you would starve because the primitive earth could only support a few million humans, not a few billions. I realize it makes no sense to talk of an IQ of 10^10 when current tests only go to about 200. But by any measure of goal achievement, such as dollars earned or number of humans that can be supported, the global brain has enormous intelligence. It is a known fact that groups of humans collectively make more accurate predictions than their members, e.g. prediction markets. http://en.wikipedia.org/wiki/Prediction_market Such markets would not work if the members did not individually think that they were smarter than the group (i.e. disagree). You may think you could run the government better than current leadership, but it is a fact that people are better off (as measured by GDP and migration) in democracies than dictatorships. Group decision making is also widely used in machine learning, e.g. the PAQ compression programs. How much of the advancement of humanity is the result of a relatively few exceptionally bright minds rather than the billions of lesser intelligences? Very little, because agents at any intelligence level cannot detect higher intelligence. Socrates was executed. Galileo was arrested. Even today, there is a span of decades between pioneering scientific work and its recognition with a Nobel prize. So I don't expect anyone to recognize the intelligence of the economy. But your ability to read this email depends more on circuit board assemblers in Malaysia than you are willing to give the world credit for. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: RSI without input (was Re: [agi] Updated AGI proposal (CMR v2.1))
Nicole, yes, Rosato I think, across the road. Ok with me. Cheers Peter Peter G Burton PhD http://homepage.mac.com/blinkcentral [EMAIL PROTECTED] intl 61 (0) 400 194 333 On Wednesday, October 15, 2008, at 09:08PM, Ben Goertzel [EMAIL PROTECTED] wrote: Matt wrote, in reply to me: An AI twice as smart as any human could figure out how to use the resources at his disposal to help him create an AI 3 times as smart as any human. These AI's will not be brains in vats. They will have resources at their disposal. It depends on what you mean by twice as smart. Do you mean twice as many brain cells? Twice as much memory? Twice as fast? Twice as much knowledge? Able to score 200 on an adult IQ test (if such a thing existed)? Unless you tell me otherwise, I have to assume that it means able to do what 2 people can do (or 3 or 10, the exact number isn't important). In that case, I have to argue it is the global brain that is creating the AI with a very tiny bit of help from the parent AI. You would get the same result by hiring more people. Whatever ... You are IMO just distracting attention from the main point, by making odd definitions... No, of course my colloquial phrase twice as smart does not mean as smart as two people put together. That is not the accepted interpretation of that colloquialism and you know it! To make my statement clearer, one approach is to forget about quantitating intelligence for the moment... Let's talk about qualitative differences in intelligence. Do you agree that a dog is qualitatively much more intelligent than a roach, and a human is qualitatively much more intelligent than a dog? In this sense I could replace An AI twice as smart as any human could figure out how to use the resources at his disposal to help him create an AI 3 times as smart as any human. These AI's will not be brains in vats. They will have resources at their disposal. with An AI that is qualitatively much smarter than any human could figure out how to use the resources at his disposal to help it create an AI that is qualitatively much smarter than it. These AI's will not be brains in vats. They will have resources at their disposal. On the other hand, if you insist on mathematical definitions of intelligence, we could talk about, say, the intelligence of a system as the total prediction difficulty of the set S of sequences, with the property that the system can predict S during a period of time of length T. We can define prediction difficulty as Shane Legg does in his PhD thesis. We can then average this over various time-lengths T, using some appropriate weighting function. (I'm not positing the above as an ideal definition of intelligence ... just throwing one definition out there... my conceptual point is quite independent of the specific definition of intelligence you choose) Using this sort of definition, my statement is surely true, though it would take work to prove it. Using this sort of definition, a system A2 that is twice as smart as system A1, if allowed to interact with an appropriate environment vastly more complex than either of the systems, would surely be capable of modifying itself into a system A3 that is twice as smart as A2. This seems extremely obvious and I don't want to spend time right now proving it formally. No doubt writing out the proof would reveal various mathematical conditions on the theorem statement... -- Ben G --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
On Wed, Oct 15, 2008 at 5:38 AM, Matt Mahoney [EMAIL PROTECTED] wrote: --- On Tue, 10/14/08, Terren Suydam [EMAIL PROTECTED] wrote: Matt, Your measure of intelligence seems to be based on not much more than storage capacity, processing power, I/O, and accumulated knowledge. This has the advantage of being easily formalizable, but has the disadvantage of missing a necessary aspect of intelligence. Usually when I say intelligence I mean amount of knowledge, which can be measured in bits. (Well not really, since Kolmogorov complexity is not computable). The other measures reduce to it. Increasing memory allows more knowledge to be stored. Increasing processing power and I/O bandwidth allows faster learning, or more knowledge accumulation over the same time period. Actually, amount of knowledge is just an upper bound. A random string has high algorithmic complexity but is not intelligent in any meaningful sense. My justification for this measure is based on the AIXI model. In order for an agent to guess an environment with algorithmic complexity K, the agent must be able to simulate the environment, so it must also have algorithmic complexity K. An agent with higher complexity can guess a superset of environments that a lower complexity agent could, and therefore cannot do worse in accumulated reward. Interstellar void must be astronomically intelligent, with all its incompressible noise... -- Vladimir Nesov [EMAIL PROTECTED] http://causalityrelay.wordpress.com/ --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
--- On Wed, 10/15/08, Vladimir Nesov [EMAIL PROTECTED] wrote: Interstellar void must be astronomically intelligent, with all its incompressible noise... How do you know it's not compressible? Compression is not computable. To give a concrete example, the output of RC4 looks like random noise if you don't know the key. Yet it is extremely simple algorithmically. http://en.wikipedia.org/wiki/RC4 More generally, the universe might be simulated by the following algorithm: enumerate all Turing machines until life is found, running the n'th machine for n steps. In this case, the universe (including your interstellar void) has a complexity of log2 H = 407 bits, where H is the Bekenstein bound of the Hubble radius, 2.91 x 10^122 bits. (Anyway, this is aside from my point that you apparently missed, that algorithmic complexity is an upper bound on intelligence only). -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
RSI without input (was Re: [agi] Updated AGI proposal (CMR v2.1))
--- On Tue, 10/14/08, Charles Hixson [EMAIL PROTECTED] wrote: It seems clear that without external inputs the amount of improvement possible is stringently limited. That is evident from inspection. But why the without input? The only evident reason is to ensure the truth of the proposition, as it doesn't match any intended real-world scenario that I can imagine. (I've never considered the Oracle AI scenario [an AI kept within a black box that will answer all your questions without inputs] to be plausible.) If input is allowed, then we can't clearly distinguish between self improvement and learning. Clearly, learning is a legitimate form of improvement, but it is not *self* improvement. What I am trying to debunk is the perceived risk of a fast takeoff singularity launched by the first AI to achieve superhuman intelligence. In this scenario, a scientist with an IQ of 180 produces an artificial scientist with an IQ of 200, which produces an artificial scientist with an IQ of 250, and so on. I argue it can't happen because human level intelligence is the wrong threshold. There is currently a global brain (the world economy) with an IQ of around 10^10, and approaching 10^12. THAT is the threshold we must cross. And that seed was already planted 3 billion years ago. To argue this point, I need to discredit certain alternative proposals, such as intelligent agents making random variations of itself and then testing the children with puzzles of the parent's choosing. My paper proves that proposals of this form cannot work. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: RSI without input (was Re: [agi] Updated AGI proposal (CMR v2.1))
What I am trying to debunk is the perceived risk of a fast takeoff singularity launched by the first AI to achieve superhuman intelligence. In this scenario, a scientist with an IQ of 180 produces an artificial scientist with an IQ of 200, which produces an artificial scientist with an IQ of 250, and so on. I argue it can't happen because human level intelligence is the wrong threshold. There is currently a global brain (the world economy) with an IQ of around 10^10, and approaching 10^12. THAT is the threshold we must cross. And that seed was already planted 3 billion years ago. To argue this point, I need to discredit certain alternative proposals, such as intelligent agents making random variations of itself and then testing the children with puzzles of the parent's choosing. My paper proves that proposals of this form cannot work. Your paper does **not** prove anything whatsoever about real-world situations. Among other reasons: Because, in the real world, the scientist with an IQ of 200 is **not** a brain in a vat with the inability to learn from the external world. Rather, he is able to run experiments in the external world (which has a far higher algorithmic information than him, by the way), which give him **new information** about how to go about making the scientist with an IQ of 220. Limitations on the rate of self-improvement of scientists who are brains in vats, are not really that interesting (And this is separate from the other critique I made, which is that using algorithmic information as a proxy for IQ is a very poor choice, given the critical importance of runtime complexity in intelligence. As an aside, note there are correlations between human intelligence and speed of neural processing!) -- Ben G --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: RSI without input (was Re: [agi] Updated AGI proposal (CMR v2.1))
On Thu, Oct 16, 2008 at 12:06 AM, Ben Goertzel [EMAIL PROTECTED] wrote: Among other reasons: Because, in the real world, the scientist with an IQ of 200 is **not** a brain in a vat with the inability to learn from the external world. Rather, he is able to run experiments in the external world (which has a far higher algorithmic information than him, by the way), which give him **new information** about how to go about making the scientist with an IQ of 220. Limitations on the rate of self-improvement of scientists who are brains in vats, are not really that interesting (And this is separate from the other critique I made, which is that using algorithmic information as a proxy for IQ is a very poor choice, given the critical importance of runtime complexity in intelligence. As an aside, note there are correlations between human intelligence and speed of neural processing!) Brain in a vat self-improvement is also interesting and worthwhile endeavor. One problem to tackle, for example, is to develop more efficient optimization algorithms, that will be able to faster find better plans according to the goals (and naturally apply these algorithms to decision-making during further self-improvement). Advances in algorithms can bring great efficiency, and looking at what modern computer science came up with, this efficiency rarely requires an algorithm of in the least significant complexity. There is plenty of ground to cover in the space of simple things, limitations on complexity are pragmatically void. -- Vladimir Nesov [EMAIL PROTECTED] http://causalityrelay.wordpress.com/ --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: RSI without input (was Re: [agi] Updated AGI proposal (CMR v2.1))
--- On Wed, 10/15/08, Ben Goertzel [EMAIL PROTECTED] wrote: Your paper does **not** prove anything whatsoever about real-world situations. You are correct. My RSI paper only applies to self improvement of closed systems. In the interest of proving the safety of AI, I think this is a good thing. It proves that various scenarios where an AI rewrites its source code or makes random changes and tests them, will not work without external input, even if computing power is unlimited. This removes one possible threat of a fast takeoff singularity. Also, you are right that it does not apply to many real world problems. Here my objection (as stated in my AGI proposal, but perhaps not clearly) is that creating an artificial scientist with slightly above human intelligence won't launch a singularity either, but for a different reason. It is not the scientist who creates a smarter scientist, but it is the whole global economy that creates it. George Will expresses the idea better than I do in http://www.newsweek.com/id/158752 Nobody can make a pencil, much less an AI. The global brain *is* self improving, both by learning and by reorganizing itself to be more efficient. Without input, the self organization would reach a maximum and stop. Growth requires input as well as increased computing power by adding people and computers. As for using algorithmic complexity as a proxy for intelligence (an upper bound, actually), perhaps you can suggest an alternative. Algorithmic complexity is how much we know. Less well-defined measures seem to break down into philosophical arguments over exactly what intelligence is. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: RSI without input (was Re: [agi] Updated AGI proposal (CMR v2.1))
Hi, Also, you are right that it does not apply to many real world problems. Here my objection (as stated in my AGI proposal, but perhaps not clearly) is that creating an artificial scientist with slightly above human intelligence won't launch a singularity either, but for a different reason. It is not the scientist who creates a smarter scientist, but it is the whole global economy that creates it. George Will expresses the idea better than I do in http://www.newsweek.com/id/158752 Nobody can make a pencil, much less an AI. This strikes me as a very, very bad argument. An AI twice as smart as any human could figure out how to use the resources at his disposal to help him create an AI 3 times as smart as any human. These AI's will not be brains in vats. They will have resources at their disposal. Also, when we can build one AI twice as smart as any human, we can build a million of them soon thereafter. Unlike humans, software can easily be copied. So don't think about just one smart AI. Think about a huge number of them, with all the resources in the world at their potential disposal. As for using algorithmic complexity as a proxy for intelligence (an upper bound, actually), perhaps you can suggest an alternative. Algorithmic complexity is how much we know. Less well-defined measures seem to break down into philosophical arguments over exactly what intelligence is. Algorithmic complexity is an abstraction of how much we know declaratively rather than procedurally. I am suggesting that one proxy for intelligence is the complexity of the problems that a system can solve within a certain, fixed period of time. This can be formalized in many ways, including using algorithmic information theory to formalize problem complexity. But the point is the incorporation of running speed... -- Ben G --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
It doesn't need to satisfy everyone, it just has to be the definition that you are using in your argument, and which you agree to stick to. E.g., if you define intelligence to be the resources used (given some metric) in solving some particular selection of problems, then that is a particular definition of intelligence. It may not be a very good one, though, as it looks like a system that knows the answers ahead of time and responds quickly would win over one that understood the problems in depth. Rather like a multiple choice test rather than an essay. I'm sure that one could fudge the definition to skirt that particular pothole, but it would be an ad hoc patch. I don't trust that entire mechanism of defining intelligence. Still, if I know what you mean, I don't have to accept your interpretations to understand your argument. (You can't average across all domains, only across some pre-specified set of domains. Infinity doesn't exist in the implementable universe.) Personally, I'm not convinced by the entire process of measuring intelligence. I don't think that there *IS* any such thing. If it were a disease, I'd call intelligence a syndrome rather than a diagnosis. It's a collection of partially related capabilities given one name to make them easy to think about, while ignoring details. As such it has many uses, but it's easy to mistake it for some genuine thing, especially as it's an intangible. As an analogy consider the gene for blue eyes. There is no such gene. There is a combination of genes that yields blue eyes, and it's characterized by the lack of genes for other eye colors. (It's more complex than that, but that's enough.) E.g., there appears to be a particular gene which is present in almost all people which enables them to parse grammatical sentences. But there have been found a few people in one family where this gene is damaged. The result is that about half the members of that family can't speak or understand language. Are they unintelligent? Well, the can't parse grammatical sentences, and they can't learn language. In most other ways they appear as intelligent as anyone else. So I'm suspicious of ALL definitions of intelligence which treat it as some kind of global thing. But if you give me the definition that you are using in an argument, then I can at least attempt to understand what you are saying. Terren Suydam wrote: Charles, I'm not sure it's possible to nail down a measure of intelligence that's going to satisfy everyone. Presumably, it would be some measure of performance in problem solving across a wide variety of novel domains in complex (i.e. not toy) environments. Obviously among potential agents, some will do better in domain D1 than others, while doing worse in D2. But we're looking for an average across all domains. My task-specific examples may have confused the issue there, you were right to point that out. But if you give all agents identical processing power and storage space, then the winner will be the one that was able to assimilate and model each problem space the most efficiently, on average. Which ultimately means the one which used the *least* amount of overall computation. Terren --- On Tue, 10/14/08, Charles Hixson [EMAIL PROTECTED] wrote: From: Charles Hixson [EMAIL PROTECTED] Subject: Re: [agi] Updated AGI proposal (CMR v2.1) To: agi@v2.listbox.com Date: Tuesday, October 14, 2008, 2:12 PM If you want to argue this way (reasonable), then you need a specific definition of intelligence. One that allows it to be accurately measured (and not just in principle). IQ definitely won't serve. Neither will G. Neither will GPA (if you're discussing a student). Because of this, while I think your argument is generally reasonable, I don't thing it's useful. Most of what you are discussing is task specific, and as such I'm not sure that intelligence is a reasonable term to use. An expert engineer might be, e.g., a lousy bridge player. Yet both are thought of as requiring intelligence. I would assert that in both cases a lot of what's being measured is task specific processing, i.e., narrow AI. (Of course, I also believe that an AGI is impossible in the true sense of general, and that an approximately AGI will largely act as a coordinator between a bunch of narrow AI pieces of varying generality. This seems to be a distinctly minority view.) Terren Suydam wrote: Hi Will, I think humans provide ample evidence that intelligence is not necessarily correlated with processing power. The genius engineer in my example solves a given problem with *much less* overall processing than the ordinary engineer, so in this case intelligence is correlated with some measure of cognitive efficiency (which I will leave undefined). Likewise, a grandmaster chess player looks at a given position and can calculate a better move in one second than you or me could come up
Re: RSI without input (was Re: [agi] Updated AGI proposal (CMR v2.1))
--- On Wed, 10/15/08, Ben Goertzel [EMAIL PROTECTED] wrote: An AI twice as smart as any human could figure out how to use the resources at his disposal to help him create an AI 3 times as smart as any human. These AI's will not be brains in vats. They will have resources at their disposal. It depends on what you mean by twice as smart. Do you mean twice as many brain cells? Twice as much memory? Twice as fast? Twice as much knowledge? Able to score 200 on an adult IQ test (if such a thing existed)? Unless you tell me otherwise, I have to assume that it means able to do what 2 people can do (or 3 or 10, the exact number isn't important). In that case, I have to argue it is the global brain that is creating the AI with a very tiny bit of help from the parent AI. You would get the same result by hiring more people. The fact is we have been creating smarter than human machines for 50 years now, depending on what intelligence test you use. And they have greatly increased our productivity by doing well the things that humans do poorly, much more than you could have gotten by hiring more people. Also, when we can build one AI twice as smart as any human, we can build a million of them soon thereafter. All of whom will know exactly the same thing. Training each of them to do a specialized task will not be cheap. And no, they will not just learn on their own without human effort. On the job training has real costs in mistakes and lost productivity. Not everything they need to know is written down. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
The small point I was trying to make was that cognitive architecture is much more important to the realization of AGI than the amount of processing power you have at your disposal, or some other such platform-related considerations. It doesn't seem like a very controversial point to me. Objecting to it on the basis of the difficulty/impossibility of measuring intelligence seems like a bit of a tangent. --- On Wed, 10/15/08, Charles Hixson [EMAIL PROTECTED] wrote: From: Charles Hixson [EMAIL PROTECTED] Subject: Re: [agi] Updated AGI proposal (CMR v2.1) To: agi@v2.listbox.com Date: Wednesday, October 15, 2008, 8:09 PM It doesn't need to satisfy everyone, it just has to be the definition that you are using in your argument, and which you agree to stick to. E.g., if you define intelligence to be the resources used (given some metric) in solving some particular selection of problems, then that is a particular definition of intelligence. It may not be a very good one, though, as it looks like a system that knows the answers ahead of time and responds quickly would win over one that understood the problems in depth. Rather like a multiple choice test rather than an essay. I'm sure that one could fudge the definition to skirt that particular pothole, but it would be an ad hoc patch. I don't trust that entire mechanism of defining intelligence. Still, if I know what you mean, I don't have to accept your interpretations to understand your argument. (You can't average across all domains, only across some pre-specified set of domains. Infinity doesn't exist in the implementable universe.) Personally, I'm not convinced by the entire process of measuring intelligence. I don't think that there *IS* any such thing. If it were a disease, I'd call intelligence a syndrome rather than a diagnosis. It's a collection of partially related capabilities given one name to make them easy to think about, while ignoring details. As such it has many uses, but it's easy to mistake it for some genuine thing, especially as it's an intangible. As an analogy consider the gene for blue eyes. There is no such gene. There is a combination of genes that yields blue eyes, and it's characterized by the lack of genes for other eye colors. (It's more complex than that, but that's enough.) E.g., there appears to be a particular gene which is present in almost all people which enables them to parse grammatical sentences. But there have been found a few people in one family where this gene is damaged. The result is that about half the members of that family can't speak or understand language. Are they unintelligent? Well, the can't parse grammatical sentences, and they can't learn language. In most other ways they appear as intelligent as anyone else. So I'm suspicious of ALL definitions of intelligence which treat it as some kind of global thing. But if you give me the definition that you are using in an argument, then I can at least attempt to understand what you are saying. Terren Suydam wrote: Charles, I'm not sure it's possible to nail down a measure of intelligence that's going to satisfy everyone. Presumably, it would be some measure of performance in problem solving across a wide variety of novel domains in complex (i.e. not toy) environments. Obviously among potential agents, some will do better in domain D1 than others, while doing worse in D2. But we're looking for an average across all domains. My task-specific examples may have confused the issue there, you were right to point that out. But if you give all agents identical processing power and storage space, then the winner will be the one that was able to assimilate and model each problem space the most efficiently, on average. Which ultimately means the one which used the *least* amount of overall computation. Terren --- On Tue, 10/14/08, Charles Hixson [EMAIL PROTECTED] wrote: From: Charles Hixson [EMAIL PROTECTED] Subject: Re: [agi] Updated AGI proposal (CMR v2.1) To: agi@v2.listbox.com Date: Tuesday, October 14, 2008, 2:12 PM If you want to argue this way (reasonable), then you need a specific definition of intelligence. One that allows it to be accurately measured (and not just in principle). IQ definitely won't serve. Neither will G. Neither will GPA (if you're discussing a student). Because of this, while I think your argument is generally reasonable, I don't thing it's useful. Most of what you are discussing is task specific, and as such I'm not sure that intelligence is a reasonable term to use. An expert engineer might be, e.g., a lousy bridge player. Yet both are thought of as requiring intelligence. I would assert that in both cases a lot of what's being measured is task specific processing, i.e., narrow AI
Re: RSI without input (was Re: [agi] Updated AGI proposal (CMR v2.1))
Matt wrote, in reply to me: An AI twice as smart as any human could figure out how to use the resources at his disposal to help him create an AI 3 times as smart as any human. These AI's will not be brains in vats. They will have resources at their disposal. It depends on what you mean by twice as smart. Do you mean twice as many brain cells? Twice as much memory? Twice as fast? Twice as much knowledge? Able to score 200 on an adult IQ test (if such a thing existed)? Unless you tell me otherwise, I have to assume that it means able to do what 2 people can do (or 3 or 10, the exact number isn't important). In that case, I have to argue it is the global brain that is creating the AI with a very tiny bit of help from the parent AI. You would get the same result by hiring more people. Whatever ... You are IMO just distracting attention from the main point, by making odd definitions... No, of course my colloquial phrase twice as smart does not mean as smart as two people put together. That is not the accepted interpretation of that colloquialism and you know it! To make my statement clearer, one approach is to forget about quantitating intelligence for the moment... Let's talk about qualitative differences in intelligence. Do you agree that a dog is qualitatively much more intelligent than a roach, and a human is qualitatively much more intelligent than a dog? In this sense I could replace An AI twice as smart as any human could figure out how to use the resources at his disposal to help him create an AI 3 times as smart as any human. These AI's will not be brains in vats. They will have resources at their disposal. with An AI that is qualitatively much smarter than any human could figure out how to use the resources at his disposal to help it create an AI that is qualitatively much smarter than it. These AI's will not be brains in vats. They will have resources at their disposal. On the other hand, if you insist on mathematical definitions of intelligence, we could talk about, say, the intelligence of a system as the total prediction difficulty of the set S of sequences, with the property that the system can predict S during a period of time of length T. We can define prediction difficulty as Shane Legg does in his PhD thesis. We can then average this over various time-lengths T, using some appropriate weighting function. (I'm not positing the above as an ideal definition of intelligence ... just throwing one definition out there... my conceptual point is quite independent of the specific definition of intelligence you choose) Using this sort of definition, my statement is surely true, though it would take work to prove it. Using this sort of definition, a system A2 that is twice as smart as system A1, if allowed to interact with an appropriate environment vastly more complex than either of the systems, would surely be capable of modifying itself into a system A3 that is twice as smart as A2. This seems extremely obvious and I don't want to spend time right now proving it formally. No doubt writing out the proof would reveal various mathematical conditions on the theorem statement... -- Ben G --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
On Tue, Oct 14, 2008 at 8:36 AM, Matt Mahoney [EMAIL PROTECTED] wrote: Ben, If you want to argue that recursive self improvement is a special case of learning, then I have no disagreement with the rest of your argument. But is this really a useful approach to solving AGI? A group of humans can generally make better decisions (more accurate predictions) by voting than any member of the group can. Did these humans improve themselves? My point is that a single person can't create much of anything, much less an AI smarter than himself. If it happens, it will be created by an organization of billions of humans. Without this organization, you would probably not think to create spears out of sticks and rocks. That is my problem with the seed AI approach. The seed AI depends on the knowledge and resources of the economy to do anything. An AI twice as smart as a human could not do any more than 2 people could. You need to create an AI that is billions of times smarter to get anywhere. We are already doing that. Human culture is improving itself by accumulating knowledge, by becoming better organized through communication and specialization, and by adding more babies and computers. You are slipping from strained interpretation of the technical argument to the informal point that argument was intended to rationalize. If interpretation of technical argument is weaker than original informal argument it was invented to support, there is no point in technical argument. Using the fact of 2+2=4 won't give technical support to e.g. philosophy of solipsism. -- Vladimir Nesov [EMAIL PROTECTED] http://causalityrelay.wordpress.com/ --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
2008/10/14 Terren Suydam [EMAIL PROTECTED]: --- On Tue, 10/14/08, Matt Mahoney [EMAIL PROTECTED] wrote: An AI that is twice as smart as a human can make no more progress than 2 humans. Spoken like someone who has never worked with engineers. A genius engineer can outproduce 20 ordinary engineers in the same timeframe. Do you really believe the relationship between intelligence and output is linear? I'm going to use this post as a place to grind one of my axes, apologies Terren. The relationship between processing power and results is not necessarily linear or even positively correlated. And as an increase in intelligence above a certain level requires increased processing power (or perhaps not? anyone disagree?). When the cost of adding more computational power, outweighs the amount of money or energy that you acquire from adding the power, there is not much point adding the computational power. Apart from if you are in competition with other agents, that can out smart you. Some of the traditional views of RSI neglects this and thinks that increased intelligence is always a useful thing. It is not very There is a reason why lots of the planets biomass has stayed as bacteria. It does perfectly well like that. It survives. Too much processing power is a bad thing, it means less for self-preservation and affecting the world. Balancing them is a tricky proposition indeed. Will Pearson --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
Hi Will, I think humans provide ample evidence that intelligence is not necessarily correlated with processing power. The genius engineer in my example solves a given problem with *much less* overall processing than the ordinary engineer, so in this case intelligence is correlated with some measure of cognitive efficiency (which I will leave undefined). Likewise, a grandmaster chess player looks at a given position and can calculate a better move in one second than you or me could come up with if we studied the board for an hour. Grandmasters often do publicity events where they play dozens of people simultaneously, spending just a few seconds on each board, and winning most of the games. Of course, you were referring to intelligence above a certain level, but if that level is high above human intelligence, there isn't much we can assume about that since it is by definition unknowable by humans. Terren --- On Tue, 10/14/08, William Pearson [EMAIL PROTECTED] wrote: The relationship between processing power and results is not necessarily linear or even positively correlated. And as an increase in intelligence above a certain level requires increased processing power (or perhaps not? anyone disagree?). When the cost of adding more computational power, outweighs the amount of money or energy that you acquire from adding the power, there is not much point adding the computational power. Apart from if you are in competition with other agents, that can out smart you. Some of the traditional views of RSI neglects this and thinks that increased intelligence is always a useful thing. It is not very There is a reason why lots of the planets biomass has stayed as bacteria. It does perfectly well like that. It survives. Too much processing power is a bad thing, it means less for self-preservation and affecting the world. Balancing them is a tricky proposition indeed. Will Pearson --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
--- On Tue, 10/14/08, Terren Suydam [EMAIL PROTECTED] wrote: --- On Tue, 10/14/08, Matt Mahoney [EMAIL PROTECTED] wrote: An AI that is twice as smart as a human can make no more progress than 2 humans. Spoken like someone who has never worked with engineers. A genius engineer can outproduce 20 ordinary engineers in the same timeframe. Do you really believe the relationship between intelligence and output is linear? You are right, it is not, but that does not detract from my main point. Two brains have twice as much storage capacity, processing power, and I/O as one brain. They have less than twice as much knowledge because some of it is shared. They can do less than twice as much work because the brain has a fixed rate of long term learning (2 bits per second), and a portion of that must be devoted to communicating with the other brain. The intelligence of 2 brains is between 1 and 2 depending on the degree to which the intelligence test can be parallelized. The degree of parallelization is generally higher for humans than it is for dogs because humans can communicate more efficiently. Ants and bees communicate to some extent, so we observe that a colony is more intelligent (at finding food) than any individual. I have said many times that humans cannot test for higher than human intelligence. Here is proof. We know from experiments that groups of humans make better predictions (by voting) than individuals. However, if individuals recognized that the group was smarter, then they would never disagree with it. But if they never disagreed, then the group would not be smarter. With regard to RSI, we now have a global economy of 10^10 brains, which I estimate is about 10^8 times smarter (and growing) than any individual. It is less than 10^10 because of less than optimal organization. I estimate the inefficiency based on the cost of replacing an employee in lost productivity. So even an AGI that is 1000 times smarter than a human would only have the impact of adding a few thousand more people, whether you measure intelligence by instructions per second, memory, I/O bandwidth, or bits of knowledge. I think Ben could see just how much their team depends on the (unrecognized) intelligence of the global brain if they imagined going back 100 years in time and asking how much progress they would be making toward AGI then? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
--- On Tue, 10/14/08, Matt Mahoney [EMAIL PROTECTED] wrote: An AI that is twice as smart as a human can make no more progress than 2 humans. Spoken like someone who has never worked with engineers. A genius engineer can outproduce 20 ordinary engineers in the same timeframe. Do you really believe the relationship between intelligence and output is linear? Terren --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
Matt, Your measure of intelligence seems to be based on not much more than storage capacity, processing power, I/O, and accumulated knowledge. This has the advantage of being easily formalizable, but has the disadvantage of missing a necessary aspect of intelligence. I have yet to see from you any acknowledgment that cognitive architecture is at all important to realized intelligence. Even your global brain requires an explanation of how cognition actually happens at each of the nodes, be they humans or AI. Cognitive architecture (whatever form that takes) determines the efficiency of an intelligence given more external constraints like processing power etc. I assume that it is this aspect that is the primary target of significant (disruptive) improvement in RSI schemes. Terren --- On Tue, 10/14/08, Matt Mahoney [EMAIL PROTECTED] wrote: Two brains have twice as much storage capacity, processing power, and I/O as one brain. They have less than twice as much knowledge because some of it is shared. They can do less than twice as much work because the brain has a fixed rate of long term learning (2 bits per second), and a portion of that must be devoted to communicating with the other brain. The intelligence of 2 brains is between 1 and 2 depending on the degree to which the intelligence test can be parallelized. The degree of parallelization is generally higher for humans than it is for dogs because humans can communicate more efficiently. Ants and bees communicate to some extent, so we observe that a colony is more intelligent (at finding food) than any individual. --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
Hi Terren, I think humans provide ample evidence that intelligence is not necessarily correlated with processing power. The genius engineer in my example solves a given problem with *much less* overall processing than the ordinary engineer, so in this case intelligence is correlated with some measure of cognitive efficiency (which I will leave undefined). Likewise, a grandmaster chess player looks at a given position and can calculate a better move in one second than you or me could come up with if we studied the board for an hour. Grandmasters often do publicity events where they play dozens of people simultaneously, spending just a few seconds on each board, and winning most of the games. What I meant was at processing power/memory Z, there is an problem solving ability Y which is the maximum. To increase the problem solving ability above Y you would have to increase processing power/memory. That is when cognitive efficiency reaches one, in your terminology. Efficiency is normally measured in ratios so that seems natural. There are things you can't model with limits of processing power/memory which restricts your ability to solve them. Of course, you were referring to intelligence above a certain level, but if that level is high above human intelligence, there isn't much we can assume about that since it is by definition unknowable by humans. Not quite what I meant. Will --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
--- On Tue, 10/14/08, Ben Goertzel [EMAIL PROTECTED] wrote: Here is how I see this exchange... You proposed a so-called *mathematical* debunking of RSI. I presented some detailed arguments against this so-called debunking, pointing out that its mathematical assumptions and its quantification of improvement bear little relevance to real-world AI now or in the future. I can only disprove a mathematical argument. I think I have disproved RSI based on a model of self introspection without input. If you want to allow input, then you need to make a clear distinction between self improvement and learning. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
If you want to argue this way (reasonable), then you need a specific definition of intelligence. One that allows it to be accurately measured (and not just in principle). IQ definitely won't serve. Neither will G. Neither will GPA (if you're discussing a student). Because of this, while I think your argument is generally reasonable, I don't thing it's useful. Most of what you are discussing is task specific, and as such I'm not sure that intelligence is a reasonable term to use. An expert engineer might be, e.g., a lousy bridge player. Yet both are thought of as requiring intelligence. I would assert that in both cases a lot of what's being measured is task specific processing, i.e., narrow AI. (Of course, I also believe that an AGI is impossible in the true sense of general, and that an approximately AGI will largely act as a coordinator between a bunch of narrow AI pieces of varying generality. This seems to be a distinctly minority view.) Terren Suydam wrote: Hi Will, I think humans provide ample evidence that intelligence is not necessarily correlated with processing power. The genius engineer in my example solves a given problem with *much less* overall processing than the ordinary engineer, so in this case intelligence is correlated with some measure of cognitive efficiency (which I will leave undefined). Likewise, a grandmaster chess player looks at a given position and can calculate a better move in one second than you or me could come up with if we studied the board for an hour. Grandmasters often do publicity events where they play dozens of people simultaneously, spending just a few seconds on each board, and winning most of the games. Of course, you were referring to intelligence above a certain level, but if that level is high above human intelligence, there isn't much we can assume about that since it is by definition unknowable by humans. Terren --- On Tue, 10/14/08, William Pearson [EMAIL PROTECTED] wrote: The relationship between processing power and results is not necessarily linear or even positively correlated. And as an increase in intelligence above a certain level requires increased processing power (or perhaps not? anyone disagree?). When the cost of adding more computational power, outweighs the amount of money or energy that you acquire from adding the power, there is not much point adding the computational power. Apart from if you are in competition with other agents, that can out smart you. Some of the traditional views of RSI neglects this and thinks that increased intelligence is always a useful thing. It is not very There is a reason why lots of the planets biomass has stayed as bacteria. It does perfectly well like that. It survives. Too much processing power is a bad thing, it means less for self-preservation and affecting the world. Balancing them is a tricky proposition indeed. Will Pearson --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
An AI that is twice as smart as a human can make no more progress than 2 humans. Actually I'll argue that we can't make predictions about what a greater-than-human intelligence would do. Maybe the summed intelligence of 2 humans would be sufficient to do the work of a dozen. Maybe --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
Charles, I'm not sure it's possible to nail down a measure of intelligence that's going to satisfy everyone. Presumably, it would be some measure of performance in problem solving across a wide variety of novel domains in complex (i.e. not toy) environments. Obviously among potential agents, some will do better in domain D1 than others, while doing worse in D2. But we're looking for an average across all domains. My task-specific examples may have confused the issue there, you were right to point that out. But if you give all agents identical processing power and storage space, then the winner will be the one that was able to assimilate and model each problem space the most efficiently, on average. Which ultimately means the one which used the *least* amount of overall computation. Terren --- On Tue, 10/14/08, Charles Hixson [EMAIL PROTECTED] wrote: From: Charles Hixson [EMAIL PROTECTED] Subject: Re: [agi] Updated AGI proposal (CMR v2.1) To: agi@v2.listbox.com Date: Tuesday, October 14, 2008, 2:12 PM If you want to argue this way (reasonable), then you need a specific definition of intelligence. One that allows it to be accurately measured (and not just in principle). IQ definitely won't serve. Neither will G. Neither will GPA (if you're discussing a student). Because of this, while I think your argument is generally reasonable, I don't thing it's useful. Most of what you are discussing is task specific, and as such I'm not sure that intelligence is a reasonable term to use. An expert engineer might be, e.g., a lousy bridge player. Yet both are thought of as requiring intelligence. I would assert that in both cases a lot of what's being measured is task specific processing, i.e., narrow AI. (Of course, I also believe that an AGI is impossible in the true sense of general, and that an approximately AGI will largely act as a coordinator between a bunch of narrow AI pieces of varying generality. This seems to be a distinctly minority view.) Terren Suydam wrote: Hi Will, I think humans provide ample evidence that intelligence is not necessarily correlated with processing power. The genius engineer in my example solves a given problem with *much less* overall processing than the ordinary engineer, so in this case intelligence is correlated with some measure of cognitive efficiency (which I will leave undefined). Likewise, a grandmaster chess player looks at a given position and can calculate a better move in one second than you or me could come up with if we studied the board for an hour. Grandmasters often do publicity events where they play dozens of people simultaneously, spending just a few seconds on each board, and winning most of the games. Of course, you were referring to intelligence above a certain level, but if that level is high above human intelligence, there isn't much we can assume about that since it is by definition unknowable by humans. Terren --- On Tue, 10/14/08, William Pearson [EMAIL PROTECTED] wrote: The relationship between processing power and results is not necessarily linear or even positively correlated. And as an increase in intelligence above a certain level requires increased processing power (or perhaps not? anyone disagree?). When the cost of adding more computational power, outweighs the amount of money or energy that you acquire from adding the power, there is not much point adding the computational power. Apart from if you are in competition with other agents, that can out smart you. Some of the traditional views of RSI neglects this and thinks that increased intelligence is always a useful thing. It is not very There is a reason why lots of the planets biomass has stayed as bacteria. It does perfectly well like that. It survives. Too much processing power is a bad thing, it means less for self-preservation and affecting the world. Balancing them is a tricky proposition indeed. Will Pearson --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
--- On Tue, 10/14/08, William Pearson [EMAIL PROTECTED] wrote: There are things you can't model with limits of processing power/memory which restricts your ability to solve them. Processing power, storage capacity, and so forth, are all important in the realization of an AI but I don't see how they limit your ability to model or solve problems except in terms of performance... i.e. can a problem be solved within time T. Those are factors outside of the black box of intelligence. Cognitive architecture is the guts of the black box. Any attempt to create AGI cannot be taken seriously if it doesn't explain what intelligence does, inside the black box, whether you're talking about an individual agent or a globally distributed one. (By the way, it's worth noting that problem solving ability Y is uncomputable since it's basically just a twist on Kolmogorov Complexity. Which is to say, you can never prove that you have the perfect (un-improvable) cognitive architecture given finite resources.) With toy problems like chess, increasing computing power can compensate for what amounts to a wildly inefficient cognitive architecture. In the real world of AGI, you have to work on efficiency first because the complexity is just too high to manage. So while you can get linear improvement on Y by increasing out-of-the-black-box factors, it's inside the box you get the non-linear, punctuated gains that are in all likelihood necessary to create AGI. Terren --- On Tue, 10/14/08, William Pearson [EMAIL PROTECTED] wrote: From: William Pearson [EMAIL PROTECTED] Subject: Re: [agi] Updated AGI proposal (CMR v2.1) To: agi@v2.listbox.com Date: Tuesday, October 14, 2008, 1:13 PM Hi Terren, I think humans provide ample evidence that intelligence is not necessarily correlated with processing power. The genius engineer in my example solves a given problem with *much less* overall processing than the ordinary engineer, so in this case intelligence is correlated with some measure of cognitive efficiency (which I will leave undefined). Likewise, a grandmaster chess player looks at a given position and can calculate a better move in one second than you or me could come up with if we studied the board for an hour. Grandmasters often do publicity events where they play dozens of people simultaneously, spending just a few seconds on each board, and winning most of the games. What I meant was at processing power/memory Z, there is an problem solving ability Y which is the maximum. To increase the problem solving ability above Y you would have to increase processing power/memory. That is when cognitive efficiency reaches one, in your terminology. Efficiency is normally measured in ratios so that seems natural. There are things you can't model with limits of processing power/memory which restricts your ability to solve them. Of course, you were referring to intelligence above a certain level, but if that level is high above human intelligence, there isn't much we can assume about that since it is by definition unknowable by humans. Not quite what I meant. Will --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
On Tue, Oct 14, 2008 at 2:41 PM, Matt Mahoney wrote: But no matter. Whichever definition you accept, RSI is not a viable path to AGI. An AI that is twice as smart as a human can make no more progress than 2 humans. I can't say I've noticed two dogs being smarter than one dog. Admittedly, a pack of dogs can do hunting better, but they are not 'smarter'. Numbers just increase capabilities. Two humans can lift a heavier object than one human, but they are not twice as smart. As Ben says, I don't see a necessary connection between RSI and 'smarts'. It's a technique applicable from very basic levels. BillK --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
--- On Tue, 10/14/08, Vladimir Nesov [EMAIL PROTECTED] wrote: On Tue, Oct 14, 2008 at 8:36 AM, Matt Mahoney [EMAIL PROTECTED] wrote: Ben, If you want to argue that recursive self improvement is a special case of learning, then I have no disagreement with the rest of your argument. You are slipping from strained interpretation of the technical argument to the informal point that argument was intended to rationalize. If interpretation of technical argument is weaker than original informal argument it was invented to support, there is no point in technical argument. Using the fact of 2+2=4 won't give technical support to e.g. philosophy of solipsism. I did not say that I agree with Ben's definition of RSI to include learning. But no matter. Whichever definition you accept, RSI is not a viable path to AGI. An AI that is twice as smart as a human can make no more progress than 2 humans. You don't have automatic self improvement until you have AI that is billions of times smarter. A team of a few people isn't going to build that. The cost of training such a system with 10^17 to 10^18 bits of useful knowledge is in the quadrillions of dollars, even if the hardware is free and the problem of brain emulation is solved. Until then, you have manual self improvement. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
Matt, But no matter. Whichever definition you accept, RSI is not a viable path to AGI. An AI that is twice as smart as a human can make no more progress than 2 humans. You don't have automatic self improvement until you have AI that is billions of times smarter. A team of a few people isn't going to build that. The cost of training such a system with 10^17 to 10^18 bits of useful knowledge is in the quadrillions of dollars, even if the hardware is free and the problem of brain emulation is solved. Until then, you have manual self improvement. Here is how I see this exchange... You proposed a so-called *mathematical* debunking of RSI. I presented some detailed arguments against this so-called debunking, pointing out that its mathematical assumptions and its quantification of improvement bear little relevance to real-world AI now or in the future. You then responded by ignoring my detailed arguments, and retreating into informal, nonmathematical generalizations ... and furthermore, ones that don't seem to make much sense to me (or others on this list, if the responses are indicative...) I don't know what you mean by twice as smart but I'm sure I can make more than twice as much progress at science and engineering as someone with half my IQ ;-p ... my IQ is around 180 whereas someone with an IQ of 90 couldn't even understand this email let alone design an AGI or a machine learning algorithm, etc. ... they probably couldn't even do my taxes for me ;-p It is not clear why you think an AGI needs to be billions of times smarter than a human to undergo dramatic RSI. It might not need to be *any* smarter than a smart human ... maybe an AGI with the same IQ as a smart human but an underlying architecture built with RSI in mind, could be able to rapidly self-improve. In fact I strongly suspect this is the case, though I can't prove it ... and nor can you disprove it, without making unrealistic assumptions that render your disproof irrelevant!! -- Ben G --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
Will:There is a reason why lots of the planets biomass has stayed as bacteria. It does perfectly well like that. It survives. Too much processing power is a bad thing, it means less for self-preservation and affecting the world. Balancing them is a tricky proposition indeed Interesting thought. But do you (or anyone else) have any further thoughts about what the proper balance between brain and body relative to what set of functions/behaviours is, or how it is determined or adjusted? (Obviously a v. difficult question)/ --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
--- On Tue, 10/14/08, Terren Suydam [EMAIL PROTECTED] wrote: Matt, Your measure of intelligence seems to be based on not much more than storage capacity, processing power, I/O, and accumulated knowledge. This has the advantage of being easily formalizable, but has the disadvantage of missing a necessary aspect of intelligence. Usually when I say intelligence I mean amount of knowledge, which can be measured in bits. (Well not really, since Kolmogorov complexity is not computable). The other measures reduce to it. Increasing memory allows more knowledge to be stored. Increasing processing power and I/O bandwidth allows faster learning, or more knowledge accumulation over the same time period. Actually, amount of knowledge is just an upper bound. A random string has high algorithmic complexity but is not intelligent in any meaningful sense. My justification for this measure is based on the AIXI model. In order for an agent to guess an environment with algorithmic complexity K, the agent must be able to simulate the environment, so it must also have algorithmic complexity K. An agent with higher complexity can guess a superset of environments that a lower complexity agent could, and therefore cannot do worse in accumulated reward. I have yet to see from you any acknowledgment that cognitive architecture is at all important to realized intelligence. Even your global brain requires an explanation of how cognition actually happens at each of the nodes, be they humans or AI. Cognitive architecture is not relevant to Legg and Hutter's universal intelligence (expected reward in random AIXI environments). It is only important for specific subsets of possible goals, like the ones that are important to us. If you define intelligence by the Turing test, then obviously the cognitive architecture should model a human brain. In my global brain model, nodes trade messages when the receivers can compress them smaller than the senders, achieving distributed data compression. In general, compression is not computable regardless of architecture. In practice the messages are natural language text, so the architecture is important. It will probably be a neural language model. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
[agi] Updated AGI proposal (CMR v2.1)
I updated my AGI proposal from a few days ago. http://www.mattmahoney.net/agi2.html There are two major changes. First I clarified the routing strategy and justified it on an information theoretic basis. An organization is optimally efficient when its members specialize with no duplication of knowledge or skills. To achieve this, we use a market economy to trade messages where information has negative value. It is mutually beneficial for peers to trade messages when the receivers can compress them more tightly than the senders. This results in convergence to an optimal mapping of peers to clusters of data in semantic space. The routing strategy is for a peer to use cached messages from its neighbors as estimates of the neighbor's database. For a message X and each neighbor j, it computes the distance D(X,Y_j) where Y_j is a concatenation of cached messages from peer j. Then it routes X to j because it estimates that j can store X most efficiently. Routing stops when j is itself. The distance function is non-mutual information: D(X,Y) = K(X|Y) + K(Y|X) where K is Kolmogorov complexity, the size of the shortest program that can output X or Y given the other message as input. When I wrote my thesis, I assumed a vector space language model, but I just now realized that D is a measure, compatible with Euclidean distance in the vector space model. K is not computable, but we can approximate K using the output size of a text compressor. The economic model rewards good compression algorithms. The second change is a new section (5) addressing long term safety. I think I have debunked RSI, proving that the friendly seed AI approach could not work even in theory. This leaves an evolutionary improvement model in which peers compete for resources in a hostile environment. The other risks I have identified are competition from uploads with property rights, intelligent worms, and a singularity that redefines humanity making the question of human extinction moot. I don't have good solutions to these risks. I did not mention all possible risks, e.g. gray goo. To answer Mike Tintner's remark, yes, $1 quadrillion is expensive, but I think that AGI will pay for itself many times over. It won't address the basic instability and unpredictability of speculative investment markets. It will probably make matters worse by enabling nonstop automated trading and waves of panic selling traveling at the speed of light. As before, comments are welcome. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
I was eager to debunk your supposed debunking of recursive self-improvement, but I found that when I tried to open that PDF file, it looked like a bunch of gibberish (random control characters) in my PDF reader (Preview on OSX Leopard) ben g On Mon, Oct 13, 2008 at 12:19 PM, Matt Mahoney [EMAIL PROTECTED] wrote: I updated my AGI proposal from a few days ago. http://www.mattmahoney.net/agi2.html There are two major changes. First I clarified the routing strategy and justified it on an information theoretic basis. An organization is optimally efficient when its members specialize with no duplication of knowledge or skills. To achieve this, we use a market economy to trade messages where information has negative value. It is mutually beneficial for peers to trade messages when the receivers can compress them more tightly than the senders. This results in convergence to an optimal mapping of peers to clusters of data in semantic space. The routing strategy is for a peer to use cached messages from its neighbors as estimates of the neighbor's database. For a message X and each neighbor j, it computes the distance D(X,Y_j) where Y_j is a concatenation of cached messages from peer j. Then it routes X to j because it estimates that j can store X most efficiently. Routing stops when j is itself. The distance function is non-mutual information: D(X,Y) = K(X|Y) + K(Y|X) where K is Kolmogorov complexity, the size of the shortest program that can output X or Y given the other message as input. When I wrote my thesis, I assumed a vector space language model, but I just now realized that D is a measure, compatible with Euclidean distance in the vector space model. K is not computable, but we can approximate K using the output size of a text compressor. The economic model rewards good compression algorithms. The second change is a new section (5) addressing long term safety. I think I have debunked RSI, proving that the friendly seed AI approach could not work even in theory. This leaves an evolutionary improvement model in which peers compete for resources in a hostile environment. The other risks I have identified are competition from uploads with property rights, intelligent worms, and a singularity that redefines humanity making the question of human extinction moot. I don't have good solutions to these risks. I did not mention all possible risks, e.g. gray goo. To answer Mike Tintner's remark, yes, $1 quadrillion is expensive, but I think that AGI will pay for itself many times over. It won't address the basic instability and unpredictability of speculative investment markets. It will probably make matters worse by enabling nonstop automated trading and waves of panic selling traveling at the speed of light. As before, comments are welcome. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] Nothing will ever be attempted if all possible objections must be first overcome - Dr Samuel Johnson --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
--- On Mon, 10/13/08, Ben Goertzel [EMAIL PROTECTED] wrote: I was eager to debunk your supposed debunking of recursive self-improvement, but I found that when I tried to open that PDF file, it looked like a bunch of gibberish (random control characters) in my PDF reader (Preview on OSX Leopard) That's odd. Maybe you should run Windows :-( Anyway I posted an HTML version. Not sure why PDF wouldn't work. I created both in OpenOffice. http://www.mattmahoney.net/rsi.pdf http://www.mattmahoney.net/rsi.html Anyone else have trouble reading the PDF version? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
I can read the pdf just fine. I am also using mac's Preview program. So it is not that... --Abram On Mon, Oct 13, 2008 at 1:29 PM, Matt Mahoney [EMAIL PROTECTED] wrote: --- On Mon, 10/13/08, Ben Goertzel [EMAIL PROTECTED] wrote: I was eager to debunk your supposed debunking of recursive self-improvement, but I found that when I tried to open that PDF file, it looked like a bunch of gibberish (random control characters) in my PDF reader (Preview on OSX Leopard) That's odd. Maybe you should run Windows :-( Anyway I posted an HTML version. Not sure why PDF wouldn't work. I created both in OpenOffice. http://www.mattmahoney.net/rsi.pdf http://www.mattmahoney.net/rsi.html Anyone else have trouble reading the PDF version? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
On Mon, Oct 13, 2008 at 1:29 PM, Matt Mahoney [EMAIL PROTECTED] wrote: That's odd. Maybe you should run Windows :-( No. You should not run Windows --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
Hi, OK, I read the supposed refutation of recursive self-improvement at http://www.mattmahoney.net/rsi.html There are at least three extremely major problems with the argument. 1) By looking only at algorithmic information (defined in terms of program length) and ignoring runtime complexity, you are ignoring much of the value to be achieved via RSI. Suppose program P1 can solve problems of class C and size 500 in 3 hours per problem. Then, suppose P1 spends 50 hours transforming itself into a new program,P2, that can solve problems of class C and size 500 in one second per problem. Furthermore, suppose the RAM available in the machine at hand cannot hold bothP1 and P2 at the same time. In this case, it's obvious there's a huge advantage involved in P1 replacing itself withP2 ... if solving problems of class C is important for P1 achieving its goals, and if P2 is oriented toward achieving the same goal. Your argument is blind to this advantage because it ignores runtime complexity. Your argument is fixated on the fact that P2 can be generated by information consistingof {P1 plus the data P1 has observed} ... but so what? Program length is not, initself, all that useful thing to be looking at in the context of real-world computing. We need to be thinking about both space and time complexity. 2) You don't consider the program as interacting with an environment. IMO you shouldbe using the mathematical setup that Hutter uses in his main theorems about AIXI and AIXItl. In this setup, the AI is an agent that takes actions in an environment, which then responds to its actions. Furthermore, you should enhance Hutter's setup to consider the case where the agenthas not only fixed RAM (together potentially with a larger amount of memory that is slower to access), but also has processing cycle that is defined in terms of the cycle time of the environment, so that it only gets N internal processing cycles per each opportunity to sense/act. Considering the argument in this kind of more realistic setting, the critical importance of runtime as I noted above would immediately become apparent. 3) You don't consider that a smarter program might be able to figure out ways to increase its processor speed or RAM capacity, thus breaking your theoretical assumptions altogether. In this case, P2 could have an arbitrarily larger algorithmic information than P1, contradicting your result (by using a different, more realistic assumption). ... In short, what you have shown is that, according to an uninteresting measure (algorithmic information), RSI is not very dramatically useful in an artificial situation (no environment, no restrictions on processor cycle consumption, no ability for intelligence to lead to hardware modification). -- Ben G p.s. I read many PDF files each day using the same OS and viewer, and have never before seen the kind of problem I did with your pdf file. But I don't know what the source of the problem was. Anyway I read the HTML file just fine, thanks! On Mon, Oct 13, 2008 at 1:29 PM, Matt Mahoney [EMAIL PROTECTED] wrote: --- On Mon, 10/13/08, Ben Goertzel [EMAIL PROTECTED] wrote: I was eager to debunk your supposed debunking of recursive self-improvement, but I found that when I tried to open that PDF file, it looked like a bunch of gibberish (random control characters) in my PDF reader (Preview on OSX Leopard) That's odd. Maybe you should run Windows :-( Anyway I posted an HTML version. Not sure why PDF wouldn't work. I created both in OpenOffice. http://www.mattmahoney.net/rsi.pdf http://www.mattmahoney.net/rsi.html Anyone else have trouble reading the PDF version? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] Nothing will ever be attempted if all possible objections must be first overcome - Dr Samuel Johnson --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
Ben, Thanks for the comments on my RSI paper. To address your comments, 1. I defined improvement as achieving the same goal (utility) in less time or achieving greater utility in the same time. I don't understand your objection that I am ignoring run time complexity. 2. I agree that an AIXI type interactive environment is a more appropriate model than a Turing machine receiving all of its input at the beginning. The problem is how to formally define improvement in a way that distinguishes it from learning. I am open to suggestions. To see why this is a problem, consider an agent that after a long time, guesses the environment's program and is able to achieve maximum reward from that point forward. The agent could improve itself by hard-coding the environment's program into its successor and thereby achieve maximum reward right from the beginning. 3. A computer's processor speed and memory have no effect on the algorithmic complexity of a program running on it. -- Matt Mahoney, [EMAIL PROTECTED] --- On Mon, 10/13/08, Ben Goertzel [EMAIL PROTECTED] wrote: From: Ben Goertzel [EMAIL PROTECTED] Subject: Re: [agi] Updated AGI proposal (CMR v2.1) To: agi@v2.listbox.com Date: Monday, October 13, 2008, 8:33 PM Hi, OK, I read the supposed refutation of recursive self-improvement at http://www.mattmahoney.net/rsi.html There are at least three extremely major problems with the argument. 1) By looking only at algorithmic information (defined in terms of program length) and ignoring runtime complexity, you are ignoring much of the value to be achieved via RSI. Suppose program P1 can solve problems of class C and size 500 in 3 hours per problem. Then, suppose P1 spends 50 hours transforming itself into a new program,P2, that can solve problems of class C and size 500 in one second per problem. Furthermore, suppose the RAM available in the machine at hand cannot hold bothP1 and P2 at the same time. In this case, it's obvious there's a huge advantage involved in P1 replacing itself withP2 ... if solving problems of class C is important for P1 achieving its goals, and if P2 is oriented toward achieving the same goal. Your argument is blind to this advantage because it ignores runtime complexity. Your argument is fixated on the fact that P2 can be generated by information consistingof {P1 plus the data P1 has observed} ... but so what? Program length is not, initself, all that useful thing to be looking at in the context of real-world computing. We need to be thinking about both space and time complexity. 2) You don't consider the program as interacting with an environment. IMO you shouldbe using the mathematical setup that Hutter uses in his main theorems about AIXI and AIXItl. In this setup, the AI is an agent that takes actions in an environment, which then responds to its actions. Furthermore, you should enhance Hutter's setup to consider the case where the agenthas not only fixed RAM (together potentially with a larger amount of memory that is slower to access), but also has processing cycle that is defined in terms of the cycle time of the environment, so that it only gets N internal processing cycles per each opportunity to sense/act. Considering the argument in this kind of more realistic setting, the critical importance of runtime as I noted above would immediately become apparent. 3) You don't consider that a smarter program might be able to figure out ways to increase its processor speed or RAM capacity, thus breaking your theoretical assumptions altogether. In this case, P2 could have an arbitrarily larger algorithmic information than P1, contradicting your result (by using a different, more realistic assumption). ... In short, what you have shown is that, according to an uninteresting measure (algorithmic information), RSI is not very dramatically useful in an artificial situation (no environment, no restrictions on processor cycle consumption, no ability for intelligence to lead to hardware modification). -- Ben G --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
On Mon, Oct 13, 2008 at 11:30 PM, Matt Mahoney [EMAIL PROTECTED] wrote: Ben, Thanks for the comments on my RSI paper. To address your comments, You seem to be addressing minor lacunae in my wording, while ignoring my main conceptual and mathematical point!!! 1. I defined improvement as achieving the same goal (utility) in less time or achieving greater utility in the same time. I don't understand your objection that I am ignoring run time complexity. OK, you are not ignoring run time completely ... BUT ... in your measurement of the benefit achieved by RSI, you're not measuring the amount of run-time improvement achieved, you're only measuring algorithmic information. What matters in practice is, largely, the amount of run-time improvement achieved. This is the point I made in the details of my reply -- which you have not counter-replied to. I contend that, in my specific example, program P2 is a *huge* improvement over P1, in a way that is extremely important to practical AGI yet is not captured by your algorithmic-information-theoretic measurement method. What is your specific response to my example?? 2. I agree that an AIXI type interactive environment is a more appropriate model than a Turing machine receiving all of its input at the beginning. The problem is how to formally define improvement in a way that distinguishes it from learning. I am open to suggestions. To see why this is a problem, consider an agent that after a long time, guesses the environment's program and is able to achieve maximum reward from that point forward. The agent could improve itself by hard-coding the environment's program into its successor and thereby achieve maximum reward right from the beginning. Recursive self-improvement **is** a special case of learning; you can't completely distinguish them. 3. A computer's processor speed and memory have no effect on the algorithmic complexity of a program running on it. Yes, I can see I didn't phrase that point properly, sorry. I typed that prior email too hastily as I'm trying to get some work done ;-) The point I *wanted* to make in my third point, was that if you take a program with algorithmic information K, and give it the ability to modify its own hardware, then it can achieve algorithmic information M K. However, it is certainly true that this can happen even without the program modifying its own hardware -- especially if you make fanciful assumptions like Turing machines with huge tapes ... but even without such fanciful assumptions. The key point, which I did not articulate properly in my prior message, is that: ** by engaging with the world, the program can intake new information, which can increase its algorithmic information ** The new information a program P1 takes in from the **external world** may be random with regard to P1, yet may not be random with regard to {P1 + the new information taken in}. As self-modification may cause the intake of new information causing algorithmic information to increase arbitrarily much, your argument does not hold in the case of a program interacting with a world that has much higher algorithmic information than it does. And this of course is exactly the situation people are in. For instance, a program may learn that In the past, on 10 occasions, I have taken in information from Bob that was vastly beyond my algorithmic information content at that time. In each case this process helped me to achieve my goals, though in ways I would not have been able to understand before taking in the information. So, once again, I am going to trust Bob to alter me with info far beyond my current comprehension and algorithmic information content. Sounds a bit like a child trusting their parent, eh? This is a separate point from my point about P1 and P2 in point 1. But the two phenomena intersect, of course. -- Ben G This intake --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: [agi] Updated AGI proposal (CMR v2.1)
Ben, If you want to argue that recursive self improvement is a special case of learning, then I have no disagreement with the rest of your argument. But is this really a useful approach to solving AGI? A group of humans can generally make better decisions (more accurate predictions) by voting than any member of the group can. Did these humans improve themselves? My point is that a single person can't create much of anything, much less an AI smarter than himself. If it happens, it will be created by an organization of billions of humans. Without this organization, you would probably not think to create spears out of sticks and rocks. That is my problem with the seed AI approach. The seed AI depends on the knowledge and resources of the economy to do anything. An AI twice as smart as a human could not do any more than 2 people could. You need to create an AI that is billions of times smarter to get anywhere. We are already doing that. Human culture is improving itself by accumulating knowledge, by becoming better organized through communication and specialization, and by adding more babies and computers. -- Matt Mahoney, [EMAIL PROTECTED] --- On Mon, 10/13/08, Ben Goertzel [EMAIL PROTECTED] wrote: From: Ben Goertzel [EMAIL PROTECTED] Subject: Re: [agi] Updated AGI proposal (CMR v2.1) To: agi@v2.listbox.com Date: Monday, October 13, 2008, 11:46 PM On Mon, Oct 13, 2008 at 11:30 PM, Matt Mahoney [EMAIL PROTECTED] wrote: Ben, Thanks for the comments on my RSI paper. To address your comments, You seem to be addressing minor lacunae in my wording, while ignoring my main conceptual and mathematical point!!! 1. I defined improvement as achieving the same goal (utility) in less time or achieving greater utility in the same time. I don't understand your objection that I am ignoring run time complexity. OK, you are not ignoring run time completely ... BUT ... in your measurement of the benefit achieved by RSI, you're not measuring the amount of run-time improvement achieved, you're only measuring algorithmic information. What matters in practice is, largely, the amount of run-time improvement achieved. This is the point I made in the details of my reply -- which you have not counter-replied to. I contend that, in my specific example, program P2 is a *huge* improvement over P1, in a way that is extremely important to practical AGI yet is not captured by your algorithmic-information-theoretic measurement method. What is your specific response to my example?? 2. I agree that an AIXI type interactive environment is a more appropriate model than a Turing machine receiving all of its input at the beginning. The problem is how to formally define improvement in a way that distinguishes it from learning. I am open to suggestions. To see why this is a problem, consider an agent that after a long time, guesses the environment's program and is able to achieve maximum reward from that point forward. The agent could improve itself by hard-coding the environment's program into its successor and thereby achieve maximum reward right from the beginning. Recursive self-improvement **is** a special case of learning; you can't completely distinguish them. 3. A computer's processor speed and memory have no effect on the algorithmic complexity of a program running on it. Yes, I can see I didn't phrase that point properly, sorry. I typed that prior email too hastily as I'm trying to get some work done ;-) The point I *wanted* to make in my third point, was that if you take a program with algorithmic information K, and give it the ability to modify its own hardware, then it can achieve algorithmic information M K. However, it is certainly true that this can happen even without the program modifying its own hardware -- especially if you make fanciful assumptions like Turing machines with huge tapes ... but even without such fanciful assumptions. The key point, which I did not articulate properly in my prior message, is that: ** by engaging with the world, the program can intake new information, which can increase its algorithmic information ** The new information a program P1 takes in from the **external world** may be random with regard to P1, yet may not be random with regard to {P1 + the new information taken in}. As self-modification may cause the intake of new information causing algorithmic information to increase arbitrarily much, your argument does not hold in the case of a program interacting with a world that has much higher algorithmic information than it does. And this of course is exactly the situation people are in. For instance, a program may learn that In the past, on 10 occasions, I have taken in information from Bob that was vastly beyond my algorithmic information content at that time. In each case this process helped me to achieve my goals, though in ways I would not have been able to understand before taking
Re: [agi] Updated AGI proposal (CMR v2.1)
OK, well now you are backing away from your claim of a mathematical disproof of RSI!! What you did IMHO was to prove there is limited value in RSI by defining RSI in a very limited way, and then measuring the value of this limited-RSI in a manner that does not capture the practical value of any kind of RSI... I don't agree that an AGI will be programmed by billions of humans. I think an AGI will be created by a fairly small team of programmers and scientists. Of course, this effort will build atop the prior work of a large number of other scientists and engineers -- the ones who built the computer chips, the Internet, the programming languages, and so forth. But I see no reason why the actual programming and design of the AGI can't be done by a fairly small team... I agree that RSI is not how human intelligence predominantly works, but my goal is not to replicate human intelligence, rather to create better forms of intelligence that can help humans better than we can help ourselves directly ... and can also move on to levels inaccessible to humans... -- Ben G On Tue, Oct 14, 2008 at 12:36 AM, Matt Mahoney [EMAIL PROTECTED] wrote: Ben, If you want to argue that recursive self improvement is a special case of learning, then I have no disagreement with the rest of your argument. But is this really a useful approach to solving AGI? A group of humans can generally make better decisions (more accurate predictions) by voting than any member of the group can. Did these humans improve themselves? My point is that a single person can't create much of anything, much less an AI smarter than himself. If it happens, it will be created by an organization of billions of humans. Without this organization, you would probably not think to create spears out of sticks and rocks. That is my problem with the seed AI approach. The seed AI depends on the knowledge and resources of the economy to do anything. An AI twice as smart as a human could not do any more than 2 people could. You need to create an AI that is billions of times smarter to get anywhere. We are already doing that. Human culture is improving itself by accumulating knowledge, by becoming better organized through communication and specialization, and by adding more babies and computers. -- Matt Mahoney, [EMAIL PROTECTED] --- On Mon, 10/13/08, Ben Goertzel [EMAIL PROTECTED] wrote: From: Ben Goertzel [EMAIL PROTECTED] Subject: Re: [agi] Updated AGI proposal (CMR v2.1) To: agi@v2.listbox.com Date: Monday, October 13, 2008, 11:46 PM On Mon, Oct 13, 2008 at 11:30 PM, Matt Mahoney [EMAIL PROTECTED] wrote: Ben, Thanks for the comments on my RSI paper. To address your comments, You seem to be addressing minor lacunae in my wording, while ignoring my main conceptual and mathematical point!!! 1. I defined improvement as achieving the same goal (utility) in less time or achieving greater utility in the same time. I don't understand your objection that I am ignoring run time complexity. OK, you are not ignoring run time completely ... BUT ... in your measurement of the benefit achieved by RSI, you're not measuring the amount of run-time improvement achieved, you're only measuring algorithmic information. What matters in practice is, largely, the amount of run-time improvement achieved. This is the point I made in the details of my reply -- which you have not counter-replied to. I contend that, in my specific example, program P2 is a *huge* improvement over P1, in a way that is extremely important to practical AGI yet is not captured by your algorithmic-information-theoretic measurement method. What is your specific response to my example?? 2. I agree that an AIXI type interactive environment is a more appropriate model than a Turing machine receiving all of its input at the beginning. The problem is how to formally define improvement in a way that distinguishes it from learning. I am open to suggestions. To see why this is a problem, consider an agent that after a long time, guesses the environment's program and is able to achieve maximum reward from that point forward. The agent could improve itself by hard-coding the environment's program into its successor and thereby achieve maximum reward right from the beginning. Recursive self-improvement **is** a special case of learning; you can't completely distinguish them. 3. A computer's processor speed and memory have no effect on the algorithmic complexity of a program running on it. Yes, I can see I didn't phrase that point properly, sorry. I typed that prior email too hastily as I'm trying to get some work done ;-) The point I *wanted* to make in my third point, was that if you take a program with algorithmic information K, and give it the ability to modify its own hardware, then it can achieve algorithmic information M K. However, it is certainly true that this can happen even without the program modifying