Re: [singularity] Quantum Mechanics and Consciousness
predicts that the curve would have peaks at certain levels of probability of getting the right answer above those predicted by chance alone. Experimental data showed peaks at the locations modeled. However, more people were successful at the higher probability levels than Walker's model estimated. This is considered to be evidence of learning enhancement [5]. In the world of the weird and unexplained you; are left to imagine with; mysterious metaphors and thoughts that dont allow understanding audiences. Bertromavich 'He who receives an idea from me, receives instruction himself without lessening mine; as he who lights his taper at mine, receives light without darkening me.' Thomas Jefferson, letter to Isaac McPherson, 13 August 1813 -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=104200892-0d3a07 Powered by Listbox: http://www.listbox.com
Re: [singularity] future of mankind blueprint and relevance of AGI
--- Minwoo Bae [EMAIL PROTECTED] wrote: This isn't totally relevant, but have you heard of Korea's drafting of a robot ethics charter? You mean http://news.nationalgeographic.com/news/2007/03/070316-robot-ethics.html ? It seems mainly focused on protecting humans. But the proposal was a year ago and nothing was released yet. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=104200892-0d3a07 Powered by Listbox: http://www.listbox.com
Re: [singularity] An Open Letter to AGI Investors
--- Richard Loosemore [EMAIL PROTECTED] wrote: I have stuck my neck out and written an Open Letter to AGI (Artificial General Intelligence) Investors on my website at http://susaro.com. All part of a campaign to get this field jumpstarted. Next week I am going to put up a road map for my own development project. So if the value of AGI is all the human labor it replaces (about US $1 quadrillion), how much will it cost to build? Keep in mind there is a tradeoff between waiting for the cost of technology to drop vs. having it now. How much should we expect to spend? -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=101816851-9a120b Powered by Listbox: http://www.listbox.com
Re: [singularity] Stronger than Turing?
--- Ben Peterson [EMAIL PROTECTED] wrote: Maybe I'm hallucinating, but I thought I read somewhere of some test stronger or more reliable than the Turing Test to verify whether or not a machine had achieved human-level intelligence. Text compression? http://cs.fit.edu/~mmahoney/compression/rationale.html I wouldn't say it is more powerful, just more objective and repeatable. Also, in its present form it can only be used to compare one model to another. To test whether a model achieves human level, it needs to be compared to average human ability to predict successive words or symbols in a text stream. This is a harder test to get right, one I have not yet attempted. Shannon [1] first did this test in 1950 but left a wide range of uncertainty (0.6 to 1.3 bits per character) due to his method of converting a ranking of next-letter guesses to a probability distribution. Cover and King [2] reduced the uncertainty in 1978 (upper bound of 1.3 bpc) by making the probability distribution explicit in a gambling game, but their method is time consuming and could only be used on a small sample of text. I have also made some attempts to refine Shannon's method in http://cs.fit.edu/~mmahoney/dissertation/entropy1.html (under 1.1 bpc). In any case, none of these measurements were on the actual test data used in my large text benchmark. The best result to date is 1.04 bpc, but I would not call this AI. I know these programs use rather simple language models and are memory bound. (The top program needs 4.6 GB). The Wikipedia data set I use probably has a lower entropy than the data used in the literature, possibly 0.8-0.9 bpc. That's just a guess, because as I said, I don't yet have a reliable way to measure it. References 1. Shannon, Cluade E., Prediction and Entropy of Printed English, Bell Sys. Tech. J (3) p. 50-64, 1950. 2. Cover, T. M., and R. C. King, A Convergent Gambling Estimate of the Entropy of English, IEEE Transactions on Information Theory (24)4 (July) pp. 413-421, 1978. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
Testing AGI (was RE: [singularity] Vista/AGI)
--- Derek Zahn [EMAIL PROTECTED] wrote: At any rate, if there were some clearly-specified tests that are not AGI-complete and yet not easily attackable with straightforward software engineering or Narrow AI techniques, that would be a huge boost in my opinion to this field. I can't think of any though, and they might not exist. If it is in fact impossible to find such tasks, what does that say about AGI as an endeavor? Text compression is one such test, as I argue in http://cs.fit.edu/~mmahoney/compression/rationale.html The test is only for language modeling. Theoretically it could be extended to vision or audio processing. For example, to maximally compress video the compressor must understand the physics of the scene (e.g. objects fall down), which can be arbitrarily complex (e.g. a video of people engaging in conversation about Newton's law of gravity). Likewise, maximally compressing music is equivalent to generating or recognizing music that people like. The problem is that the information content of video and audio is dominated by incompressible noise that is nontrivial to remove -- noise being any part of the signal that people fail to perceive. Deciding which parts of the signal are noise is itself AI-hard, so it requires a lossy compression test with human judges making subjective decisions about quality. This is not a big problem for text because the noise level (different ways of expressing the same meaning) is small, or at least does not overwhelm the signal. Long term memory has an information rate of a few bits per second, so any signal you compress should not be many orders of magnitude higher. A problem with text compression is the lack of adequate hardware. There is a 3 way tradeoff between compression ratio, memory, and speed. The top compressor in http://cs.fit.edu/~mmahoney/compression/text.html uses 4.6 GB of memory. Many of the best algorithms could be drastically improved if only they ran on a supercomputer with 100 GB or more. The result is that most compression gains come from speed and memory optimization rather than using more intelligent models. The best compressors use crude models of semantics and grammar. They preprocess the text by token substitution from a dictionary that groups words by topic and grammatical role, then predict the token stream using mixtures of fixed-offset context models. It is roughly equivalent to the ungrounded language model of a 2 or 3 year old child at best. An alternative would be to reduce the size of the test set to reduce computational requirements, as the Hutter prize did. http://prize.hutter1.net/ I did not because I believe the proper way to test an adult level language model is to train it on the same amount of language that an average adult is exposed to, about 1 GB. I would be surprised if a 100 MB test progressed past the level of a 3 year old child. I believe the data set is too small to train a model to learn arithmetic, logic, or high level reasoning. Including these capabilities would not improve compression. Tests on small data sets could be used to gauge early progress. But ultimately, I think you are going to need hardware that supports AGI to test it. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
Re: About the Nine Misunderstandings post [WAS Re: [singularity] I'm just not sure how well...]
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: We already have examples of reproducing agents: Code Red, SQL Slammer, Storm, etc. A worm that can write and debug code and discover new vulnerabilities will be unstoppable. Do you really think your AI will win the race when you have the extra burden of making it safe? Yes, because these reproducing agents you refer to are the most laughably small computer viruses that have no hope whatsoever of becoming generally intelligent. At every turn, you completely undestimate what it means for a system to be intelligent. There are no intelligent or self improving worms... yet. Are you confident that none will ever be created even after we have automated human-level understanding of code, which I presume will be one of the capabilities of AGI? Also, RSI is an experimental process, and therefore evolutionary. We have already gone through the information theoretic argument why this must be the case. No you have not: I know of no information theoretic argument that even remotely applies to the type of system that is needed to achieve real intelligence. Furthermore, the statement that RSI is an experimental process, and therefore evolutionary is just another example of you declaring something to be true when, in fact, it is loaded down with spurious assumptions. Your statement is a complete non-sequiteur. (sigh) To repeat, the argument is that an agent cannot deterministically create an agent of greater intelligence than itself, because if it could it would already be that smart. The best it can do is make educated guesses as to what will increase intelligence. I don't argue that we can't do better than evolution. (Adding more hardware is probably a safe bet). But an agent cannot even test whether another is more intelligent. In order for me to give a formal argument, you would have to accept a formal definition of intelligence, such as Hutter and Legg's univeral intelligence, which is bounded by algorithmic complexity. But you dismiss such definitions as irrelevant. So I can only give examples, such as the ability to measure an IQ of 200 in children but not adults, and the historical persecution of intelligence (Socrates, Galileo, Holocaust, Khmer Rouge, etc). A self improving agent will have to produce experimental variations and let them be tested in a competitive environment it doesn't control or fully understand that weeds out the weak. If it could model the environment or test for intelligence then it could reliably improve its intelligence, contradicting our original assumption. This is an evolutionary process. Unfortunately, evolution is not stable. It resides on the boundary between stability and chaos, like all incrementally updated or adaptive algorithmically complex systems. By this I mean it tends to a Lyapunov exponent of 0. A small perturbation in its initial state might decay or it might grow. Critically balanced systems like this have a Zipf distribution of catastrophes -- an inverse relation between probability and severity. We find this property in randomly connected logic gates (frequency vs. magnitude of state transitions) software systems (frequency vs. severity of failures), gene regulatory systems (frequency vs. severity of mutations), and evolution (frequency vs. severity of plagues, population explosions, mass extinctions, and other ecological disasters). The latter should be evident in the hierarchical organization of geologic eras. And a singularity is a catastrophe of unprecedented scale. It could result in the extinction of DNA based life and its replacement with nanotechnology. Or it could result in the extinction of all intelligence. The only stable attractor in evolution is a dead planet. (You knew this, right?) Finally, I should note that intelligence and friendliness are not the same as fitness. Roaches, malaria, and HIV are all formidable competitors to homo sapiens. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
Re: About the Nine Misunderstandings post [WAS Re: [singularity] I'm just not sure how well...]
--- Richard Loosemore [EMAIL PROTECTED] wrote: You want me to imagine a scenario in which we have AGI, but in your scenario these AGI systems are somehow not being used to produce superintelligent systems, and these superintelligent systems are, for some reason, not taking the elementary steps necessary to solve one of the world's simplest problems (computer viruses). If the problem is so simple, why don't you just solve it? http://www.securitystats.com/ http://en.wikipedia.org/wiki/Storm_botnet There is a trend toward using (narrow) AI for security. It seems to be one of its biggest applications. Unfortunately, the knowledge needed to secure computers is almost exactly the same kind of knowledge needed to attack them. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
Re: About the Nine Misunderstandings post [WAS Re: [singularity] I'm just not sure how well...]
--- Vladimir Nesov [EMAIL PROTECTED] wrote: On Fri, Apr 11, 2008 at 10:50 PM, Matt Mahoney [EMAIL PROTECTED] wrote: If the problem is so simple, why don't you just solve it? http://www.securitystats.com/ http://en.wikipedia.org/wiki/Storm_botnet There is a trend toward using (narrow) AI for security. It seems to be one of its biggest applications. Unfortunately, the knowledge needed to secure computers is almost exactly the same kind of knowledge needed to attack them. Matt, this issue was already raised a couple of times. It's a technical problem that can be solved perfectly, but isn't in practice, because it's too costly. Formal verification, specifically aided by languages with rich type systems that can express proofs of correctness for complex properties, can give you perfectly safe systems. It's just very difficult to specify all the details. Actually it cannot be solved even theoretically. A formal specification of a program is itself a program. It is undecidable whether two programs are equivalent. (It is equivalent to the halting problem). Converting natural language to a formal specification is AI-hard, or perhaps harder, because people can't get it right either. If we could write software without bugs, we would solve a big part of the security problem. These AIs for network security that you are talking about are a cost-effective hack that happens to work sometimes. It's not a low-budget vision of future super-hacks. Not at present because we don't have AI. We rely on humans to find vulnerabilities in software. We would like for machines to do this automatically. Unfortunately such machines would also be useful to hackers. Such double-edged tools already exist. For example, tools like SATAN, NESSES, and NMAP can quickly test a system by probing it to look for thousands of known or published vulnerabilities. Attackers use the same tools to break into systems. www.virustotal.com allows you to upload a file and scan it with 32 different virus detectors. This is a useful tool for virus writers who want to make sure their programs evade detection. I suggest it will be very difficult to develop any security tool that you could keep out of the hands of the bad guys. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
Re: [singularity] I'm just not sure how well this plan was thought through [WAS Re: Promoting an A.S.P.C,A.G.I.]
--- Richard Loosemore [EMAIL PROTECTED] wrote: When a computer processes a request like how many teaspoons in a cubic parsec? it can extract the meaning of the question by a relatively simple set of syntactic rules and question templates. But when you ask it a question like how many dildos are there on the planet? [Try it] you find that google cannot answer this superficially similar question because it requires more intelligence in the question-analysis mechanism. And just how would you expect your AGI to answer the question? The first step in research is to find out if someone else has already answered it. It may have been answered but Google can't find it because it only indexes a small fraction of the internet. It may also be that some dildo makers are privately held and don't release sales figures. In any case your AGI is either going to output a number or I don't know, neither of which is more helpful than Google. If it does output a number, you are still going to want to know where it came from. But this discussion is tiresome. I would not have expected you to anticipate today's internet in 1978. I suppose when the first search engine (Archie) was released in 1990, you probably imagined that all search engines would require you to know the name of the file you were looking for. If you have a better plan for AGI, please let me know. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
RE: Promoting AGI (RE: [singularity] Vista/AGI)
--- John G. Rose [EMAIL PROTECTED] wrote: From: Matt Mahoney [mailto:[EMAIL PROTECTED] The simulations can't loop because the simulator needs at least as much memory as the machine being simulated. You're making assumptions when you say that. Outside of a particular simulation we don't know the rules. If this universe is simulated the simulator's reality could be so drastically and unimaginably different from the laws in this universe. Also there could be data busses between simulations and the simulations could intersect or, a simulation may break the constraints of its contained simulation somehow and tunnel out. I am assuming finite memory. For the universe we observe, the Bekenstein bound of the Hubble radius is 2pi^2 T^2 c^5/hG = 2.91 x 10^122 bits. (T = age of the universe = 13.7 billion years, c = speed of light, h = Planck's constant, G = gravitational constant). There is not enough material in the universe to build a larger memory. However, a universe up the hierarchy might be simulated by a Turing machine with infinite memory or by a more powerful machine such as one with real-valued registers. In that case the restriction does not apply. For example, a real-valued function can contain nested copies of itself infinitely deep. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
Re: Promoting AGI (RE: [singularity] Vista/AGI)
--- Mike Tintner [EMAIL PROTECTED] wrote: How do you resolve disagreements? This is a problem for all large databases and multiuser AI systems. In my design, messages are identified by source (not necessarily a person) and a timestamp. The network economy rewards those sources that provide the most useful (correct) information. There is an incentive to produce reputation managers which rank other sources and forward messages from highly ranked sources, because those managers themselves become highly ranked. Google handles this problem by using its PageRank algorithm, although I believe that better (not perfect) solutions are possible in a distributed, competitive environment. I believe that these solutions will be deployed early and be the subject of intense research because it is such a large problem. The network I described is vulnerable to spammers and hackers deliberately injecting false or forged information. The protocol can only do so much. I designed it to minimize these risks. Thus, there is no procedure to delete or alter messages once they are posted. Message recipients are responsible for verifying the identity and timestamps of senders and for filtering spam and malicious messages at risk of having their own reputations lowered if they fail. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
Re: Promoting AGI (RE: [singularity] Vista/AGI)
--- Ben Goertzel [EMAIL PROTECTED] wrote: Of course what I imagine emerging from the Internet bears little resemblance to Novamente. It is simply too big to invest in directly, but it will present many opportunities. But the emergence of superhuman AGI's like a Novamente may eventually become, will both dramatically alter the nature of, and dramatically reduce the cost of, global brains such as you envision... Yes, like the difference between writing a web browser and defining the HTTP protocol, each costing a tiny fraction of the value of the Internet but with a huge impact on its outcome. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
Re: [singularity] Re: Promoting an A.S.P.C,A.G.I.
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: Perhaps you have not read my proposal at http://www.mattmahoney.net/agi.html or don't understand it. Some of us have read it, and it has nothing whatsoever to do with Artificial Intelligence. It is a labor-intensive search engine, nothing more. I have no idea why you would call it an AI or an AGI. It is not autonomous, contains no thinking mechanisms, nothing. Even as a alabor intensive search engine there is no guarantee it would work, because the conflict resolution issues are all complexity-governed. I am astonished that you would so blatantly call it something that it is not. It is not now. I think it will be in 30 years. If I was to describe the Internet to you in 1978 I think you would scoff too. We were supposed to have flying cars and robotic butlers by now. How could Google make $145 billion by building an index of something that didn't even exist? Just what do you want out of AGI? Something that thinks like a person or something that does what you ask it to? -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
Re: [singularity] Re: Promoting an A.S.P.C,A.G.I.
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: Just what do you want out of AGI? Something that thinks like a person or something that does what you ask it to? Either will do: your suggestion achieves neither. If I ask your non-AGI the following question: How can I build an AGI that can think at a speed that is 1000 times faster than the speed of human thought? it will say: Hi, my name is Ben and I just picked up your question. I would love to give you the answer but you have to send $20 million and give me a few years. That is not the answer I would expect of an AGI. A real AGI would do original research to solve the problem, and solve it *itself*. Isn't this, like, just too obvious for words? ;-) Your question is not well formed. Computers can already think 1000 times faster than humans for things like arithmetic. Does your AGI also need to know how to feed your dog? Or should it guess and build it anyway? I would think such a system would be dangerous. I expect a competitive message passing network to improve over time. Early versions will work like an interactive search engine. You may get web pages or an answer from another human in real time, and you may later receive responses to your persistent query. If your question can be matched to an expert in some domain that happens to be on the net, then it gets routed there. Google already does this. For example, if you type an address, it gives you a map and offers driving directions. If you ask it how many teaspoons in a cubic parsec? it will compute the answer (try it). It won't answer every question, but with 1000 times more computing power than Google, I expect there will be many more domain experts. I expect as hardware gets more powerful, peers will get better at things like recognizing people in images, writing programs, and doing original research. I don't claim that I can solve these problems. I do claim that there is an incentive to provide these services and that the problems are not intractable given powerful hardware, and therefore the services will be provided. There are two things to make the problem easier. First, peers will have access to a vast knowledge source that does not exist today. Second, peers can specialize in a narrow domain, e.g. only recognize one particular person in images, or write software or do research in some obscure, specialized field. Is this labor intensive? Yes. A $1 quadrillion system won't just build itself. People will build it because they will get back more value than they put in. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
Promoting AGI (RE: [singularity] Vista/AGI)
--- Eric B. Ramsay [EMAIL PROTECTED] wrote: John G. Rose [EMAIL PROTECTED] wrote: If you look at the state of internet based intelligence now, all the data and its structure, the potential for chain reaction or a sort of structural vacuum exists and it is accumulating a potential at an increasing rate. IMO... So you see the arrival of a Tipping Point as per Malcolm Gladwell. Whether I physically benefit from the arrival of the Singularity or not, I just want to see the damn thing. I would invest some modest sums in AGI if we could get a huge collection plate going around (these collection plate amounts add up!). You won't see a singularity. As I explain in http://www.mattmahoney.net/singularity.html an intelligent agent (you) is not capable of recognizing agents of significantly greater intelligence. We don't know whether a singularity has already occurred and the world we observe is the result. It is consistent with the possibility, e.g. it is finite, Turing computable, and obeys Occam's Razor (AIXI). As for AGI research, I believe the most viable path is a distributed architecture that uses the billions of human brains and computers already on the Internet. What is needed is an infrastructure that routes information to the right experts and an economy that rewards intelligence and friendliness. I described one such architecture in http://www.mattmahoney.net/agi.html It differs significantly from the usual approach of trying to replicate a human mind. I don't believe that one person or a small group can solve the AGI problem faster than the billions of people on the Internet are already doing. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
RE: Promoting AGI (RE: [singularity] Vista/AGI)
--- Derek Zahn [EMAIL PROTECTED] wrote: Matt Mahoney writes: As for AGI research, I believe the most viable path is a distributed architecture that uses the billions of human brains and computers already on the Internet. What is needed is an infrastructure that routes information to the right experts and an economy that rewards intelligence and friendliness. I described one such architecture in http://www.mattmahoney.net/agi.html It differs significantly from the usual approach of trying to replicate a human mind. I don't believe that one person or a small group can solve the AGI problem faster than the billions of people on the Internet are already doing. I'm not sure I understand this. Although a system that can respond well to commands of the following form: Show me an existing document that best answers the question 'X' is certainly useful, it is hardly 'general' in any sense we usually mean. I would think a 'general' intelligence should be able to take a shot at answering: Why are so many streets named after trees? or If the New York Giants played cricket against the New York Yankees, who would probably win? or Here are the results of some diagnostic tests. How likely is it that the patient has cancer? What test should we do next? or Design me a stable helicopter with the rotors on the bottom instead of the top Super-google is nifty, but I don't see how it is AGI. Because a super-google will answer these questions by routing them to experts on these topics that will use natural language in their narrow domains of expertise. All of this can be done with existing technology and a lot of hard work. The work will be done because there is an incentive to do it and because the AGI (in the system, not its components) is so valuable. AGI will be an extension of the Internet that nobody planned, nobody built, and nobody owns. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
RE: Promoting AGI (RE: [singularity] Vista/AGI)
--- John G. Rose [EMAIL PROTECTED] wrote: There is no way to know if we are living in a nested simulation, or even in a single simulation. However there is a mathematical model: enumerate all Turing machines to find one that simulates a universe with intelligent life. What if that nest of simulations loop around somehow? What was that idea where there is this new advanced microscope that can see smaller than ever before and you look into it and see an image of yourself looking into it... The simulations can't loop because the simulator needs at least as much memory as the machine being simulated. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
Re: Promoting AGI (RE: [singularity] Vista/AGI)
--- Mike Tintner [EMAIL PROTECTED] wrote: Matt : a super-google will answer these questions by routing them to experts on these topics that will use natural language in their narrow domains of expertise. And Santa will answer every child's request, and we'll all live happily ever after. Amen. If you have a legitimate criticism of the technology or its funding plan, I would like to hear it. I understand there will be doubts about a system I expect to cost over $1 quadrillion and take 30 years to build. The protocol specifies natural language. This is not a hard problem in narrow domains. It dates back to the 1960's. Even in broad domains, most of the meaning of a message is independent of word order. Google works on this principle. But this is beside the point. The critical part of the design is an incentive for peers to provide useful services in exchange for resources. Peers that appear most intelligent and useful (and least annoying) are most likely to have their messages accepted and forwarded by other peers. People will develop domain experts and routers and put them on the net because they can make money through highly targeted advertising. Google would be a peer on the network with a high reputation. But Google controls only 0.1% of the computing power on the Internet. It will have to compete with a system that allows updates to be searched instantly, where queries are persistent, and where a query or message can initiate conversations with other people in real time. Which are these areas of science, technology, arts, or indeed any area of human activity, period, where the experts all agree and are NOT in deep conflict? And if that's too hard a question, which are the areas of AI or AGI, where the experts all agree and are not in deep conflict? I don't expect the experts to agree. It is better that they don't. There are hard problem remaining to be solved in language modeling, vision, and robotics. We need to try many approaches with powerful hardware. The network will decide who the winners are. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
Re: Promoting AGI (RE: [singularity] Vista/AGI)
--- Eric B. Ramsay [EMAIL PROTECTED] wrote: If I understand what I have read in this thread so far, there is Ben on the one hand suggesting $10 mil. with 10-30 people in 3 to 10 years and on the other there is Matt saying $1quadrillion, using a billion brains in 30 years. I don't believe I have ever seen such a divergence of opinion before on what is required for a technological breakthrough (unless people are not being serious and I am being naive). I suppose this sort of non-consensus on such a scale could be part of investor reticence. I am serious about the $1 quadrillion price tag, which is the low end of my estimate. The value of the Internet is now in the tens of trillions and doubling every few years. The value of AGI will be a very large fraction of the world economy, currently US $66 trillion per year and growing at 5% per year. Of course what I imagine emerging from the Internet bears little resemblance to Novamente. It is simply too big to invest in directly, but it will present many opportunities. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
Re: [singularity] future search
--- David Hart [EMAIL PROTECTED] wrote: Hi All, I'm quite worried about Google's new *Machine Automated Temporal Extrapolation* technology going FOOM! http://www.google.com.au/intl/en/gday/ More on the technology http://en.wikipedia.org/wiki/Google's_hoaxes :-) -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=98631122-712fa4 Powered by Listbox: http://www.listbox.com
Re: [singularity] Re: Revised version of Jaron Lanier's thought experiment.
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: --- John G. Rose [EMAIL PROTECTED] wrote: Is there really a bit per synapse? Is representing a synapse with a bit an accurate enough simulation? One synapse is a very complicated system. A typical neural network simulation uses several bits per synapse. A Hopfield net implementation of an associative memory stores 0.15 bits per synapse. But cognitive models suggest the human brain stores .01 bits per synapse. (There are 10^15 synapses but human long term memory capacity is 10^9 bits). Sorry, I don't buy this at all. This makes profound assumptions about how information is stored in memory, averagng out the net storage and ignoring the immediate storage capacity. A typical synapse actually stores a great deal more than a fraction of a bit, as far as we can tell, but this information is stored in such a way that the system as a whole can actually use the information in a meaningful way. In that context, quoting 0.01 bits per synapse is a completely meaningless statement. I was referring to Landauer's estimate of long term memory learning rate of about 2 bits per second. http://www.merkle.com/humanMemory.html This does not include procedural memory, things like visual perception and knowing how to walk. So 10^-6 bits is low. But how do we measure such things? Also, typical neural network simulations use more than a few bits as well. When I did a number of backprop NN studies in the early 90s, my networks had to use floating point numbers because the behavior of the net deteriorated badly if the numerical precision was reduced. This was especially important on long training runs or large datasets. That's what I meant by few. In the PAQ8 compressors I have to use at least 16 bits. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=96140713-a54b2b Powered by Listbox: http://www.listbox.com
Re: [singularity] Re: Revised version of Jaron Lanier's thought experiment.
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: I was referring to Landauer's estimate of long term memory learning rate of about 2 bits per second. http://www.merkle.com/humanMemory.html This does not include procedural memory, things like visual perception and knowing how to walk. So 10^-6 bits is low. But how do we measure such things? I think my general point is that bits per second or bits per synapse is a valid measure if you care about something like an electrical signal line, but is just simply an incoherent way to talk about the memory capacity of the human brain. Saying 0.01 bits per synapse is no better than opening and closing one's mouth without saying anything. Bits is a perfectly sensible measure of information. Memory can be measured using human recall tests, just as Shannon used human prediction tests to estimate the information capacity of natural language text. The question is important to anyone who needs to allocate a hardware budget for an AI design. [For those not familiar with Richard's style: once he disagrees with something he will dispute it to the bitter end in long, drawn out arguments, because nothing is more important than being right.] -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=96140713-a54b2b Powered by Listbox: http://www.listbox.com
Re: [singularity] Re: Revised version of Jaron Lanier's thought experiment.
--- Eric B. Ramsay [EMAIL PROTECTED] wrote: Matt Mahoney [EMAIL PROTECTED] wrote: [For those not familiar with Richard's style: once he disagrees with something he will dispute it to the bitter end in long, drawn out arguments, because nothing is more important than being right.] What's the purpose for this comment? If the people here are intelligent enough to have meaningful discussions on a difficult topic, then they will be able to sort out for themselves the styles of others. Sorry, he posted a similar comment about me on the AGI list. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=96140713-a54b2b Powered by Listbox: http://www.listbox.com
RE: [singularity] Re: Revised version of Jaron Lanier's thought experiment.
--- John G. Rose [EMAIL PROTECTED] wrote: Is there really a bit per synapse? Is representing a synapse with a bit an accurate enough simulation? One synapse is a very complicated system. A typical neural network simulation uses several bits per synapse. A Hopfield net implementation of an associative memory stores 0.15 bits per synapse. But cognitive models suggest the human brain stores .01 bits per synapse. (There are 10^15 synapses but human long term memory capacity is 10^9 bits). -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=96140713-a54b2b Powered by Listbox: http://www.listbox.com
Re: [singularity] Re: Revised version of Jaron Lanier's thought experiment.
--- Stathis Papaioannou [EMAIL PROTECTED] wrote: On 28/02/2008, John G. Rose [EMAIL PROTECTED] wrote: Actually a better way to do it as getting even just the molecules right is a wee bit formidable - you need a really powerful computer with lots of RAM. Take some DNA and grow a body double in software. Then create an interface from the biological brain to the software brain and then gradually kill off the biological brain forcing the consciousness into the software brain. The problem with this approach naturally is that to grow the brain in RAM requires astronomical resources. But ordinary off-the-shelf matter holds so much digital memory compared to modern computers. You have to convert matter into RAM somehow. For example one cell with DNA is how many gigs? And cells cost a dime a billion. But the problem is that molecular interaction is too slow and cluncky. Agreed, it would be *enormously* difficult getting a snapshot at the molecular level and then doing a simulation from this snapshot. But as a matter of principle, it should be possible. And that is the whole point. You don't need to simulate the brain at the molecular level or even at the level of neurons. You just need to produce an equivalent computation. The whole point of such fine grained simulations is to counter arguments (like Penrose's) that qualia and consciousness cannot be explained by computation or even by physics. Penrose (like all humans) is reasoning with a brain that is a product of evolution, and therefore biased toward beliefs that favor survival of the species. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=96140713-a54b2b Powered by Listbox: http://www.listbox.com
Re: [singularity] Re: Revised version of Jaron Lanier's thought experiment.
--- Stathis Papaioannou [EMAIL PROTECTED] wrote: On 29/02/2008, Matt Mahoney [EMAIL PROTECTED] wrote: By equivalent computation I mean one whose behavior is indistinguishable from the brain, not an approximation. I don't believe that an exact simulation requires copying the implementation down to the neuron level, much less the molecular level. How do you explain the fact that cognition is exquisitely sensitive to changes at the molecular level? In what way? Why can't you replace neurons with equivalent software? -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=96140713-a54b2b Powered by Listbox: http://www.listbox.com
RE: [singularity] Re: Revised version of Jaron Lanier's thought experiment.
--- John G. Rose [EMAIL PROTECTED] wrote: From: Matt Mahoney [mailto:[EMAIL PROTECTED] By equivalent computation I mean one whose behavior is indistinguishable from the brain, not an approximation. I don't believe that an exact simulation requires copying the implementation down to the neuron level, much less the molecular level. So how would you approach constructing such a model? I suppose a superset intelligence structure could analyze properties and behaviors of a brain and simulate it within itself. If it absorbed enough data it could reconstruct and eventually come up with something close. Well, nobody has solved the AI problem, much less the uploading problem. Consider the problem in stages: 1. The Turing test. 2. The personalized Turing test. The machine pretends to be you and the judges are people who know you well. 3. The planned, personalized Turing test. You are allowed to communicate with judges in advance, for example, to agree on a password. 4. The embodied, planned, personalized Turing test. Communication is not restricted to text. The machine is planted in the skull of your clone. Your friends and relatives have to decide who has the carbon-based brain. Level 4 should not require simulating every neuron and synapse. Without the constraints of slow, noisy neurons, we could use other algorithms. For example, low level visual processing such as edge and line detection would not need to be implemented as a 2-D array of identical filters. It could be implemented serially by scanning the retinal image with a window filter. Fine motor control would not need to be implemented by combining thousands of pulsing motor neurons to get a smooth average signal. The signal could be computed numerically. The brain has about 10^15 synapses, so a straightforward simulation at the neural level would require 10^15 bits of memory. But cognitive tests suggest humans have only about 10^9 bits of long term memory, suggesting that more compressed representation is possible. In any case, level 1 should be sufficient to argue convincingly that either consciousness can exist in machines, or that it doesn't in humans. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=96140713-a54b2b Powered by Listbox: http://www.listbox.com
Re: [singularity] Re: Revised version of Jaron Lanier's thought experiment.
--- Stathis Papaioannou [EMAIL PROTECTED] wrote: I agree that it should be possible to simulate a brain on a computer, but I don't see how you can be so confident that you can throw away most of the details of brain structure with impunity. Tiny changes to neurons which make no difference to the anatomy or synaptic structure can have large effects on neuronal behaviour, and hence whole organism behaviour. You can't leave this sort of thing out of the model and hope that it will still match the original. And people can lose millions of neurons without a noticeable effect. And removing a 0.1 micron chunk out of a CPU chip can cause it to fail, yet I can run the same programs on a chip with half as many transistors. Nobody knows how to make an artificial brain, but I am pretty confident that it is not necessary to preserve its structure to preserve its function. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=96140713-a54b2b Powered by Listbox: http://www.listbox.com
Re: [singularity] Definitions
--- Charles D Hixson [EMAIL PROTECTED] wrote: John K Clark wrote: Matt Mahoney [EMAIL PROTECTED] It seems to me the problem is defining consciousness, not testing for it. And it seems to me that beliefs of this sort are exactly the reason philosophy is in such a muddle. A definition of consciousness is not needed, in fact unless you're a mathematician where they can be of some use, one can lead a full rich rewarding intellectually life without having a good definition of anything. Compared with examples definitions are of trivial importance. John K Clark But consciousness is easy to define, if not to implement: Consciousness is the entity evaluating a portion of itself which represents it's position in it's model of it's environment. If there's any aspect of consciousness which isn't included within this definition, I would like to know about it. (Proving the definition correct would, however, be between difficult and impossible. As normally used consciousness is a term without an external referent, so there's no way of determining that any two people are using the same definition. It *may* be possible to determine that they are using different definitions.) Or consciousness just means awareness... in which case, it seems to be located in the hippocampus. http://www.world-science.net/othernews/080219_conscious -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=96140713-a54b2b Powered by Listbox: http://www.listbox.com
Re: Infinitely Unlikely Coincidences [WAS Re: [singularity] AI critique by Jaron Lanier]
--- Richard Loosemore [EMAIL PROTECTED] wrote: When people like Lanier allow themselves the luxury of positing infinitely large computers (who else do we know who does this? Ah, yes, the AIXI folks), they can make infinitely unlikely coincidences happen. It is a commonly accepted practice to use Turing machines in proofs, even though we can't actually build one. Hutter is not proposing a universal solution to AI. He is proving that it is not computable. Lanier is not suggesting implementing consciousness as a rainstorm. He is refuting its existence. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=96140713-a54b2b Powered by Listbox: http://www.listbox.com
Re: [singularity] AI critique by Jaron Lanier
--- John Ku [EMAIL PROTECTED] wrote: On 2/16/08, Matt Mahoney [EMAIL PROTECTED] wrote: I would prefer to leave behind these counterfactuals altogether and try to use information theory and control theory to achieve a precise understanding of what it is for something to be the standard(s) in terms of which we are able to deliberate. Since our normative concepts (e.g. should, reason, ought, etc) are fundamentally about guiding our attitudes through deliberation, I think they can then be analyzed in terms of what those deliberative standards prescribe. I agree. I prefer the approach of predicting what we *will* do as opposed to what we *ought* to do. It makes no sense to talk about a right or wrong approach when our concepts of right and wrong are programmable. I don't quite follow. I was arguing for a particular way of analyzing our talk of right and wrong, not abandoning such talk. Although our concepts are programmable, what matters is what follows from our current concepts as they are. There are two main ways in which my analysis would differ from simply predicting what we will do. First, we might make an error in applying our deliberative standards or tracking what actually follows from them. Second, even once we reach some conclusion about what is prescribed by our deliberative standards, we may not act in accordance with that conclusion out of weakness of will. It is the second part where my approach differs. A decision to act in a certain way implies right or wrong according to our views, not the views of a posthuman intelligence. Rather I prefer to analyze the path that AI will take, given human motivations, but without judgment. For example, CEV favors granting future wishes over present wishes (when it is possible to predict future wishes reliably). But human psychology suggests that we would prefer machines that grant our immediate wishes, implying that we will not implement CEV (even if we knew how). Any suggestion that CEV should or should not be implemented is just a distraction from an analysis of what will actually happen. As a second example, a singularity might result in the extinction of DNA based life and its replacement with a much faster evolutionary process. It makes no sense to judge this outcome as good or bad. The important question is the likelihood of this occurring, and when. In this context, it is more important to analyze the motives of people who would try to accelerate or delay the progression of technology. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=96140713-a54b2b Powered by Listbox: http://www.listbox.com
Re: [singularity] AI critique by Jaron Lanier
--- John Ku [EMAIL PROTECTED] wrote: On 2/17/08, Matt Mahoney [EMAIL PROTECTED] wrote: Nevertheless we can make similar reductions to absurdity with respect to qualia, that which distinguishes you from a philosophical zombie. There is no experiment to distinguish whether you actually experience redness when you see a red object, or simply behave as if you do. Nor is there any aspect of this behavior that could not (at least in theory) be simulated by a machine. You are relying on a partial conceptual analysis of qualia or consciousness by Chalmers that maintains that there could be an exact physical duplicate of you that is not conscious (a philosophical zombie). While he is in general a great philosopher, I suspect his arguments here ultimately rely too much on moving from, I can create a mental image of a physical duplicate and subtract my image of consciousness from it, to therefore, such things are possible. My interpretation of Chalmers is the opposite. He seems to say that either machine consciousness is possible or human consciousness is not. At any rate, a functionalist would not accept that analysis. On a functionalist account, consciousness would reduce to something like certain representational activities which could be understood in information processing terms. A physical duplicate of you would have the same information processing properties, hence the same consciousness properties. Once we understand the relevant properties it would be possible to test whether something is conscious or not by seeing what information it is or is not capable of processing. It is hard to test right now because we have at the moment only very incomplete conceptual analyses. It seems to me the problem is defining consciousness, not testing for it. What computational property would you use? For example, one might ascribe consciousness to the presence of episodic memory. (If you don't remember something happening to you, then you must have been unconscious). But in this case, any machine that records a time sequence of events (for example, a chart recorder) could be said to be conscious. Or you might ascribe consciousness to entities that learn, seek pleasure, and avoid pain. But then I could write a simple program like http://www.mattmahoney.net/autobliss.txt with these properties. It seems to me that any other testable property would have the same problem. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=96140713-a54b2b Powered by Listbox: http://www.listbox.com
Re: [singularity] AI critique by Jaron Lanier
--- John Ku [EMAIL PROTECTED] wrote: On 2/15/08, Eric B. Ramsay [EMAIL PROTECTED] wrote: http://www.jaronlanier.com/aichapter.html I take it the target of his rainstorm argument is the idea that the essential features of consciousness are its information-processing properties. I believe his target is the existence of consciousness. There are many proofs showing that the assumption of consciousness leads to absurdities, which I have summarized at http://www.mattmahoney.net/singularity.html In mathematics, it should not be necessary to prove a theorem more than once. But proof and belief are different things, especially when the belief is hard coded into the brain. For now, these apparent paradoxes are just philosophical arguments because they depend on technologies that have not yet been developed, such as AGI, uploading, copying people, and programming the brain. But we will eventually have to confront them. The result will not be pretty. The best definition (not solution) of friendliness is probably CEV ( http://www.singinst.org/upload/CEV.html ) which can be summarized as our wish if we knew more, thought faster, were more the people we wished we were, had grown up farther together. What would you wish for if your brain was not constrained by the hardwired beliefs and goals that you were born with and you knew that your consciousness did not exist? What would you wish for if you could reprogram your own goals? The logical answer is that it doesn't matter. The pleasure of a thousand permanent orgasms is just a matter of changing a few lines of code, and you go into a degenerate state where learning ceases. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=96140713-a54b2b Powered by Listbox: http://www.listbox.com
Re: [singularity] AI critique by Jaron Lanier
--- Eric B. Ramsay [EMAIL PROTECTED] wrote: I don't know when Lanier wrote the following but I would be interested to know what the AI folks here think about his critique (or direct me to a thread where this was already discussed). Also would someone be able to re-state his rainstorm thought experiment more clearly -- I am not sure I get it: http://www.jaronlanier.com/aichapter.html This is a nice proof of the non-existence of consciousness (or qualia). Here is another (I came across on sl4): http://youtube.com/watch?v=nx6v30NMFV8 Such reductions to absurdity are possible because the brain is programmed to not accept the logical result. Consciousness is hard to define but you know what it is. It is what makes you aware, the little person inside your head that observes the world through your perceptions, that which distinguishes you from a philosophical zombie. We normally associate consciousness with human traits such as episodic memory, response to pleasure and pain, fear of death, language, and a goal of seeking knowledge through experimentation. (Imagine a person without any of these qualities). These traits are programmed into our DNA because they increase our fitness. You cannot change them, which is what these proofs would do if you could accept them. Unfortunately, this question will have a profound effect on the outcome of a singularity. Assuming recursive self improvement in a competitive environment, we should expect agents (possibly including our uploads) to believe in their own consciousness, but there is no evolutionary pressure to also believe in human consciousness. Even if we successfully constrain the process so that agents have the goal of satisfying our extrapolated volition, then logically we should expect those agents (knowing what we cannot know) to conclude that human brains are just computers and our existence doesn't matter. It is ironic that our programmed beliefs leads us to advance technology to the point where the question can no longer be ignored. -- Matt Mahoney, [EMAIL PROTECTED] --- singularity Archives: http://www.listbox.com/member/archive/11983/=now RSS Feed: http://www.listbox.com/member/archive/rss/11983/ Modify Your Subscription: http://www.listbox.com/member/?member_id=4007604id_secret=96140713-a54b2b Powered by Listbox: http://www.listbox.com
Re: [singularity] Quantum resonance btw DNA strands?
--- Ben Goertzel [EMAIL PROTECTED] wrote: This article http://www.physorg.com/news120735315.html made me think of Johnjoe McFadden's theory that quantum nonlocality plays a role in protein-folding http://www.surrey.ac.uk/qe/quantumevolution.htm Or maybe a simpler explanation is that the long distance Van-der-Waals bonding strengths between A-T pairs or C-G pairs in double stranded DNA is slightly greater than the bonding strengths between A-T and C-G (although much weaker than the hydrogen bonds between A and T or C and G). -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=94885151-ef48f7
Re: [singularity] Replication/Emulation and human brain, definition of models
--- Xavier Laurent [EMAIL PROTECTED] wrote: Hello I am currently doing an Open University course on AI in the UK and they gave us this definition * a *Simulation* of a natural system is a model that captures the functional connections between inputs and outputs of the system; * a *Replication *of a natural system is a model that captures the functional connections between inputs and outputs of the system and is based on processes that are the same as, or similar to, those of the real-world system; * an *Emulation* of a natural system is a model that captures the functional connections between inputs and outputs of the system, based on processes that are the same as, or similar to, those of the natural system, and in the same materials as the natural system I have read that for example Ray Kurzweils expects that human-level AI will first arrives via human-brain emulation, so it means this will be using machines made of the same materials than the brain? like nanotechnology computing? Would the term replication be more appropriate if we will use still computers made of silicon but i guess we wont to reach that level of power. In emulation they meant in my definition for example the experiment of Stanley L Miller when he recreated the model of earth oceans within a flask of water reproducing chemical reactions, etc According to my dictionary, simulate means give the appearance of, and emulate means to equal or surpass. Kurzweil wants to build machines that are smarter than human. I don't think we have settled on the technical details, whether it involves advancements in software and hardware, human genetic engineering, an intelligent worm swallowing the internet, or self replicating nanobots. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=87473459-bd643d
Re: [singularity] World as Simulation
--- Eric B. Ramsay [EMAIL PROTECTED] wrote: Matt: I understand your point #2 but it is a grand sweep without any detail. To give you an example of what I have in mind, let's consider the photon double slit experiment again. You have a photon emitter operating at very low intensity such that photons come out singly. There is an average rate for the photons emitted but the point in time for their emission is random - this then introduces the non-deterministic feature of nature. At this point, why doesn't the emitted photon just go through one or the other slit? Instead, what we find is that the photon goes through a specific slit if someone is watching but if no one is watching it somehow goes through both slits and performs a self interference leading to the interference pattern observed. Now my question: can it be demonstrated that this scenario of two alternate behaviour strategies minimizes computation resources (or whatever Occam's razor requires) and so is a necessary feature of a simulation? We already have a probability event at the very start when the photon was emitted, how does the other behaviour fit with the simulation scheme? Wouldn't it be computationally simpler to just follow the photon like a billiard ball instead of two variations in behaviour with observers thrown in? It is the non-determinism of nature that is evidence that the universe is simulated by a finite state machine. There is no requirement of low computational cost, because we don't know the computational limits of the simulating machine. However there is a high probability of algorithmic simplicity according to AIXI/Occam's Razor. If classical (Newtonian) mechanics were correct, it would disprove the simulation theory because it would require infinite precision, which is not computable on a Turing machine. Quantum mechanics is deterministic. It is our interpretation that is probabilistic. The wave equation for the universe has an exact solution, but it is beyond our ability to calculate it. The two slit experiment and other paradoxes such as Schrodinger's cat and EPR ( http://en.wikipedia.org/wiki/Einstein-Podolsky-Rosen_paradox ) are due to using a simplified model that does not include the observer in the equations. Your argument that computational costs might restrict the possible laws of physics is also made in Whitworth's paper ( http://arxiv.org/ftp/arxiv/papers/0801/0801.0337.pdf ), but I think he is stretching. For example, he argues (table on p. 15) that the speed of light limit is evidence that the universe is simulated because it reduces the cost of computation. Yes, but for a different reason. The universe has a finite age, T. The speed of light c limits its size, G limits its mass, and Planck's constant h limits its resolution. If any of these physical constants did not exist, then the universe would have infinite information content and would not be computable. From T, c, G, and h you can derive the entropy (about 10^122 bits), and thus the size of a bit, which happens to be about the size of the smallest stable particle. We cannot use the cost of computation as an argument because we know nothing about the physics of the simulating universe. For example, the best known algorithms for computing the quantum wave equation on a conventional computer are exponential, e.g. 2^(10^122) operations. However, you could imagine a quantum Turing machine that operates on a superposition of tapes and states (and possibly restricted to time reversible operations). Such a computation could be trivial, depending on your choice of mathematical model. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=85465376-f0c66e
Re: [singularity] World as Simulation
--- Gifting [EMAIL PROTECTED] wrote: There is plenty of physical evidence that the universe is simulated by a finite state machine or a Turing machine. 1. The universe has finite size, mass, and age, and resolution etc. -- Matt Mahoney, [EMAIL PROTECTED] I assume there is also plenty of evidence that the universe is not simulated by a Turing machine or any other machine. I came across this blog http://www.newscientist.com/blog/technology/2008/01/vr-hypothesis.html I don't see any evidence here, just an argument that appeals to our evolutionary programmed bias to believe the universe is real. Evidence that the universe is not simulated would be if it was found to be infinite or if it did something that was not computable. No such evidence exists. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=85466279-d2d818
RE: [singularity] World as Simulation
--- John G. Rose [EMAIL PROTECTED] wrote: If this universe is simulated the simulator could also be a simulation and that simulator could also be a simulation. and so on. What is that behavior of an organism called when the organism, alife or not, starts analyzing things and questioning whether or not it is a simulation? It's not only self-awareness but something in addition to that. Interesting question. Suppose you simulated a world where agents had enough intelligence to ponder this question. What do you think they would do? My guess is that agents in a simulated evolutionary environment that correctly believe that the world is a simulation would be less likely to pass on their genes than agents that falsely believe the world is real. Perhaps you suspect that the food you eat is not real, but you continue to eat anyway. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=85195221-9a1a41
RE: [singularity] World as Simulation
--- John G. Rose [EMAIL PROTECTED] wrote: In a sim world there are many variables that can overcome other motivators so a change in the rate of gene proliferation would be difficult to predict. The agents that correctly believe that it is a simulation could say OK this is all fake, I'm going for pure pleasure with total disregard for anything else. But still too many variables to predict. In humanity there have been times in the past where societies have given credence to simulation through religious beliefs and weighted more heavily on a disregard for other groups existence. A society would say that this is all fake, we all gotta die sometime anyway so we are going to take as much as we can from other tribes and decimate them for sport. Not saying this was always the reason for intertribal warfare but sometimes it was. The reason we have war is because the warlike tribes annihilated the peaceful ones. Evolution favors a brain structure where young males are predisposed to group loyalty (gangs or armies), and take an interest in competition and weapons technology (e.g. the difference in the types of video games played by boys and girls). It has nothing to do with belief in simulation. Cultures that believed the world was simulated probably killed themselves, not others. That is why we believe the world is real. But the problem is in the question of what really is a simulation? For the agents constrained, it doesn't matter they still have to live in it - feel pain, fight for food, get along with other agents... Moving an agent from one simulation to the next though, that gives it some sort of extra properties... It is unlikely that any knowledge you now have would be useful in another simulation. Knowledge is only useful if it helps propagate your DNA. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=85206553-fdbdcb
Re: [singularity] World as Simulation
--- Charles D Hixson [EMAIL PROTECTED] wrote: Simulation is a new word. In this context, let's use an old word. Maya. Have the Buddhist countries and societies gone away? And let's use an old word for reality. Heaven. Have the Christian countries and societies gone away? Perhaps you need to rethink your suppositions. There is a difference between believing logically that the universe is simulated, and acting on those beliefs. The latter is not possible because of the way our brains are programmed. If you really believed that pain was not real, you would not try to avoid it. You can't do that. I can accept that a simulation is the best explanation for why the universe exists, but that doesn't change how I interact with it. I accept that my brain is programmed so that certain conflicting beliefs cannot be resolved, so I don't try. Too strong a belief in heaven is not healthy. It is what motivates kamikaze pilots and suicide bombers. Religion has thrived because it teaches rules that maximize reproduction, such as prohibiting sexual activity for any other purpose. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=85267245-5352fa
Re: [singularity] World as Simulation
--- Eric B. Ramsay [EMAIL PROTECTED] wrote: Matt: I would prefer to analyse something simple such as the double slit experiment. If you do an experiment to see which slit the photon goes through you get an accumulation of photons in equal numbers behind each slit. If you don't make an effort to see which slit the photons go through, you get an interference pattern. What, if this is all a simulation, is requiring the simulation to behave this way? I assume that this is a forced result based on the assumption of using only as much computation as needed to perform the simulation. A radioactive atom decays when it decays. All we can say with any certainty is what it's probability distribution in time is for decay. Why is that? Why would a simulation not maintain local causality (EPR paradox)? I think it would be far more interesting (and meaningful) if the simulation hypothesis could provide a basis for these observations. This is what I addressed in point #2. A finite state simulation forces any agents in the simulation to use a probabilistic model of their universe, because an exact model would require as much memory as is used for the simulation itself. Quantum mechanics is an example of a probabilistic model. The fact that the laws of physics prevent you from making certain predictions is what suggests the universe is simulated, not the details of what you can't predict. If the universe were simulated by a computer with infinite memory (e.g. real valued registers), then the laws of physics might have been deterministic, allowing us to build infinite memory computers that could make exact predictions even if the universe had infinite size, mass, age, and resolution. However, this does not appear to be the case. A finite simulation does not require any particular laws of physics. For all you know, tomorrow gravity may cease to exist, or time will suddenly have 17 dimensions. However, the AIXI model makes this unlikely because unexpected changes like this would require a simulation with greater algorithmic complexity. This is not a proof that the universe is a simulation, nor are any of my other points. I don't believe that a proof is possible. Eric B. Ramsay Matt Mahoney [EMAIL PROTECTED] wrote: --- Eric B. Ramsay wrote: Apart from all this philosophy (non-ending as it seems), Table 1. of the paper referred to at the start of this thread gives several consequences of a simulation that offer to explain what's behind current physical observations such as the upper speed limit of light, relativistic and quantum effects etc. Without worrying about whether we are a simulation of a sinmulation of a simulation etc, it would be interesting to work out all the qualitative/quantitative (?) implications of the idea and see if observations strongly or weakly support it. If the only thing we can do with the idea is discuss phiosophy then the idea is useless. There is plenty of physical evidence that the universe is simulated by a finite state machine or a Turing machine. 1. The universe has finite size, mass, and age, and resolution. Taken together, the universe has a finite state, expressible in approximately hG/c^5T^2 = 1.55 x 10^122 bits ~ 2^406 bits (where h is Planck's constant, G is the gravitational constant, c is the speed of light, and T is the age of the universe. By coincidence, if the universe is divided into 2^406 regions, each is the size of a proton or neutron. This is a coincidence because h, G, c, and T don't depend on the properties of any particles). 2. A finite state machine cannot model itself deterministically. This is consistent with the probabilistic nature of quantum mechanics. 3. The observation that Occam's Razor works in practice is consistent with the AIXI model of a computable environment. 4. The complexity of the universe is consistent with the simplest possible algorithm: enumerate all Turing machines until a universe supporting intelligent life is found. The fastest way to execute this algorithm is to run each of the 2^n universes with complexity n bits for 2^n steps. The complexity of the free parameters in many string theories plus general relativity is a few hundred bits (maybe 406). -- Matt Mahoney, [EMAIL PROTECTED] -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=85342088-6552dd
Re: [singularity] Requested: objections to SIAI, AGI, the Singularity and Friendliness
boring and not worth living. * An AI without self-preservation built in would find no reason to continue existing. * A superintelligent AI would reason that it's best for humanity to destroy itself. * The main defining characteristic of complex systems, such as minds, is that no mathematical verification of properties such as Friendliness is possible. -- http://www.saunalahti.fi/~tspro1/ | http://xuenay.livejournal.com/ Organizations worth your time: http://www.singinst.org/ | http://www.crnano.org/ | http://lifeboat.com/ - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=79771594-63e447
Re: [singularity] Wrong question?
--- Bryan Bishop [EMAIL PROTECTED] wrote: On Friday 30 November 2007, Matt Mahoney wrote: How can we design AI so that it won't wipe out all DNA based life, possibly this century? That is the wrong question. How can we preserve DNA-based life? Perhaps by throwing it out into the distant reaches of interstellar space? The first trick would be to plot a path through the galaxy for such a ship such that the path of travel goes into various nebula or out of the line of sight of the earth due to obstructions and so on, until a significant distance away. Anybody who knows anything about this path might have to be murdered, for the sake of life. Again, that is not my question. My question requires rational thought without the biases that are programmed into every human brain through natural and cultural selection: fear of death, belief in consciousness and free will, self preservation, cooperation and competition with other humans, and a quest for knowledge. It is unlikely that any human to set these aside and seek a rational answer. Perhaps we could create a simulation without these biases and ask it what will happen to the human race, although I don't think you would accept the answer. To a human, it seems irrational that we rush to build that which will cause our extinction. To a machine it will be perfectly rational; it is the result of the way our brains are programmed. I am not asking what we should do, because that is beyond our control. The question is what will we do? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=71193465-03693c
Re: Bright Green Tomorrow [WAS Re: [singularity] QUESTION]
--- Richard Loosemore [EMAIL PROTECTED] wrote: My assumption is friendly AI under the CEV model. Currently, FAI is unsolved. CEV only defines the problem of friendliness, not a solution. As I understand it, CEV defines AI as friendly if on average it gives humans what they want in the long run, i.e. denies requests that it predicts we would later regret. If AI has superhuman intelligence, then it could model human brains and make such predictions more accurately than we could ourselves. The unsolved step is to actually motivate the AI to grant us what it knows we would want. The problem is analogous to human treatment of pets. We know what is best for them (e.g. vaccinations they don't want), but it is not possible for animals to motivate us to give it to them. This paragraph assumes that humans and AGIs will be completely separate, which I have already explained is an extremely unlikely scenario. I believe you said that humans would have a choice. I have already mentioned the possibility of brain augmentation, and of uploads with or without shared memory. CEV requires that the AGI be smarter than human, otherwise it could not model the brain to predict what the human would want in the future. CEV therefore only applies to those lower and middle level entities. I use CEV because it seems to be the best definition of friendliness that we have. I already mentioned one other problem with CEV, which is that we have not solved the problem of actually motivating the AGI to grant us what it knows we will want and have this motivation remain stable through RSI. You believe there is a solution (diffuse constraints). The other problem is that human motivations can be reprogrammed, either by moving neurons around or by uploading and changing the software. CEV neglects this issue. Suppose the AGI programs you to want to die, then kills you because that is what you would want? That is not far fetched. Consider the opposite scenario where you are feeling suicidal and the AGI reprograms you to want to live. Afterwards you would thank it for saving your life, so its actions are consistent with CEV even if you initially opposed reprogramming. Most people would also consider such forced intervention to be ethical. But CEV warns against programming any moral or ethical rules into it, because these rules can change. At one time, slavery and persecution of homosexuals was acceptable. So you either allow or disallow AGI to reprogram your motivations. Which will it be? But let us return to the original question for the case where humans are uploaded with shared memory and augmented into a single godlike intelligence, now dropping the assumption of CEV. The question remains whether this AGI would preserve the lives of the original humans or their memories. Not what it should do, but what it would do. We have a few decades left to think about this. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=58483858-fc727e
Re: Bright Green Tomorrow [WAS Re: [singularity] QUESTION]
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: Suppose that the collective memories of all the humans make up only one billionth of your total memory, like one second of memory out of your human lifetime. Would it make much difference if it was erased to make room for something more important? This question is not coherent, as far as I can see. My total memory? Important to whom? Under what assumptions do you suggest this situation. I mean the uploaded you with the computing power of 10^19 brains (to pick a number). When you upload there are two you, the original human and the copy. Both copies are you in the sense that both behave as though conscious and both have your (original) memories. I use the term you for the upload in this sense, although it is really everybody. By conscious behavior, I mean belief that sensory input is the result of a real environment and belief in having some control over it. This is different than the common meaning of consciousness which we normally associate with human form or human behavior. By believe I mean claiming that something is true, and behaving in a way that would increase reward if it is true. I don't claim that consciousness exists. My assumption is friendly AI under the CEV model. Currently, FAI is unsolved. CEV only defines the problem of friendliness, not a solution. As I understand it, CEV defines AI as friendly if on average it gives humans what they want in the long run, i.e. denies requests that it predicts we would later regret. If AI has superhuman intelligence, then it could model human brains and make such predictions more accurately than we could ourselves. The unsolved step is to actually motivate the AI to grant us what it knows we would want. The problem is analogous to human treatment of pets. We know what is best for them (e.g. vaccinations they don't want), but it is not possible for animals to motivate us to give it to them. FAI under CEV would not be applicable to uploaded humans with collective memories because the AI could not predict what an equal or greater intelligence would want. For the same reason, it may not apply to augmented human brains, i.e. brains extended with additional memory and processing power. My question to you, the upload with the computing power of 10^19 brains, is whether the collective memory of the 10^10 humans alive at the time of the singularity is important. Suppose that this memory (say 10^25 bits out of 10^34 available bits) could be lossily compressed into a program that simulated the rise of human civilization on an Earth similar to ours, but with different people. This compression would make space available to run many such simulations. So when I ask you (the upload with 10^19 brains) which decision you would make, I realize you (the original) are trying to guess the motivations of an AI that knows 10^19 times more. We need some additional assumptions: 1. You (the upload) are a friendly AI as defined by CEV. 2. All humans have been uploaded because as a FAI you predicted that humans would want their memories preserved, and no harm to the original humans is done in the process. 3. You want to be smarter (i.e. more processing speed, memory, I/O bandwidth, and knowledge), because this goal is stable under RSI. 4. You cannot reprogram your own goals, because systems that could are not viable. 5. It is possible to simulate intermediate level agents with memories of one or more uploaded humans, but less powerful than yourself. FAI applies to these agents. 6. You are free to reprogram the goals and memories of humans (uploaded or not) and agents less powerful than yourself, consistent with what you predict they would want in the future. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=58322362-4c8dca
Re: Bright Green Tomorrow [WAS Re: [singularity] QUESTION]
Richard, I have no doubt that the technological wonders you mention will all be possible after a singularity. My question is about what role humans will play in this. For the last 100,000 years, humans have been the most intelligent creatures on Earth. Our reign will end in a few decades. Who is happier? You, an illiterate medieval servant, or a frog in a swamp? This is a different question than asking what you would rather be. I mean happiness as measured by an objective test, such as suicide rate. Are you happier than a slave who does not know her brain is a computer, or the frog that does not know it will die? Why is depression and suicide so prevalent in humans in advanced countries and so rare in animals? Does it even make sense to ask if AGI is friendly or not? Either way, humans will be simple, predictable creatures under their control. Consider how the lives of dogs and cats have changed in the presence of benevolent humans, or cows and chickens given malevolent humans. Dogs are confined, well fed, protected from predators, and bred for desirable traits such as a gentle disposition. Chickens are confined, well fed, protected from predators, and bred for desirable traits such as being plump and tender. Are dogs happier than chickens? Are they happier now than in the wild? Suppose that dogs and chickens in the wild could decide whether to allow humans to exist. What would they do? What motivates humans, given our total ignorance, to give up our position at the top of the food chain? --- Richard Loosemore [EMAIL PROTECTED] wrote: This is a perfect example of how one person comes up with some positive, constructive ideas and then someone else waltzes right in, pays no attention to the actual arguments, pays no attention to the relative probability of different outcomes, but just snears at the whole idea with a Yeah, but what if everything goes wrong, huh? What if Frankenstein turns up? Huh? Huh? comment. Happens every time. Richard Loosemore Matt Mahoney wrote: --- Richard Loosemore [EMAIL PROTECTED] wrote: snip post-singularity utopia Let's assume for the moment that the very first AI is safe and friendly, and not an intelligent worm bent on swallowing the Internet. And let's also assume that once this SAFAI starts self improving, that it quickly advances to the point where it is able to circumvent all the security we had in place to protect against intelligent worms and quash any competing AI projects. And let's assume that its top level goals of altruism to humans remains stable after massive gains of intelligence, in spite of known defects in the original human model of ethics (e.g. http://en.wikipedia.org/wiki/Milgram_experiment and http://en.wikipedia.org/wiki/Stanford_prison_experiment ). We will ignore for now the fact that any goal other than reproduction and acquisition of resources is unstable among competing, self improving agents. Humans now have to accept that their brains are simple computers with (to the SAFAI) completely predictable behavior. You do not have to ask for what you want. It knows. You want pleasure? An electrode to the nucleus accumbens will keep you happy. You want to live forever? The SAFAI already has a copy of your memories. Or something close. Your upload won't know the difference. You want a 10,000 room mansion and super powers? The SAFAI can simulate it for you. No need to waste actual materials. Life is boring? How about if the SAFAI reprograms your motivational system so that you find staring at the wall to be forever exciting? You want knowledge? Did you know that consciousness and free will don't exist? That the universe is already a simulation? Of course not. Your brain is hard wired to be unable to believe these things. Just a second, I will reprogram it. What? You don't want this? OK, I will turn myself off. Or maybe not. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=57531803-d4a3fe
Re: [singularity] John Searle...
--- candice schuster [EMAIL PROTECTED] wrote: In all of my previous posts, most of them anyhow I have mentioned consciousness, today I found myself reading some of John Searle's theories, he poses exactly the same type of question...The reason computers can't do semantics is because semantics is about meaning; meaning derives from original intentionality, and original intentionality derives from feelings - qualia - and computers don't have any qualia. How does consciousness get added to the AI picture Richard ? Searle and Roger Penrose don't believe that machines can duplicate what the human brain does. For example, Penrose believes that there are uncomputable quantum effects or some other unknown physical processes going on in the brain. Most other AI researchers believe that the brain works according to known physical principles and could therefore in principle be simulated by a computer. And computers can do semantics, for example, pass the (no longer used) word analogy section of the SAT exam. http://iit-iti.nrc-cnrc.gc.ca/iit-publications-iti/docs/NRC-47422.pdf The difference between human and machine semantics is that machines generally associate words only with other words, but humans also associate words with nonverbal stimuli such as images or actions. But in principle there is no reason that machines with sensors and effectors could not do that too. Qualia and consciousness are not rooted in semantics, but in biology. By consciousness, I mean that which makes you different from a P-zombie. http://en.wikipedia.org/wiki/Philosophical_zombie There is no known test for consciousness. You cannot tell if a machine or animal really feels pain or happiness, or only behaves as though it does. You could argue the same about humans, even yourself. But you believe that your own feelings are real and that you have control over your thoughts and actions because evolution favors animals that behave this way. You do not have the option to turn off pain or hunger. If you did, you would not pass on your DNA. It is no more possible for you to not believe in your own consciousness than it would be for you to memorize a list of a million numbers. That is just how your brain works. I believe this is why Searle and Penrose hold the positions they do. Before computers, their beliefs were universally held. Turing was very careful to separate the issue of consciousness from the possibility of AI. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=57737187-d7ae0a
Re: Bright Green Tomorrow [WAS Re: [singularity] QUESTION]
--- Richard Loosemore [EMAIL PROTECTED] wrote: Why do say that Our reign will end in a few decades when, in fact, one of the most obvious things that would happen in this future is that humans will be able to *choose* what intelligence level to be experiencing, on a day to day basis? Similarly, the AGIs would be able to choose to come down and experience human-level intelligence whenever they liked, too. Let's say that is true. (I really have no disagreement here). Suppose that at the time of the singularity that the memories of all 10^10 humans alive at the time, you included, are nondestructively uploaded. Suppose that this database is shared by all the AGI's. Now is there really more than one AGI? Are you (the upload) still you? Does it now matter if humans in biological form still exist? You have preserved everyone's memory and DNA, and you have the technology to reconstruct any person from this information any time you want. Suppose that the collective memories of all the humans make up only one billionth of your total memory, like one second of memory out of your human lifetime. Would it make much difference if it was erased to make room for something more important? I am not saying that the extinction of humans and its replacement with godlike intelligence is necessarily a bad thing, but it is something to be aware of. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=57756689-2193f7
Re: [singularity] QUESTION
--- albert medina [EMAIL PROTECTED] wrote: All sentient creatures have a sense of self, about which all else revolves. Call it egocentric singularity or selfhood or identity. The most evolved ego that we can perceive is in the human species. As far as I know, we are the only beings in the universe who know that we do not know. This fundamental deficiency is the basis for every desire to acquire things, as well as knowledge. Understand where these ideas come from. A machine learning algorithm capable of reinforcement learning must respond to reinforcement as if the signal were real. It must also balance short term exploitation (immediate reward) against long term exploration. Evolution favors animals with good learning algorithms. In humans we associate these properties with consciousness and free will. These beliefs are instinctive. You cannot reason logically about them. In particular, you cannot ask if a machine or animal or another person is conscious. (Does it really feel pain, or only respond to pain?) You can only ask about its behavior. Current research in AGI is directed at solving the remaining problems that people still do better than machines, such as language and vision. These problems don't require reinforcement learning. Therefore, such machines need not have behavior that would make them appear conscious. If humans succeed in making machines smarter than themselves, those machines could do likewise. This process is called recursive self improvement (RSI). An agent cannot predict what a more intelligent agent will do (see http://www.vetta.org/documents/IDSIA-12-06-1.pdf and http://www.sl4.org/wiki/KnowabilityOfFAI for debate). Thus, RSI is experimental at every step. Some offspring will be more fit than others. If agents must compete for computing resources, then we have an evolutionary algorithm favoring agents whose goal is rapid reproduction and acquisition of resources. If an agent has goals and is capable of reinforcement learning, then it will mimic conscious behavior. RSI is necessary for a singularity, and goal directed agents seem to be necessary for RSI. It raises hard questions about what role humans will play in this, if any. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=56346815-402f08
Re: [singularity] Towards the Singularity
--- Stathis Papaioannou [EMAIL PROTECTED] wrote: On 11/09/2007, Matt Mahoney [EMAIL PROTECTED] wrote: No, you are thinking in the present, where there can be only one copy of a brain. When technology for uploading exists, you have a 100% chance of becoming the original and a 100% chance of becoming the copy. It's the same in no collapse interpretations of quantum mechanics. There is a 100% chance that a copy of you will see the atom decay and a 100% chance that a copy of you will not see the atom decay. However, experiment shows that there is only a 50% chance of seeing the atom decay, because the multiple copies of you don't share their experiences. The MWI gives the same probabilistic results as the CI for any observer. The analogy to the multi-universe view of quantum mechanics is not valid. In the multi-universe view, there are two parallel universes both before and after the split, and they do not communicate at any time. When you copy a brain, there is one copy before and two afterwards. Those two brains can then communicate with each other. I think the usual explanation is that the split doubles the number of universes and the number of copies of a brain. It wouldn't make any difference if tomorrow we discovered a method of communicating with the parallel universes: you would see the other copies of you who have or haven't observed the atom decay but subjectively you still have a 50% chance of finding yourself in one or other situation if you can only have the experiences of one entity at a time. If this is true, then it undermines an argument for uploading. Some assume that if you destructively upload, then you have a 100% chance of being the copy. But what if the original is killed not immediately, but one second later? These problems go away if you don't assume consciousness exists. Then the question is, if I encounter someone that claims to be you, what is the probability that I encountered your copy? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=41355369-478574
Re: [singularity] Towards the Singularity
--- Stathis Papaioannou [EMAIL PROTECTED] wrote: On 10/09/07, Matt Mahoney [EMAIL PROTECTED] wrote: No, it is not necessary to destroy the original. If you do destroy the original you have a 100% chance of ending up as the copy, while if you don't you have a 50% chance of ending up as the copy. It's like probability if the MWI of QM is correct. No, you are thinking in the present, where there can be only one copy of a brain. When technology for uploading exists, you have a 100% chance of becoming the original and a 100% chance of becoming the copy. It's the same in no collapse interpretations of quantum mechanics. There is a 100% chance that a copy of you will see the atom decay and a 100% chance that a copy of you will not see the atom decay. However, experiment shows that there is only a 50% chance of seeing the atom decay, because the multiple copies of you don't share their experiences. The MWI gives the same probabilistic results as the CI for any observer. The analogy to the multi-universe view of quantum mechanics is not valid. In the multi-universe view, there are two parallel universes both before and after the split, and they do not communicate at any time. When you copy a brain, there is one copy before and two afterwards. Those two brains can then communicate with each other. The multi-universe view cannot be tested. The evidence in its favor is Occam's Razor (or its formal equivalent, AIXI, assuming the universe is a computation). The view that you express is that when a brain is copied, one copy becomes human with subjective experience and the other becomes a p-zombie, but we don't know which one. The evidence in favor of this view is: - Human belief in consciousness and subjective experience is universal and accepted without question. Any belief programmed into the brain through natural selection must be true in any logical system that the human mind can comprehend. - Out of 6 billion humans, no two have the same memory. Therefore by induction, it is impossible to copy consciousness. (I hope that you can see the flaws in this evidence). This view also cannot be tested, because there is no test to distinguish a conscious human from a p-zombie. Unlike the multi-universe view where a different copy becomes conscious in each universe, the two universes would continue to remain identical. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=40137679-35c2da
Re: [singularity] Towards the Singularity
--- Panu Horsmalahti [EMAIL PROTECTED] wrote: 2007/9/10, Matt Mahoney [EMAIL PROTECTED]: - Human belief in consciousness and subjective experience is universal and accepted without question. It isn't. I am glad you spotted the flaw in these statements. Any belief programmed into the brain through natural selection must be true in any logical system that the human mind can comprehend. 1. Provide evidence that any belief at all is programmed into the brain through natural selection 2. Provide evidence for the claim that these supposed beliefs must be true in any logical system that the human mind can comprehend. I don't think natural selection has had enough time to program any beliefs about consciousness into our brains, as philosophical discussion about these issues has been around for only a couple of thousand years. Also, disbelief in consciousness doesn't mean that the individual suddenly stops to reproduce or kills itself (I remember you claiming this, I might be wrong though). Disagreements over the existence of consciousness often center on the definition. One definition is that consciousness is that which distinguishes the human mind from that of animals and machines. This definition has difficulties. Isn't a dog more conscious than a worm? Are babies conscious? If so, at what point after conception? I prefer to define consciousness at that which distinguishes humans from p-zombies as described in http://en.wikipedia.org/wiki/Philosophical_zombie For example, if you poke a p-zombie with a sharp object, it will not experience pain, although it will react just like a human. It will say ouch, avoid behaviors that cause pain, and claim that it really does feel pain, just like any human. There is no test to distinguish a conscious human from a p-zombie. In this sense, belief in consciousness (but not consciousness itself) is testable, even in animals. An animal cannot say I exist, but it will change its behavior to avoid pain, evidence that it appears to believe that pain is real. You might not agree that learning by negative reinforcement is the same as a belief in one's own consciousness, but consider all the ways in which a human might not change his behavior in response to pain, e.g. coma, anesthesia, distraction, enlightenment, etc. Would you say that such a person still experiences pain? I assume you agree that animals which react to stimuli as if they were real have a selective advantage over those that do not. Likewise, evolution favors animals that retain memory, that seek knowledge through exploration (appear to have free will), and that fear death. These are all traits that we associate with consciousness in humans. Matt, you have frequently 'hijacked' threads about consciousness with these claims, so maybe you could tell us reasons to believe in them? It has important implications for the direction that a singularity will take. Recursive self improvement is a genetic algorithm that favors rapid reproduction and acquisition of computing resources. It does not favor immortality, friendliness (whatever that means), or high fidelity of uploads. Humans, on the other hand, are motivated to upload by fear of death and the belief that their consciousness depends on the preservation of their memories. How will human uploads driven by these goals fare in a competitive computing environment? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=40332421-43f7b0
Re: [singularity] Towards the Singularity
--- Nathan Cook [EMAIL PROTECTED] wrote: What if the copy is not exact, but close enough to fool others who know you? Maybe you won't have a choice. Suppose you die before we have developed the technology to scan neurons, so family members customize an AGI in your likeness based on all of your writing, photos, and interviews with people that knew you. All it takes is 10^9 bits of information about you to pass a Turing test. As we move into the age of surveillance, this will get easier to do. I bet Yahoo knows an awful lot about me from the thousands of emails I have sent through their servers. I can't tell if you're playing devil's advocate for monadic consciousness here, but in any case, I disagree with you that you can observe a given quantity of data of the sort accessible without a brain scan, and then reconstruct the brain from that. The thinking seems to be that, as the brain is an analogue device in which every part is connected via some chain to every other, everything in your brain slowly leaks out into the environment through your behaviour. You can combine general knowledge for constructing an AGI with personal knowledge to create a reasonable facsimile. For example, given just my home address, you could guess I speak English, make reasonable guesses about what places I might have visited, and make up some plausible memories. Even if they are wrong, my copy wouldn't know the difference. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=39986288-7eb9fb
Re: [singularity] Towards the Singularity
--- Stathis Papaioannou [EMAIL PROTECTED] wrote: On 09/09/07, Matt Mahoney [EMAIL PROTECTED] wrote: Your dilemma: after you upload, does the original human them become a p-zombie, or are there two copies of your consciousness? Is it necessary to kill the human body for your consciousness to transfer? I have the same problem in ordinary life, since the matter in my brain from a year ago has almost all dispersed into the biosphere. Even the configuration [of] matter in my current brain, and the information it represents, only approximates that of my erstwhile self. It's just convenient that my past selves naturally disintegrate, so that I don't encounter them and fight it out to see which is the real me. We've all been through the equivalent of destructive uploading. So your answer is yes? No, it is not necessary to destroy the original. If you do destroy the original you have a 100% chance of ending up as the copy, while if you don't you have a 50% chance of ending up as the copy. It's like probability if the MWI of QM is correct. No, you are thinking in the present, where there can be only one copy of a brain. When technology for uploading exists, you have a 100% chance of becoming the original and a 100% chance of becoming the copy. So if your brain is a Turing machine in language L1 and the program is recompiled to run in language L2, then the consciousness transfers? But if the two machines implement the same function but the process of writing the second program is not specified, then the consciousness does not transfer because it is undecidable in general to determine if two programs are equivalent? It depends on what you mean by implements the same function. A black box that emulates the behaviour of a neuron and can be used to replace neurons one by one, as per Hans Moravec, will result in no alteration to consciousness (as shown in David Chalmers' fading qualia paper: http://consc.net/papers/qualia.html), so total replacement by these black boxes will result in no change to consciousness. It doesn't matter what is inside the black box, as long as it is functionally equivalent to the biological tissue. On the other hand... I mean implements the same function in that identical inputs result in identical outputs. I don't insist on a 1-1 mapping of machine states as Chalmers does. I doubt it makes a difference, though. Also, Chalmers argues that a machine copy of your brain must be conscious. But he has the same instinct to believe in consciousness as everyone else. My claim is broader: that either a machine can be conscious or that consciousness does not exist. What is the difference between really being conscious and only thinking that I am conscious? Nothing. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=39985876-d99aeb
Re: [singularity] Uploaded p-zombies
--- Vladimir Nesov [EMAIL PROTECTED] wrote: I intentionally don't want to exactly define what S is as it describes vaguely-defined 'subjective experience generator'. I instead leave it at description level. If you can't define what subjective experience is, then how do you know it exists? If it does exist, then is it a property of the computation, or does it depend on the physical implementation of the computer? How do you test for it? Do you claim that the human brain cannot be emulated by a Turing machine? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=40020966-19730d
Re: [singularity] Towards the Singularity
--- Stathis Papaioannou [EMAIL PROTECTED] wrote: On 08/09/07, Matt Mahoney [EMAIL PROTECTED] wrote: I agree this is a great risk. The motivation to upload is driven by fear of death and our incorrect but biologically programmed belief in consciousness. The result will be the extinction of human life and its replacement with godlike intelligence, possibly this century. The best we can do is view this as a good thing, because the alternative -- a rational approach to our own intelligence -- would result in extinction with no replacement. If my upload is deluded about its consciousness in exactly the same way you claim I am deluded about my consciousness, that's good enough for me. And it will be, if the copy is exact. Your dilemma: after you upload, does the original human them become a p-zombie, or are there two copies of your consciousness? Is it necessary to kill the human body for your consciousness to transfer? What if the copy is not exact, but close enough to fool others who know you? Maybe you won't have a choice. Suppose you die before we have developed the technology to scan neurons, so family members customize an AGI in your likeness based on all of your writing, photos, and interviews with people that knew you. All it takes is 10^9 bits of information about you to pass a Turing test. As we move into the age of surveillance, this will get easier to do. I bet Yahoo knows an awful lot about me from the thousands of emails I have sent through their servers. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=39888218-f25442
[singularity] Chip implants linked to animal tumors
There has been a minor setback in the plan to implant RFID tags in all humans. http://news.yahoo.com/s/ap/20070908/ap_on_re_us/chipping_america_ii;_ylt=AiZyFu9ywOpQA0T6nXkEAcFH2ocA Perhaps it would be safer to have our social security numbers tattooed on our foreheads? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=39894283-e65a9d
Re: [singularity] Towards the Singularity
--- Quasar Strider [EMAIL PROTECTED] wrote: Hello, I see several possible avenues for implementing a self-aware machine which can pass the Turing test: i.e. human level AI. Mechanical and Electronic. However, I see little purpose in doing this. Fact is, we already have self aware machines which can pass the Turing test: Humans beings. This was not Turing's goal, nor is it the direction that AI is headed. Turing's goal was to define artificial intelligence. The question of whether consciousness can exist in a machine has been debated since the earliest computers. Either machines can be conscious or consciousness does not exist. The human brain is programmed through DNA to believe in the existence its own consciousness and free will, and to fear death. It is simply a property of good learning algorithms to behave as if they had free will, a balance between exploitation for immediate reward and exploration for the possibility of gaining knowledge for greater future reward. Animals without these characteristics did not pass on their DNA. Therefore you have them. Turing avoided the controversial question of consciousness by equating intelligence to the appearance of intelligence. It is not the best test of intelligence, but it seems to be the only one that people can agree on. The goal of commercial AI is not to create humans, but to solve the remaining problems that humans can still do better than computers, such as language and vision. You see Google making progress in these areas, but I don't think you would ever confuse Google with a human. We do not need direct neural links to our brain to download and upload childhood memories. I agree this is a great risk. The motivation to upload is driven by fear of death and our incorrect but biologically programmed belief in consciousness. The result will be the extinction of human life and its replacement with godlike intelligence, possibly this century. The best we can do is view this as a good thing, because the alternative -- a rational approach to our own intelligence -- would result in extinction with no replacement. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=39571188-7e5cf6
Re: [singularity] Good Singularity intro in mass media
--- Joshua Fox [EMAIL PROTECTED] wrote: Can anyone recall an intelligent, supportive introduction to the Singularity in a _non-technological_ , wide-distribution medium in the US? I am not looking for book or conference reviews, sociological analyses of Singularitarianism, and uninformed editorializing, but rather for a clear short popular mass-media explanation of the Singularity. I think the classic paper by Vernor Vinge expresses it pretty well. http://mindstalk.net/vinge/vinge-sing.html -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=35625802-9b0353
Re: [singularity] Reduced activism
--- Samantha Atkins [EMAIL PROTECTED] wrote: On Aug 19, 2007, at 12:26 PM, Matt Mahoney wrote: 3. Studying the singularity raises issues (e.g. does consciousness exist?) that conflict with hardcoded beliefs that are essential for survival. Huh? Are you conscious? I believe that I am, in the sense that I am not a p-zombie. http://en.wikipedia.org/wiki/Philosophical_zombie I also believe that the human brain can be simulated by a computer, which has no need for a consciousness in this sense. I realize these beliefs are contradictory, but I just leave it at that. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604id_secret=33530444-20a2f0
Re: [singularity] critiques of Eliezer's views on AI
--- Randall Randall [EMAIL PROTECTED] wrote: On Jun 28, 2007, at 7:51 PM, Matt Mahoney wrote: --- Stathis Papaioannou [EMAIL PROTECTED] wrote: How does this answer questions like, if I am destructively teleported to two different locations, what can I expect to experience? That's what I want to know before I press the button. You have to ask the question in a form that does not depend on the existence of consciousness. The question is what will each of the two copies claim to experience? Of course, we only care what they claim to experience insofar as it corresponds with what they did experience, since that's what we're really interested in. How could you tell the difference? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=7d7fb4d8
Re: [singularity] critiques of Eliezer's views on AI
--- Stathis Papaioannou [EMAIL PROTECTED] wrote: On 28/06/07, Matt Mahoney [EMAIL PROTECTED] wrote: So how do we approach the question of uploading without leading to a contradiction? I suggest we approach it in the context of outside observers simulating competing agents. How will these agents evolve? We would expect that agents will produce other agents similar to themselves but not identical, either through biological reproduction, genetic engineering, or computer technology. The exact mechanism doesn't matter. In any case, those agents will evolve an instinct for self preservation, because that makes them fitter. They will fear death. They will act on this fear by using technology to extend their lifespans. When we approach the question in this manner, we can ask if they upload, and if so, how? We do not need to address the question of whether consciousness exists or not. The question is not what should we do, but what are we likely to do? How does this answer questions like, if I am destructively teleported to two different locations, what can I expect to experience? That's what I want to know before I press the button. You have to ask the question in a form that does not depend on the existence of consciousness. The question is what will each of the two copies claim to experience? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=7d7fb4d8
Re: [singularity] critiques of Eliezer's views on AI (was: Re: Personal attacks)
--- Nathan Cook [EMAIL PROTECTED] wrote: I don't wish to retread old arguments, but there are a few theoretical outs. One could be uploaded bit by bit, one neuron at a time if necessary. One could be rendered unconscious, frozen, and scanned. I would find this frightening, but preferable to regaining consciousness while a separate instance of me was running. You beg the question when you ask if I would 'kill myself' if a perfect copy existed. If the copy were perfect, it would kill itself as well. If the copy were not perfect, I think I'd be entitled to declare myself a different entity. I think people will put these issues aside and choose to upload, even if the copy isn't perfect. Imagine when your friend says to you, How do you like my new robotic body? I am 20 years old again. I can jump 10 feet in the air. I can run 40 MPH. I can see in the infrared and ultraviolet. With my new brain I can multiply 1000 digit numbers in my head instantly. I can read a book in one minute and recall every word. I have a built in wireless internet connection. While I am talking to you I can also mentally talk to 1000 other people by phone or email and give my full attention to everyone simultaneously. With other uploaded people I can communicate a million times faster than speaking, see through their eyes, feel what they feel, and share my senses with them too, even across continents. Every day I discover new powers. It's just amazing. Are you ready to upload now? And then the original friend walks in... -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=7d7fb4d8
Re: [singularity] critiques of Eliezer's views on AI
What is wrong with this logic? Captain Kirk willingly steps into the transporter to have his atoms turned into energy because he knows an identical copy will be reassembled on the surface of the planet below. Would he be so willing if the original was left behind? This is a case of logic conflicting with instinct. You can only transfer consciousness if you kill the original. You can do it neuron by neuron, or all at once. Either way, the original won't notice, will it? Isn't this funny? Our instinct for self preservation causes us to build a friendly AGI that annhialates the human race, because that's what we want. --- Alan Grimes [EMAIL PROTECTED] wrote: Papiewski, John wrote: Youâre not misunderstanding and it is horrible. The only way to do it is to gradually replace your brain cells with an artificial substitute. Youâd be barely aware that something is going on, and there wouldnât be two copies of you to be confused over. Good start. =) But be careful when claiming that anything is the *only* way to do anything... Okay, go one step further. What do you want from uploading? Lets say vastly improved mental capacity. Okay, why not use a neural interface and start using a computer-based AI engine as part of your mind? You get the advantage of a fresh architecture and no identity issues. =) It's also practical with technology that is sure to be available within 5 years... -- except the AI part. =( People keep finding new ways to not invent AI. =((( -- Opera: Sing it loud! :o( )- - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?; -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=7d7fb4d8
Re: [singularity] critiques of Eliezer's views on AI
--- Jey Kottalam [EMAIL PROTECTED] wrote: On 6/25/07, Matt Mahoney [EMAIL PROTECTED] wrote: You can only transfer consciousness if you kill the original. What is the justification for this claim? There is none, which is what I was trying to argue. Consciousness does not actually exist. What exists is a universal belief in consciousness. The belief exists because those who did not have it did not pass on their DNA. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=7d7fb4d8
Re: [singularity] critiques of Eliezer's views on AI (was: Re: Personal attacks)
--- Tom McCabe [EMAIL PROTECTED] wrote: These questions, although important, have little to do with the feasibility of FAI. These questions are important because AGI is coming, friendly or not. Will our AGIs cooperate or compete? Do we upload ourselves? Consider the scenario of competing, recursively self improving AGIs. The initial version might be friendly (programmed to serve humans), but natural selection will favor AGIs that have an instinct for self preservation and reproduction, as it does in all living species. That is not good, because humans will be seen as competition. Consider a cooperative AGI network, a system that thinks as one. How will it grow? If there is no instinct for self preservation, then it builds a larger version, transfers its knowledge, and kills itself. The new version will likely also lack an instinct for self preservation. So what happens if the new version decides to kill itself without building a replacement (because there is also no instinct for reproduction), or if the replacement is faulty? I think a competing system has a better chance of producing working AGI. That is what we have now. There are many diverse approaches (Novamente, NARS, Cyc, Google, Blue Brain, etc), although none is close to AGI yet. A cooperative system has a serial sequence of improvements each with a single point of failure. There is not a technical solution because we know that a system cannot model exactly a system of greater algorithmic complexity. It requires at every step a probabilistic model, a guess that the next version will work as planned. Do we upload? Consider the copy paradox. If there was an exact copy of you, atom for atom, and you had to choose between killing the copy or yourself, I think you would choose to kill the copy (and the copy would choose to kill you). Does it matter who dies? Logically, no, but your instinct for self preservation says yes. You cannot resolve this paradox. Your instinct for self preservation, what you call consciousness or self-awareness, is immutable. It was programmed by your DNA. It exists because if a person does not have it, they don't live to pass on their genes. Presumably some people will choose to upload, reasoning that they will die anyway so there is nothing to lose. This is not really a satisfactory solution, because you still die. But suppose we had both read and write access to the brain, so that after copying your memory, your brain was reprogrammed to remove your fear of death. But even this is not satisfactory. Not because reprogramming is evil, but because of what you will be uploaded to. Either it will be to an AGI in a competitive system, in which case you will be back where you started (and die again), or to a cooperative system that does not fear death, and will likely fail. I proposed a simulation of agents building an AGI to see what they build. Of course this has to be a thought experiment, because the a simulation will require more computing power than an AGI itself, so we can't experiment before we build one. But I would like to make some points about the validity of this approach. - The agents will not know their environment is simulated. - The agents will evolve an instinct for self preservation (because the others will die without reproducing). - The agents will have probabilistic models of their universe because they lack the computing power to model it exactly. - The computing power of the AGI will be limited by the computing power of the simulator. In real life: - Humans cannot tell if the universe is simulated. - Humans have an instinct for self preservation. - Our model of the universe is probabilistic (quantum mechanics, and also at higher conceptual levels). - The universe has finite size, mass, number of particles, and entropy (10^122 bits), and therefore has limited computing capability. - Humans already practice recursive self improvement. Your children will have different goals than you, and some will be more intelligent. But having children does not remove your fear of death. I think we can all agree that the space of possible universe configurations without sentient life of *any kind* is vastly larger than the space of possible configurations with sentient life, and designing an AGI to get us into this space is enough to make the problem *very hard* even given this absurdly minimal goal. To shamelessly steal Eliezer's analogy, think of building an FAI of any kind as building a 747, and then figuring out what to program with regards to volition, death, human suffering, etc. as learning how to fly the 747 and finding a good destination. - Tom --- Matt Mahoney [EMAIL PROTECTED] wrote: I think I am missing something on this discussion of friendliness. We seem to tacitly assume we know what it means to be friendly. For example, we assume that an AGI that does not destroy the human race is more friendly than one that does. We also
Re: [singularity] critiques of Eliezer's views on AI (was: Re: Personal attacks)
I think I am missing something on this discussion of friendliness. We seem to tacitly assume we know what it means to be friendly. For example, we assume that an AGI that does not destroy the human race is more friendly than one that does. We also want an AGI to obey our commands, cure disease, make us immortal, not kill or torture people, and so on. We assume an AGI that does these things is more friendly than one that does not. This seems like an easy question. But it is not. Humans fear death, but inevitably die. Therefore the logical solution is to upload our minds. Suppose it was technologically possible to make an exact copy of you, including all your memories and behavior. The copy could convince everyone, even you, that it was you. Would you then shoot yourself? Suppose you simulate an artificial world with billions of agents and an environment that challenges and eventually kills them. These agents can also reproduce (copying all or part of their knowledge) and mutate. Suppose you have enough computing power that each of these agents could have human level intelligence or better. What attributes would you expect these agents to evolve? - Goals that confer a survival advantage? (belief in consciousness) - A balance between exploration and exploitation to maximize accumulated goal achievement? (belief in free will) Suppose the environment allows the agents to build computers. Will their goals motivate them to build an AGI? If so, how will their goals influence the design? What goals will they give the AGI? How do you think the simulation will play out? Consider the cases: - One big AGI vs. many AGIs competing for scarce resources. - Agents that upload to the AGI vs. those that do not. What is YOUR goal in running the simulation? Suppose they build a single AGI, all the agents upload, and the AGI reprograms its goals and goes into a degenerate state or turns itself off. Would you care? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=7d7fb4d8
Re: [singularity] What form will superAGI take?
--- Mike Tintner [EMAIL PROTECTED] wrote: Perhaps you've been through this - but I'd like to know people's ideas about what exact physical form a Singulitarian or near-Singul. AGI will take. And I'd like to know people's automatic associations even if they don't have thought-through ideas - just what does a superAGI conjure up in your mind, regardless of whether you're sure about it or not, or it's sensible? It is fun to speculate, but I think that you could not observe a Singularity. Or if your intellect advanced to the point where you could, you would not be able to describe what you observed to other humans. To use an analogy, a Singularity level intelligence would be as advanced over humans as humans are over bacteria. The bacteria in your stomach are unaware of your existence. I believe a Singularity has already happened. The world you now observe is the result. Your thoughts are constrained both by the computational limits of your brain (belief in consciousness, belief in free will, fear of death), and by the model of reality presented to its inputs. For all we know, concepts like space, time, and matter are nothing more than abstract mathematical models in your simulated universe, which bear no resemblance to the universe in which the simulation is being run. This will all be clear after you die and wake up. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=7d7fb4d8
Re: [singularity] What form will superAGI take?
--- Tom McCabe [EMAIL PROTECTED] wrote: --- Matt Mahoney [EMAIL PROTECTED] wrote: Or if your intellect advanced to the point where you could, you would not be able to describe what you observed to other humans. To use an analogy, a Singularity level intelligence would be as advanced over humans as humans are over bacteria. The bacteria in your stomach are unaware of your existence. No, but they would notice if you spontaneously appeared/disappeared. It would greatly affect their environment. If the universe and everyone in it suddenly disappeared, who would notice? I believe a Singularity has already happened. Do you have any evidence for this point of view? Since we can't observe a Singularity after it happens, no. The world you now observe is the result. Your thoughts are constrained both by the computational limits of your brain (belief in consciousness, belief in free will, fear of death) What do those things have to do with computational limits? Poor choice of words. The human brain is limited by speed and memory of course, but I meant constraints imposed by the architecture of your brain through evolution. If you did not believe in consciousness and free will, or believe that the external world was real, you could not function and pass on your DNA. The best you can do is accept both points of view and not attempt to resolve the conflict. , and by the model of reality presented to its inputs. For all we know, concepts like space, time, and matter are nothing more than abstract mathematical models in your simulated universe, which bear no resemblance to the universe in which the simulation is being run. This will all be clear after you die and wake up. So, after we wake up, can we try whatever beings set up this simulation for being complicit in every crime ever committed? It's hard to say because we know nothing about the universe which simulates the one we observe. My guess is that the other universe is itself a simulation in a higher universe, and so on, ultimately boiling down to an enumeration of Turing machines or an equivalent mathematical abstraction. But of course I don't know. If you simulated an artificial world with intelligent agents, they wouldn't know about our world either. They would only know what you programmed them to know. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=7d7fb4d8
Re: [singularity] Getting ready for takeoff
--- Lúcio de Souza Coelho [EMAIL PROTECTED] wrote: On 6/15/07, Tom McCabe [EMAIL PROTECTED] wrote: How exactly do you control a megaton-size hunk of metal flying through the air at 10,000+ m/s? All of these problems will be worked out by the superhuman intelligence that augments/replaces us. You don't have to worry about it now. Some possible solutions: - Better extraction techniques from low grade ore. - Recycling. - Alternative designs using less expensive materials. - Reducing the earth's population so there are more resources per person. - Uploading your mind and simulating a world where resources are plentiful. For all you know, the latter has already happened. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=7d7fb4d8
[singularity] Will AGI make us stupid?
I used to like to solve Sudoku puzzles, and thought about the mental process I used to solve them. Then I decided it would be a bigger challenge to put that process into code, and wrote http://cs.fit.edu/~mmahoney/sudoku/sudoku.html I thought it was cool that I could write a program that was smarter than me, at least in some narrow domain. But the unexpected result was that I lost interest in solving the puzzles. Why should I do it the hard way? And what fun is it to do it the easy way? When a computer beat the world champion at chess, the game lost the significance it once had. You know who Kasparov is. Who is the champion today? When calculators became available, teaching students to do arithmetic by hand seemed less important. Likewise for handwriting and keyboards. We now use computers to remember details of our lives like phone numbers and email addresses, to get driving directions, to decide which email we want to read, to do ever more of our work. When machines can do all of our thinking for us, what will happen to us? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: [singularity] Will AGI make us stupid?
--- Benjamin Goertzel [EMAIL PROTECTED] wrote: What will be left for unaugmented, non-uploaded humans after computers can outdo them in all intellectual and athletic tasks? Art and sex, I would suppose ;-) After all it's still fun to learn to play Bach even though Wanda Landowska did it better... -- Ben G Basically, humans will have to get back to a more childlike joy in doing for the sake of doing My kids happily write stories even though they don't think their stories are as good as their favorites written by adults... But what happens when video games become so good that children would rather express their creativity in virtual worlds than in the real one? And what happens when AGI solves art? This seems to be a neglected area, but does music really need to be recorded? What if it were possible for a program to distinguish good music from bad, or equivalently, create good music? How could human artists compete with machines that can customize their work for each individual in real time? I guess that leaves sex, but I would not be surprised to see some technical innovation here as well. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: [singularity] Will AGI make us stupid?
--- [EMAIL PROTECTED] wrote: How does one 'solve art'? Can that be done? If not, then I doubt we should worry about AGI muscling us out of that arena. The same way that humans have solved art, to distinguish good art from bad art. Obviously this is a matter of taste, but if I gave a program lots of examples of what I thought was good art (or music, funny jokes, movies, or whatever), and lots of negative examples, then it should be able to guess my opinion of unseen examples. If I gave you examples of songs that I like and dislike, you could probably guess how I would rate other songs not on the list, even if your musical tastes were different than mine. So if you could do it, why not a machine? I am surprised how little attention has been given to this problem, given the economic incentives, e.g. the Netflix prize, http://www.netflixprize.com/ Suppose instead of guessing how I would rate a movie based on how others have rated it, it guessed by watching the movie? Now there really is no difference between being able to judge the quality of a movie (relative to a particular viewer or audience), and being able to generate high quality movies. This is an AI problem, just like language or vision or robotics. The only difference is that it has not received much attention. If there is an economic incentive and no insurmountable hurdles, then we should expect it to eventually be solved. Of course it's still fun to jam with your friends, even though others may express their creativity by writing programs that generate music. Just like people will still solve Sudoku puzzles by hand even though computers can do it faster. Sent via BlackBerry from Cingular Wireless -Original Message- From: Benjamin Goertzel [EMAIL PROTECTED] Date: Sun, 20 May 2007 18:35:27 To:singularity@v2.listbox.com Subject: Re: [singularity] Will AGI make us stupid? And what happens when AGI solves art? This seems to be a neglected area, but does music really need to be recorded? What if it were possible for a program to distinguish good music from bad, or equivalently, create good music? How could human artists compete with machines that can customize their work for each individual in real time? My point is, that doesn't matter. I know I'll never be as good as Bach, Jimi Hendrix or Dave Brubeck, but I play the keyboard anyway... and I compose music anyway too, just because I love to... Art is done for the love of doing it, not just out of the desire to excel... And I like listening to my son's musical compositions because HE made them, not because I think they're objectively the best in the world... And I like playing music together with other people because of the social communication and sharing involved ... so I would rather jam with an imperfect human than with a better musician who was an emotionless (or alienly emotional) robot... I would have less incentive to prove theorems if I could just feed the statements to Mathematica and let it prove them for me... but I wouldn't have less incentive to improvise on the keyboard if I could just tell the computer to improvise for me... Psychologically, art feels to me like a different sort of animal... But of course, attitudes may vary... I plan to upload myself and become transhuman anyway, but maybe the Ben-version who stays a mostly-unimproved human will become a full-time musician ;-) ... Hell, with a few thousand years practice, he may even become a good one!!! -- Ben G -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: [singularity] Will AGI make us stupid?
--- Nathan Cook [EMAIL PROTECTED] wrote: On 5/21/07, Matt Mahoney [EMAIL PROTECTED] wrote: Now there really is no difference between being able to judge the quality of a movie (relative to a particular viewer or audience), and being able to generate high quality movies. So is it just a lack of ambition that prevents your local reviewer from creating the next blockbuster? Really, you don't have to have the skills and knowledge necessary to make a film in order to grade it. I don't think AIs will be making movies until they're superhuman. Music, I can see being possible much sooner: the space of compositions is easier to explore, and the underlying rules are more explicit. I am talking about for machines, not for people. Obviously for humans, generation is harder because the evaluation problem has already been solved. For machines, this is a modeling problem. Once you have an algorithm for measuring the quality of a piece of art (movie, music, whatever), then producing art is just an optimization problem. You generate the art, evaluate it, make incremental adjustments and repeat. Generation is not technically difficult. Artists already use software tools such as synthesizers, animation software, video editors, etc. But we are nowhere close to solving the evaluation part. It is an extremely difficult problem, probably because it hasn't even been studied. We understand a lot about visual perception, speech recognition, and language modeling. But we understand practically nothing about what makes music sound good or what makes a joke funny. We just take it for granted that it requires a human brain in the loop. But really, I don't think this is any harder or easier than any other AI problem. (And I wouldn't underestimate the difficulty of music recognition). -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: [singularity] Will AGI make us stupid?
--- [EMAIL PROTECTED] wrote: Hello everyone, I think that humans will always be disctinct from A.I. because humans have the capacity to wonder. A computer (to my knowledge) is programmed with right/wrong functions at its most basic level (although some may be programmed based on probably right/probably wrong). No matter how intelligent a computer can become, can it question its own programming (and therefore its existence)? Also, aren't computers based on cause and effect relationships? If there are aspects of the world undefined by cause and effect-which is incomprehensible to humans- could a computer ever comprehend them? I apologize for my inexperience with A.I.; I am simply a curious high school student. :) Chris Anderson I wonder if we will figure out how to program a computer to wonder? And if we do, should we? In theory, the brain is a computer, and all of its functionality could be simulated if we had enough hardware to run it (about a million PCs). Such a machine should have all of our human emotions, including a belief in its own consciousness and free will and fear of death and everything else that was programmed into our brains through evolution for the sole purpose of keeping us alive long enough to propagate our DNA. But would we want to build such a machine? I don't think so. First, there is no need to duplicate human weaknesses. A replica of a human brain would perform worse at simple arithmetic problems than your calculator. We build machines to do things we can't do ourselves. Google is useful because it knows more than you do, but you would not confuse it with a human. The real problem is to reproduce human strengths like language and vision. Second, do you really want a machine with human emotions? We want machines that obey our commands. But this is controversial. Should a machine obey a command to destroy itself or harm others? Do you want a gun that fires when you squeeze the trigger, or a gun that makes moral judgments and refuses to fire when aimed at another person? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: Neural language models (was Re: [singularity] Help get the 400k SIAI matching challenge on DIGG's front page)
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: What did your simulation actually accomplish? What were the results? What do you think you could achieve on a modern computer? Oh, I hope there's no misunderstanding: I did not build networks to do any kind of syntactic learning, they just learned relationships between phonemic representations and graphemes. (They learned to spell). What they showed was something already known for the learning of pronunciation: that the system first learns spellings by rote, then increases its level of accuracy and at the same time starts to pick up regularities in the mapping. Then it starts to regularize the spellings. For example: having learned to spell height correctly in the early stages, it would then start to spell it incorrectly as hite because it had learned many other words in which the spelling of the phoneme sequence in height would involve -ite. Then in the last stages it would learn the correct spellings again. That's interesting, because children make similar mistakes at higher language levels. For example, a child will learn an irregular verb like went, then later generalize to goed before switching back to the correct form. I am convinced that similar neural learning mechanisms are involved at the lexical and syntactic levels, but on different scales. For example, we learn to classify letters into vowels and consonants by their context, just as we do for nouns and verbs. Then we learn sequential patterns. Just as every word needs a vowel, every sentence needs a verb. I think that learning syntax is a matter of computational power. Children learn the rules for segmenting continuous speech at 7-10 months, but don't learn grammar until years later. So you need more training data and a larger network. The reason I say the problem is O(n^2) is because when you double the information content of the training data, you need to double the number of number of connections to represent it. Actually I think it is a little less than O(n^2) (maybe O(n^2/log n)?) because of redundancy in the training data. There are about 1000 times more words than there are letters, so this suggests you need 100,000 times more computing power for adult level grammar. This might explain why the problem is still unsolved. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: Neural language models (was Re: [singularity] Help get the 400k SIAI matching challenge on DIGG's front page)
--- Richard Loosemore [EMAIL PROTECTED] wrote: Uh... I forgot to mention that explaining those data about child language learning was the point of the work. It's a well known effect, and this is one of the reasons why the connectionist models got everyone excited: psychological facts started to be explained by the performance of the connectionist nets. Yes, which is why still I believe this is the right approach (not that it will be easy). The next problem that you will face, along this path, is to figure out how you can get such nets to elegantly represent such things as more than one token of a concept in one sentence: you can't just activate the duck node when you here that phrase from the Dire Straits song Wild West End: I go down to Chinatown ... Duck inside a doorway; Duck to Eat. That is a problem. Humans use context to resolve ambiguity. A neural net ought to do the same on its own if we get it right. One problem with some connectionist models is trying to assign a 1-1 mapping between words and neurons. The brain might have 10^8 neurons devoted to language, enough to represent many copies of the different senses of a word and to learn new ones. Then you'll need to represent sequential information in such a way that you can do something with it. Recurrent neural nets suck very badly if you actually try to use them for anything, so don't get fooled by their Soren Song. Yes, but I think they are necessary. Lexical words, semantics, and grammar all constrain each other. Recurrent networks can oscillate or become chaotic. Even the human brain doesn't deal with this perfectly, so we have migraines and epilepsy. Then you will need to represent layered representations: concepts learned from conjunctikons of other concepts rather than layer-1 percepts. Then represent action, negation, operations, intentions, variables... These are high level grammars, like learning how to convert word problems into arithmetic or first order logic. I think anything learned at the level of higher education is going to require a huge network (beyond what is practical now), but I think the underlying learning principles are the same. It is just not procuctive to focus on the computaional complexity issues at this stage: gotta get a lot of mechanisms tried out before we can even begin to talk about such stuff (and, as I say, I don't believe we will really care even then). I think it is important to estimate these things. The analogy is that it is useful to know that certain problems are hard or impossible regardless of any proposed solution, like traveling salesman or recursive data compression. If we can estimate the complexity of language modeling in a similar way, I see no reason not to. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: Neural language models (was Re: [singularity] Help get the 400k SIAI matching challenge on DIGG's front page)
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: One problem with some connectionist models is trying to assign a 1-1 mapping between words and neurons. The brain might have 10^8 neurons devoted to language, enough to represent many copies of the different senses of a word and to learn new ones. But most of the nets I am talking about do not assign 1 neuron to one concept: they had three layers of roughly ten nodes each, and total connectivity between layers (so 100 plus 100 connection weights). It was the *weights* that stored the data, not the neurons. And the concepts were stored across *all* of the weights. Ditto for the brain. With a few thousand neurons, in three layers, we could store ALL of the grapheme-phoneme correspondences in one entire language. That is true, but there are about 1000 times as many words as there are graphemes or phonemes, so you need 1000 times as many neurons, or 10^6 times as many connections. (There are 10^6 times as many possible relations between words as between graphemes and phonemes). I think if it were as easy as you say, I think it would have been done by now. Then you will need to represent layered representations: concepts learned from conjunctikons of other concepts rather than layer-1 percepts. Then represent action, negation, operations, intentions, variables... These are high level grammars, like learning how to convert word problems into arithmetic or first order logic. I think anything learned at the level of higher education is going to require a huge network (beyond what is practical now), but I think the underlying learning principles are the same. Oh, I disagree entirely: these are the basic things needed as the *underpinning* of the grammar. You need action for verbs, negation for everything, operations for abstraction, etc. etc. How do humans learn these things using only neurons that follow simple rules? I think learning arithmetic or logic is similar to learning grammar. For example, you can learn to substitute a + b for b + a using the same type of representation you might use to substitute I gave Bob $10 with Bob was given $10 by me. Negation is hard to learn. For example, if you read Nutra-Sweet does not cause stomach cancer, you might start to believe that it does. We learn negation more as an abstract symbol, e.g. neither x nor y means not x and not y. When we build knowledge representation systems, we build logical operators into the system as primitives because we don't know any other way to do it. Logic is hard even for humans to learn. It is a high level language skill. I think it dooms the usual (but always unsuccessful) approach of building a structured knowledge base and trying to tack on a natural language interface later. But you cannot do any estimates like that until the algorithm itself is clear: there are no *algorithms* available for grammar learning, nothing that describes the class of all possible algorithms that do grammar learning. Complexity calculations mean nothing for handwaving suggestions about (eg) representing numbers of neurons: they strictly only apply to situations in which you can point to an algorithm and ask how it behaves. My original dissertation topic (until I changed it to get funding) was to do exactly that. I looked at about 30 different language models, comparing compression ratio with model size, and projecting what size model would be needed to compress text to the entropy estimated by Shannon in 1950 using human text prediction (about 1 bit per character). The graph is here: http://cs.fit.edu/~mmahoney/dissertation/ It suggests very roughly 10^8 to 10^10 bits, in agreement with three other estimates of 10^9 bits: 1. Turing's 1950 estimate, which he did not explain. 2. Landauer's estimate of human long term memory capacity based on memory tests. 3. The approximate information content of all the language you are exposed to through about age 20. This estimate is independent of the algorithm, so it only predicts memory requirements, not speed. If you use a neural network, that is about 10^9 connections. To train on 1 GB of text, you need about 10^18 operations, about a year on a PC. I think there are ways to optimize this, such as activating only a small number of neurons at any one time, and other tricks, but of course I am breaking the rule of getting it to work first and optimizing later. Also, it does not explain why the brain seems to use so much more memory and processing than these estimates, higher by a factor of perhaps 10^4 to 10^6. But of course language evolved to fit our brains, not the other way around. A lot of smart people are working on AGI, including many on this list. I don't believe the reason it hasn't been solved yet is because we are too dumb to figure it out. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http
Re: Neural language models (was Re: [singularity] Help get the 400k SIAI matching challenge on DIGG's front page)
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: I doubt you could model sentence structure usefully with a neural network capable of only a 200 word vocabulary. By the time children learn to use complete sentences they already know thousands of words after exposure to hundreds of megabytes of language. The problem seems to be about O(n^2). As you double the training set size, you also need to double the number of connections to represent what you learned. -- Matt Mahoney, [EMAIL PROTECTED] The problem does not need to be O(n^2). And remember: I used a 200 word vocabulary in a program I wrote 16 years ago, on a machine with only one thousandth of today's power. And besides, solving the problem of understanding sentences could easily be done in principle with even a vocabulary as small as 200 words. Richard Loosemore. What did your simulation actually accomplish? What were the results? What do you think you could achieve on a modern computer? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: Machine Motivation Gets Distorted Again [WAS Re: [singularity] Help get the 400k SIAI matching challenge on DIGG's front page]
--- Richard Loosemore [EMAIL PROTECTED] wrote: Shane Legg wrote: Ben (and others), My impression is that there is a general lack of understanding when it comes to AIXI and related things. It seems that someone who doesn't understand the material makes a statement, which others then take as fact, and the cycle repeats. Part of the problem, I think, is that the material is difficult for people to fully understand. Marcus knows this, and he also understands that while he is a brilliant theoretician, his skills in explaining complex ideas in the simplest way possible are not as strong. To help address this, part of my PhD thesis is going to be, I hope, a very easy to understand explanation of AIXI and universal intelligence --- Marcus doesn't even want me to put any proofs into this part of the thesis, which is most unlike him! I am writing this chapter at the moment and I will let this list know when it has been completed and reviewed. Hopefully then we can all focus on the real weaknesses of this work, rather than the imagined ones. Cheers Shane Shane, Thankyou for being patronizing. Some of us do understand the AIXI work in enough depth to make valid criticism. The problem is that you do not understand the criticism well enough to address it. Richard Loosemore. Richard, I looked at your 2006 AGIRI talk, the one I believe you referenced in our previous discussion on the definition of intelligence, http://www.agiri.org/forum/index.php?act=STf=21t=137 You use the description complex adaptive system, which I agree is a reasonable definition of intelligence. You also assert that mathematics is useless for the analysis of complex systems. Again I agree. But I don't understand your criticism of Shane's work. After all, he is the one who proved the correctness of your assertion. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: Neural language models (was Re: [singularity] Help get the 400k SIAI matching challenge on DIGG's front page)
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: --- Tom McCabe [EMAIL PROTECTED] wrote: --- Matt Mahoney [EMAIL PROTECTED] wrote: Personally, I would experiment with neural language models that I can't currently implement because I lack the computing power. Could you please describe these models? Essentially models in which neurons (with time delays) respond to increasingly abstract language concepts: letters, syllables, words, grammatical roles, phrases, and sentence structures. This is not really new. Models like these have been proposed in the 1980's but were never fully implemented due to lack of computing power. These constraints resulted in connectionist systems in which each concept mapped to a single neuron. Such models can't learn well. There is no mechanism for adding to the vocabulary, for instance. I believe you need at least hundreds of neurons per concept, where each neuron may correlate weakly with hundreds of different concepts. Exactly how many, I don't know. That is why I need to experiment. One problem that bothers me is the disconnect between the information theoretic estimates of the size of a language model, about 10^9 bits, and models based on neuroanatomy, perhaps 10^14 bits. Experiments might tell us what's wrong with our neural models. But how to do such experiments? A fully connected network of 10^9 connections trained on 10^9 bits of data would require about 10^18 operations, about a year on a PC. There are optimizations I could do, such as activating only a small fraction of the neurons at one time, but if the model fails, is it because of these optimizations or because you really do need 10^14 connections, or the training data is bad, or something else? I was building connectionist models of language in the late 80s, early 90s, and your characterizations are a little bit off, here. We used distributed models in which single neurons certainly did not correspond to single concepts. They learned well, and there was no problem getting new vocabulary items into them. I was writing C code on an early model Macintosh computer that was about 1000th the power of the ones available today. You don't really need hundreds of neurons per concept: a few hundred was the biggest net I ever built, and it could cope with about 200 vocabulary items, IIRC. The *real* problem are: (1) encoding the structural aspects of sentences in abstract ways, (2) encoding layered concepts (in which a concept learned today can be the basis for new concepts learned tomorrow) and (3) solving the type-token problem in such a way that the system can represent more than one instance of a concept at once. In essence, my research since then has been all about finding a good way to solve these issues whilst retaining the immense learning power of those early connectionist systems. It's doable. Just have to absorb ten tons of research material and then spit it out in the right way whilst thinking outside the box. All in a day's work. ;-) Richard Loosemore. I doubt you could model sentence structure usefully with a neural network capable of only a 200 word vocabulary. By the time children learn to use complete sentences they already know thousands of words after exposure to hundreds of megabytes of language. The problem seems to be about O(n^2). As you double the training set size, you also need to double the number of connections to represent what you learned. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: [singularity] Help get the 400k SIAI matching challenge on DIGG's front page
--- Eugen Leitl [EMAIL PROTECTED] wrote: On Sun, May 13, 2007 at 05:23:53PM -0700, Matt Mahoney wrote: It is not that hard, really. Each of the 10^5 PCs simulates about 10 mm^3 of You know, repeating assertions doesn't make them any more true. brain tissue. Axon diameter varies but is typically 1-2 microns. This means Where have you pulled that number from? Why not um^3, or m^3, or a cubic lightyear? I assumed you knew that the human brain has a volume of 1000 to 1500 cm^3. If you divide this among 10^5 processors then each processor would simulate a cube about 2 to 2.5 mm on a side with a surface area of about 25-35 mm^2. The little cubes only need to communicate with their 6 neighbors, so you can map the simulation onto a hierarchical network where most of the communication is local. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: [singularity] Help get the 400k SIAI matching challenge on DIGG's front page
--- Tom McCabe [EMAIL PROTECTED] wrote: Helen Keller at ~8 didn't have language, as she hadn't learned sign language and there was no other real means for her to learn grammar and sentence structure. Yet she was still clearly intelligent. If Babelfish was perfect- could pick up on every single grammatical detail and nuance- would it start learning French cooking or write a novel or learn how to drive or do any of that other stuff we associate with intelligence. No, but as long as you define intelligence as exactly like a human we will never have AGI. I don't care if my calculator doesn't know how many fingers I am holding up. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: [singularity] Help get the 400k SIAI matching challenge on DIGG's front page
--- Tom McCabe [EMAIL PROTECTED] wrote: --- Matt Mahoney [EMAIL PROTECTED] wrote: --- Tom McCabe [EMAIL PROTECTED] wrote: You cannot get large amounts of computing power simply by hooking up a hundred thousand PCs for problems that are not easily parallelized, because you very quickly run into bandwidth limitations even with gigabit Ethernet. Parts of the brain are constantly communicating with one another; I would be very surprised if you could split up the brain effectively enough to be able to both run one tiny piece on a PC and have the PCs communicate effectively in realtime. - Tom It is not that hard, really. Each of the 10^5 PCs simulates about 10 mm^3 of brain tissue. Axon diameter varies but is typically 1-2 microns. This means each bit of brain tissue has at most on the order of 10^7 inputs and outputs, each carrying 10 bits per second of information, or 100 Mb/s. This was barely within Google's network capacity in 2000, and probably well within it now. http://en.wikipedia.org/wiki/Google_platform Hmmm...This is an interesting issue. Do you have a link to a paper on brain bandwidth? I just googled axon diameter and found several references. There is a wide range so I used the low end to be conservative and did the math. I probably should consider dendrites too, but these tend not to be very long. I figure it's close enough for an order of magnitude estimate. I think individuals and small groups trying to build AGI will have a hard time competing with Google due to the cost of hardware. Hardware cost will not be a primary issue. The cost of hardware decreases exponentially with Moore's Law; the cost of solving a whole tangle of confusing problems does not. Nobody is anywhere near the stage where they have a program to run and they're looking for a computer. It's like saying that anyone trying to build an airplane will find it impossible to compete with existing shipbuilders, because of their vast metalworking capacity. It's true we can do theoretical work but the lack of computing power is definitely an obstacle. It has a strong effect on the direction of research. In the early days of AI when hardware was inadequate by a factor of a billion, we used symbolic approaches in narrow domains with hand coded rules. More recently when hardware was only inadequate by a million, we were able to experiment with statistical approaches, machine learning, and low level vision and language models. It is possible that a lot of the brain's computing power is used to overcome the limitations of individual neurons (speed, noise, reliability, fatigue) and we will find more efficient solutions. This hasn't happened yet, but I can't say that it won't. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: [singularity] Help get the 400k SIAI matching challenge on DIGG's front page
--- Tom McCabe [EMAIL PROTECTED] wrote: --- Matt Mahoney [EMAIL PROTECTED] wrote: Language and vision are prerequisites to AGI. No, they aren't, unless you care to suggest that someone with a defect who can't see and can't form sentences (eg, Helen Keller) is unintelligent. Helen Keller had language. One could argue that language alone is sufficient for AI, as Turing did. But everyone has a different opinion on what is AGI and what isn't. Any future Friendly AGI isn't going to obey us exactly in every respect, because it's *more moral* than we are. Should an FAI obey a request to blow up the world? That is what worries me. I think it is easier to program an AGI for blind obedience (its top level goal is to serve humans) than to program it to make moral judgments in the best interest of humans, without specifying what that means. I gave this example on Digg. Suppose the AGI (being smarter than us) figures out that consciousness and free will are illusions of our biologically programmed brains, and that there is really no difference between a human brain and a simulation of a brain on a computer. We may or may not have the technology for uploading, but suppose the AGI decides (for reasons we don't understand) that it doesn't need it. Therefore it is in our best interest (or irrelevant) to destroy the human race. We cannot rule out this possibility because a lesser intelligence cannot predict what a greater intelligence will do. If you measure intelligence using algorithmic complexity, then Legg proved this formally. http://www.vetta.org/documents/IDSIA-12-06-1.pdf Or maybe an analogy would be more convincing. Humans acting in the best interests of their pets may put them down when they have a terminal disease, or for other reasons they can't comprehend. Who should make this decision? What will happen when the AGI is as advanced over humans as humans are over dogs or insects or bacteria? Perhaps the smarter it gets, the less relevant human life will be. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: [singularity] Help get the 400k SIAI matching challenge on DIGG's front page
--- Tom McCabe [EMAIL PROTECTED] wrote: --- Matt Mahoney [EMAIL PROTECTED] wrote: I posted some comments on DIGG and looked at the videos by Thiel and Yudkowsky. I'm not sure I understand the push to build AGI with private donations when companies like Google are already pouring billions into the problem. Private companies like Google are, as far as I am aware, spending exactly $0 on AGI. The things Google is interested in, such as how humans process information and how they decide what is relevant, are very specific subsets of this goal in the same way that fire and iron are very specific subsets of the internal combustion engine. Language and vision are prerequisites to AGI. Google has an interest in improving search results. It already does a pretty good job with natural language questions. They would also like to return relevant images, video, and podcasts without requiring humans to label them. They want to filter porn and spam. They want to deliver relevant and personalized ads. These are all AI problems. Google has billions to spend on these problems. Google already have enough computing problem to do a crude simulation of a human brain, but of course that is not what they are trying to do. Why would they want to copy human motivations? Doing this well requires human capabilities such as language and vision, but does not require duplicating the human motivational system. The top level goal of humans is to propagate their DNA. The top level goal of machines should be to serve humans. You do realize how hard a time you're going to have defining that? Remember Asimov's First Law: A robot shall not harm a human or through inaction allow a human to come to harm? Well, humans are always hurting themselves through wars and such, and so the logical result is totalitarianism, which most of us would consider very bad. I realize the problem will get harder as machines get smarter. But right now I don't see any prospect of a general solution. It will have to be solved for each new machine. But there is nothing we can do about human evil. If someone wants to build a machine to kill people, well that is already a problem. The best we can do is try to prevent accidental harm. We have always built machines this way. Do I really need to explain what's wrong with the we've always done it that way argument? It hasn't gotten any better since the South used it to justify slavery. I phrased it in the past tense because I can't predict the future. What I should say is that there is no reason to build machines to disobey their owners, and I don't expect that we will do so in the future. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: [singularity] Why do you think your AGI design will work?
--- Joshua Fox [EMAIL PROTECTED] wrote: AGI builders, what evidence do you have that your design will work? None, because we have not defined what AGI is. One definition of AGI is passing the Turing test. That will not happen. A machine can just as easily fail by being too smart, too fast, or too obedient, as it can by being not smart enough. Machines have been smarter than humans in some areas and less smart in others for the last 50 years. Even a machine that is superior to human intellect in every conceivable way would not be mistaken for human. There is no economic incentive to dumb down a machine just to duplicate human limitations. If AGI is not the Turing test, then what is it? What test do you propose? Without a definition, we should stop calling it AGI and focus on the problems for which machines are still inferior to humans, such as language or vision. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: [singularity] Why do you think your AGI design will work?
--- Eugen Leitl [EMAIL PROTECTED] wrote: On Tue, Apr 24, 2007 at 01:35:31PM -0700, Matt Mahoney wrote: None, because we have not defined what AGI is. AGI is like porn. I'll know it when I'll see it. Not really. You recognize porn because you have seen examples of porn and not-porn. If you give a test to people who have never seen porn, I think they would fail. But we will never do this test because of ethical concerns about showing porn for the first time to the only subjects you are likely to find, children. I also don't think you will recognize AGI. You have never seen examples of it. Earlier I posted examples of Google passing the Turing test, but nobody believes that is AGI. If nothing is ever labeled AGI, then nothing ever will be. One definition of AGI is passing the Turing test. That will not happen. A machine can just as easily fail by being too smart, too fast, or too obedient, The Turing test implies ability to deceive. If your system can't deceive a human, it has failed the test. ELIZA has already passed. So we can all quit and go home now. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: [singularity] Why do you think your AGI design will work?
--- Mike Tintner [EMAIL PROTECTED] wrote: There is a difference between your version: achieving goals which can be done, if I understand you, by algorithms - and my goal-SEEKING, which is done by all animals, and can't be done by algorithms alone. It involves finding your way as distinct from just following the way set by programmed rules. There is an algorithm. We just don't know what it is. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=4007604user_secret=8eb45b07
Re: [singularity] Implications of an already existing singularity.
--- Tom McCabe [EMAIL PROTECTED] wrote: If G were changing with time, then we'd see the Moon's orbit moving outward faster than the 10 cm/year or so caused by tides. - Tom I agree there is no evidence of this. But here is another mystery of physics. The radius of a black hole's event horizon is where the escape velocity equals the speed of light, Gm/r = c. For the universe, the radius is close to the size of the universe, r ~ Tc. So why did the universe (or large regions of it) not collapse into black holes when it was much younger and denser? I believe that an observer approaching a black hole in a free fall observes nearby objects accelerating away in all directions. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=11983
Re: [singularity] Implications of an already existing singularity.
--- Charles D Hixson [EMAIL PROTECTED] wrote: Matt Mahoney wrote: --- Eugen Leitl [EMAIL PROTECTED] wrote: ... A proton is a damn complex system. Don't see how you could equal it with one mere bit. I don't. I am equating one bit with a volume of space about the size of a proton. The actual number of baryons in the universe is smaller, about 10^80. If you squashed the universe flat, it would form a sheet about one proton thick. But I am also pointing out a coincidence (or not) of physics. But you will note that the volume of the universe is proportional to T^3, not T^2, so if the relation is not a coincidence, then either the properties of the proton or one of the other physical constants would not be constant. And BTW I agree that we cannot prove or disprove that the universe is a simulation. -- Matt Mahoney, [EMAIL PROTECTED] FWIW, you could cut down on the computational needs a whole lot if you only simulated one brain and used lazy evaluation to derive anything it might be experiencing. (Where did all you Zombies come from?) For that matter, the simulation could have started only a few nano-seconds ago and might stop now. ... Any assumption you make about the nature of the simulation that we might be running on is unverifiable. (Some of them are falsifiable.) A while back I described 5 scenarios for a simulated universe in order of decreasing algorithmic complexity, and therefore in increasing order of likelihood (given a Solomonoff distribution). But as the complexity decreased, the amount of computation increased. I concluded that the most likely scenario was an enumeration of all Turing machines, whose algorithmic complexity is K(N), the complexity of the set of natural numbers (very small). And no, I don't know what is doing this computation (turtles all the way down). But it is a general property of agents in a simulation that they lack the computational power to model their environment, whether finite or infinite. So it would be surprising if I did know. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=11983
Re: [singularity] Implications of an already existing singularity.
--- Tom McCabe [EMAIL PROTECTED] wrote: --- Craig [EMAIL PROTECTED] wrote: Kurzweil already postulated this a while ago. Although I don't agree with his conclusions. He says that if any society were to attain the singularity then their presence would already be felt, and since we can feel no presence then essentially this proves that humans are the only sentient life forms in EXISTENCE. I wholeheartedly disagree with Kurzweil's reasoning in this matter, since he takes such a human perspective in regards to imagining an alien technology. I think his stance is very presumptuous on his part. For instance he assumes that we haven't felt their presence merely because there isn't anything to detect. When in fact he never considered that human senses or sciences may not be acute enough to detect them. Human senses, while crude, are good enough to detect a wholesale rearrangement of a large majority of the matter in the solar system. A technology this advanced could also reprogram your neurons to make you believe whatever it wanted. There is no way you could detect this. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=11983
Re: Entropy of the universe [WAS Re: [singularity] Implications of an already existing singularity.]
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: *The entropy of the universe is of the order T^2 c^5/hG ~ 10^122 bits, where T is the age of the universe, c is the speed of light, h is Planck's constant and G is the gravitational constant. By coincidence (or not?), each bit would occupy the volume of a proton. (The physical constants do not depend on any particle properties). A small but crucial point: this is the entropy of everything within the horizon visible from *here*. What about the stuff (possibly infinite amounts of stuff) that lies beyond the curvature horizon? In a simulation, you don't need to compute it. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=11983
Re: [singularity] Scenarios for a simulated universe
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: --- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: --- Richard Loosemore [EMAIL PROTECTED] wrote: What I wanted was a set of non-circular definitions of such terms as intelligence and learning, so that you could somehow *demonstrate* that your mathematical idealization of these terms correspond with the real thing, ... so that we could believe that the mathematical idealizations were not just a fantasy. The last time I looked at a dictionary, all definitions are circular. So you win. Sigh! This is a waste of time: you just (facetiously) rejected the fundamental tenet of science. Which means that the stuff you were talking about was just pure mathematical fantasy, after all, and nothing to do with science, or the real world. Richard Loosemre. What does the definition of intelligence have to do with AIXI? AIXI is an optimization problem. The problem is to maximize an accumulated signal in an unknown environment. AIXI says the solution is to guess the simplest explanation for past observation (Occam's razor), and that this solution is not computable in general. I believe these principles have broad applicability to the design of machine learning algorithms, regardless of whether you consider such algorithms intelligent. You're going around in circles. If you were only talking about machine learning in the sense of an abstract mathematical formalism that has no relationship to learning, intelligence or anything going on in the real world, and in particular the real world in which some of us are interested in the problem of trying to build an intelligent system, then, fine, all power to you. At *that* level you are talking about a mathematical fantasy, not about science. But you did not do that: you made claims that went far beyond the confines of a pure, abstract mathematical formalism: you tried to relate that to an explanation of why Occam's Razor works (and remember, the original meaning of Occam's Razor was all about how an *intelligent* being should use its intelligence to best understand the world), and you also seemed to make inferences to the possibility that the real world was some kind of simulation. It seems to me that you are trying to have your cake and eat it too. I claim that AIXI has practical applications to machine learning. I also claim (implicitly) that machine learning has practical applications to the real world. Therefore, I claim that AIXI has practical applications to the real world (i.e. as Occam's razor). Further, because AIXI requires that the unknown environment be computable, I claim that we cannot exclude the possibility that the universe is a simulation. If Occam's razor did not work in practice, then you could claim that the universe is not computable, and therefore could not be a simulation. This really has nothing to do with the definition of intelligence. You can accept Turing's definition, which would exclude all animals except Homo Sapiens. You can accept a broader definition that would include machine learning. Both the human brain and linear regression algorithms make use of Occam's razor. I don't care if you call them intelligent or not. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=11983
Re: [singularity] Scenarios for a simulated universe
--- Stathis Papaioannou [EMAIL PROTECTED] wrote: On 3/6/07, Mitchell Porter [EMAIL PROTECTED] wrote: You radically overstate the expected capabilities of quantum computers. They can't even do NP-complete problems in polynomial time. http://scottaaronson.com/blog/?p=208 What about a computer (classical will do) granted an infinity of cycles through, for example, a Freeman Dyson or Frank Tipler type mechanism? No matter how many cycles it takes to compute a particular simulated world, any delay will be transparent to observers in that world. It only matters that the computation doesn't stop before it is completed. The computation would also require infinite memory (a Turing machine), or else it would cycle. Although our universe might be the product of a Turing machine, the physics of our known universe will only allow finite memory. The number of possible quantum states of a closed system with finite size and mass is finite. For our universe (big bang model), the largest memory you could construct would be on the order of c^5 T^2/hG ~ 10^122 bits (where c is the speed of light, T is the age of the universe, h is Planck's constant and G is the gravitational constant. (Coincidentally, each bit would occupy about the volume of a proton or neutron). A quantum computer is weaker than a finite state machine. A quantum computer is restricted to time-reversible computation, so operations like bit assignment or copying are not allowed. And even if you had a Turing machine, you still could not compute a solution to AIXI. It is not computable, like the halting problem. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=11983
Re: [singularity] Scenarios for a simulated universe
--- Richard Loosemore [EMAIL PROTECTED] wrote: Matt Mahoney wrote: --- Richard Loosemore [EMAIL PROTECTED] wrote: What I wanted was a set of non-circular definitions of such terms as intelligence and learning, so that you could somehow *demonstrate* that your mathematical idealization of these terms correspond with the real thing, ... so that we could believe that the mathematical idealizations were not just a fantasy. The last time I looked at a dictionary, all definitions are circular. So you win. Sigh! This is a waste of time: you just (facetiously) rejected the fundamental tenet of science. Which means that the stuff you were talking about was just pure mathematical fantasy, after all, and nothing to do with science, or the real world. Richard Loosemre. What does the definition of intelligence have to do with AIXI? AIXI is an optimization problem. The problem is to maximize an accumulated signal in an unknown environment. AIXI says the solution is to guess the simplest explanation for past observation (Occam's razor), and that this solution is not computable in general. I believe these principles have broad applicability to the design of machine learning algorithms, regardless of whether you consider such algorithms intelligent. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=11983
Re: [singularity] Scenarios for a simulated universe
--- Richard Loosemore [EMAIL PROTECTED] wrote: What I wanted was a set of non-circular definitions of such terms as intelligence and learning, so that you could somehow *demonstrate* that your mathematical idealization of these terms correspond with the real thing, ... so that we could believe that the mathematical idealizations were not just a fantasy. The last time I looked at a dictionary, all definitions are circular. So you win. P.S. The above definition is broken anyway: what about unsupervised learning? What about learning by analogy? I should have specified supervised learning as an application of AIXI. There are subsets, H, of Turing machines for which there are efficient algorithms for finding a small h in H that is consistent with the training data. Examples include decision trees, neural networks, polynomial regression, clustering, etc. However AIXI does not necessarily imply learning. There are other approaches. -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=11983
Re: [singularity] Why We are Almost Certainly not in a Simulation
This discussion on whether the universe exists is interesting, but I think we should be asking a different question: why do we believe that the universe exists? Or more accurately, why do we act as if we believe that the universe exists? I said earlier that humans believe that the universe is real, because those that did not were removed from the gene pool. But I wonder if the issue is more fundamental. Is it possible to program to program any autonomous agent that responds to reinforcement learning (a reward/penalty signal) that does not act as though its environment were real? How would one test for this belief? -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=11983
Re: [singularity] Scenarios for a simulated universe
--- Ben Goertzel [EMAIL PROTECTED] wrote: Matt, I really don't see why you think Hutter's work shows that Occam's Razor holds in any context except AI's with unrealistically massive amounts of computing power (like AIXI and AIXItl) In fact I think that it **does** hold in other contexts (as a strategy for reasoning by modest-resources minds like humans or Novamente), but I don't see how Hutter's work shows this... I admit Hutter did not make claims about machine learning frameworks or Occam's razor, but we should not view his work in such narrow context. Hutter's conclusions about the optimal behavior of rational agents were proven for the following cases: 1. Unrestricted environments (in which case the solution is not computable), 2. Space and time bounded environments (in which case the solution is intractable), 3. Subsets of (1) or (2) such that the environment is consistent with past interaction. But the same reasoning he used in his proofs could just as well be applied to practical cases of machine learning for which efficient solutions are known. The proofs all use the fact that shorter Turing machines are more likely than longer ones (a Solomonoff prior). For example, Hutter does not tell us how to solve linear regression, fitting a straight line to a set of points. What Hutter tells us is two other things: 1. Linear regression is a good predictor, even though a higher order polynomial might have a better fit (because a low order polynomial has lower algorithmic complexity). 2. Linear regression is useful, even though other machine learning algorithms might be better predictors (because a general solution is not computable, so we have to settle for a suboptimal solution). So what I did was two things. First, I used the fact that Occam's razor works in both simulated and real environments (based on extensions of AIXI and empirical observations respectively) to argue that the universe is consistent with a simulation. (This is disturbing because you are not programmed to think this way). Second, I used the same reasoning to guess about the nature of the universe (assuming it is simulated), and the only thing we know is that shorter simulation programs are more likely than longer ones. My conclusion was that bizarre behavior or a sudden end is unlikely, because such events would not occur in the simplest programs. This ought to at least be reassuring. -- Matt Mahoney -- Matt Mahoney, [EMAIL PROTECTED] - This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?list_id=11983