Re: [agi] US PATENT ISSUED for the TEN ETHICAL LAWS OF ROBOTICS
This is a real patent, unfortunately... http://patft.uspto.gov/netacgi/nph-Parser?Sect2=PTO1&Sect2=HITOFF&p=1&u=%2Fnetahtml%2FPTO%2Fsearch-bool.html&r=1&f=G&l=50&d=PALL&RefSrch=yes&Query=PN%2F6587846 But I think it will expire before anyone has the technology to implement it. :-) -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
Re: [agi] US PATENT ISSUED for the TEN ETHICAL LAWS OF ROBOTICS
--- On Tue, 7/22/08, John LaMuth <[EMAIL PROTECTED]> wrote: > Assuming I'm a Troll is pretty harsh, isnt it ? I looked at your patent. Nowadays you can patent any kind of nonsense. USPTO finds it easier to just grant the patent and let the courts sort it out. The plaintiff hires an expert who says "X infringes on Y". The defendant hires an expert who says "X does not infringe on Y". The judge, who doesn't know anything about X or Y, tries to figure out who's lying. So perhaps if you actually have a contribution to AGI, you can point us to a published paper describing the experimental results of the AI you have built? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
Re: [agi] How do we know we don't know?
This is not a hard problem. A model for data compression has the task of predicting the next bit in a string of unknown origin. If the string is an encoding of natural language text, then modeling is an AI problem. If the model doesn't know, then it assigns a probability of about 1/2 to each of 0 and 1. Probabilities can be easily detected from outside the model, regardless of the intelligence level of the model. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
Re: [agi] How do we know we don't know?
Words compress smaller than non-words. -- Matt Mahoney, [EMAIL PROTECTED] - Original Message From: Mike Tintner <[EMAIL PROTECTED]> To: agi@v2.listbox.com Sent: Thursday, July 31, 2008 5:47:54 AM Subject: Re: [agi] How do we know we don't know? Vlad: >I think Hofstadter's exploration of jumbles ( > http://en.wikipedia.org/wiki/Jumble ) covers this ground. You don't > just recognize the word, you work on trying to connect it to what you > know, and if set of letters didn't correspond to any word, you give > up. There's still more to word recognition though than this. How do we decide what is and isn't, may or may not be a word? A neologism? What may or may not be words from: cogrough dirksilt thangthing artcop coggourd cowstock or "fomlepaung" or whatever? --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
Re: [agi] META: do we need a stronger "politeness code" on this list?
I think the "sniper based" moderation policy for SL4 works pretty well and might be appropriate for this list. http://www.sl4.org/intro.html -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
Re: [agi] Groundless reasoning
- Original Message From: Ben Goertzel <[EMAIL PROTECTED]> >My perspective on grounding is partially summarized here > >www.goertzel.org/papers/PostEmbodiedAI_June7.htm I agree that AGI should ideally have multiple sources of knowledge as you describe: explicitly taught, learned from conversation (or reading), and learned from embodied experience. But again, we must start with a specification. The goal of AGI is not to produce artificial humans, but to do useful work with the lowest possible cost. Depending on their jobs, AGI will have a huge range of capabilities and non-human like environments. They will need concepts that don't translate into English. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
Re: [agi] brief post on possible path to agi
rick the ponderer wrote: >There was a brief post on a possible path to agi at >http://news.ycombinator.com/item?id=271202 The problem with speech recognition is not converting speech into words, but converting words into useful actions. "Press 1 or say 'yes'" is not a solution to the speech recognition problem. Your proposal looks similar to my proposal for competitive message routing, although lacking in detail. http://www.mattmahoney.net/agi.html Either way, it will be expensive. AGI is worth the labor it replaces, valued at over US $1 quadrillion worldwide over the next 30 years. When I see proposals that purport to solve AGI on a budget of $1 million or $1 billion or even $1 trillion, I can only shake my head. -- Matt Mahoney, [EMAIL PROTECTED] - Original Message From: rick the ponderer <[EMAIL PROTECTED]> To: agi@v2.listbox.com Sent: Saturday, August 9, 2008 4:18:34 PM Subject: Re: [agi] brief post on possible path to agi On 8/9/08, Brad Paulsen <[EMAIL PROTECTED]> wrote: rick, Except that the author bases his argument on an inaccurate premise. See: http://www.speech.cs.cmu.edu/ for an excellent speech recognizer. It's open source (has been for at least a decade). In fact, the Apple (who the author omitted) and Microsoft both based their speech recognizers on Sphinx (I know, I worked in the ATG research team at Apple that developed their speech recognizer). The Festival project (Google it), also partly hosted at CMU, is a world-class speech synthesizer. Also open source and free to all. And, there are others (eSpeak - based on Festival, GPL and free). Cheers, Brad rick the ponderer wrote: There was a brief post on a possible path to agi at http://news.ycombinator.com/item?id=271202 yesterday. Essentially it involves masses of people creating binary classifiers in a economic market system, similar to how content is created on the web today (though with a micropayment system rather than advertising supported model). *agi* | Archives <https://www.listbox.com/member/archive/303/=now> <https://www.listbox.com/member/archive/rss/303/> | Modify <https://www.listbox.com/member/?&;;> Your Subscription [Powered by Listbox] <http://www.listbox.com> --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?&;; Powered by Listbox: http://www.listbox.com I'm the author of the post - I don't really know about speech recognizers (i had heard of those and have tried using them on my computer, I just meant products widely available to the nontechnical public), But my argument is they're not good enought yet (otherwise human transcription services wouldn't exist) because enough human labelled data hasn't been used to create them, and such an undertaking would require many thousands/millions of people (if you include video and text recognition). agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
Re: [agi] brief post on possible path to agi
rick the ponderer <[EMAIL PROTECTED]> wrote: >I'm not saying that speech recognition is equivalent to human language understanding, I'm arguing that speech >recognition could be improved to the point it can recognise almost every speaker at an speed in any accent. To get human level word error rates, you need human level AI. Humans use vast knowledge to fill in missing words, e.g. "the cat caught a m". >I'm not proposing to solve it with a millions dollars or trillion etc. My argument is the exact opposite, that it is >too large to even be attempted by any set pool of funding. That is my proposal too, but my economic model is different. Information has negative value on average. People don't compete to buy information. Rather they compete for attention. This is not just advertising, but human nature. We have personal websites and write blogs to satisfy our egos. Why would I spend tens of dollars worth of time to bother posting this noncommercial message? A system where you buy the services of a classifier can be exploited. I could put up a server that charged no money, then inject advertising into its output. How do you know which servers to trust? In CMR, peers have to authenticate their identities and establish a reputation for providing useful information. Peers have an incentive to filter messages routed through them and block spam, otherwise they will be blocked by more intelligent peers. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
Re: [agi] brief post on possible path to agi
rick the ponderer <[EMAIL PROTECTED]> wrote: >Regarding cempeting to buy information - I'm not suggesting that at >all, people would be competing to sell the services of their classifier >(and shopping around for the best classifier to consume or build on). >It would be like the web services model - like for example at strikeiron.com My point is that for most information, free is too expensive. Then how is your model funded? You have to collect money from the information providers and increase its value up to at least zero by filtering out all but the most useful, like for example, Google. The missing technology is distributed indexing. This has a number of problems. First, it is very expensive to compete with Google. Its servers make up about 0.1% of the world's computing power. Second, competing web services would be inefficient because of the duplication of network traffic (spiders) and index storage. A centralized model favors a monopoly. Third, Google it is very limited. After a web page update, Google may take days to find it and update its index. Distributed indexing would solve these problems. Nobody would control the index. Everyone would have an incentive to contribute computing power (storage and bandwidth) and high quality information in exchange for the ability to send messages. There would be no distinction between queries and updates. You just send a message and it is routed to anyone who cares. Imagine if a Google query could initiate a conversation in real time. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=108809214-a0d121 Powered by Listbox: http://www.listbox.com
Re: [agi] Meet the world's first robot controlled exclusively by living brain tissue
Mike Tintner <[EMAIL PROTECTED]> wrote: > http://www.wired.com/wired/archive/8.02/warwick.html An interesting perspective. Instead of brain tissue controlling a machine, we have a brain wanting to be controlled by a machine. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] How We Look At Faces
Mike Tintner wrote:
>{I wonder whether the difference below *is* biological - due to narrower
>eyes taking that little bit longer to process?]
Or there is a learned difference in the way Caucasians and Asians process
visual information due to written language differences (a larger alphabet).
Or there is a genetic difference, such as a broader fovea in Asians, or
differences in the eye muscles resulting in a lower saccade rate.
Or as the paper suggests, it is rude to stare at people in Asian cultures, so
they learn to recognize faces without looking directly at the eyes.
You can't tell from the paper, but perhaps you could conclude that for an AI,
getting the low level features right is not critical for face recognition.
-- Matt Mahoney, [EMAIL PROTECTED]
- Original Message
From: Mike Tintner <[EMAIL PROTECTED]>
To: agi@v2.listbox.com
Sent: Wednesday, August 20, 2008 10:16:42 AM
Subject: [agi] How We Look At Faces
{I wonder whether the difference below *is* biological - due to narrower
eyes taking that little bit longer to process?]
Culture Shapes How We Look at Faces
Caroline Blais1,2, Rachael E. Jack1, Christoph Scheepers1, Daniel Fiset1,2,
Roberto Caldara1
1 Department of Psychology, University of Glasgow, Glasgow, United Kingdom,
2 Département de Psychologie, Université de Montréal, Montréal, Canada
Abstract
Background
Face processing, amongst many basic visual skills, is thought to be
invariant across all humans. From as early as 1965, studies of eye movements
have consistently revealed a systematic triangular sequence of fixations
over the eyes and the mouth, suggesting that faces elicit a universal,
biologically-determined information extraction pattern.
Methodology/Principal Findings
Here we monitored the eye movements of Western Caucasian and East Asian
observers while they learned, recognized, and categorized by race Western
Caucasian and East Asian faces. Western Caucasian observers reproduced a
scattered triangular pattern of fixations for faces of both races and across
tasks. Contrary to intuition, East Asian observers focused more on the
central region of the face.
Conclusions/Significance
These results demonstrate that face processing can no longer be considered
as arising from a universal series of perceptual events. The strategy
employed to extract visual information from faces differs across cultures.
Source: PLoS One [Open Access]
http://www.plosone.org/article/info:doi/10.1371/journal.pone.0003022
---
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/
Modify Your Subscription: https://www.listbox.com/member/?&;
Powered by Listbox: http://www.listbox.com
---
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51
Powered by Listbox: http://www.listbox.com
Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)
Valentina Poletti <[EMAIL PROTECTED]> wrote:
> I was wondering why no-one had brought up the information-theoretic aspect of
> this yet.
It has been studied. For example, Hutter proved that the optimal strategy of a
rational goal seeking agent in an unknown computable environment is AIXI: to
guess that the environment is simulated by the shortest program consistent with
observation so far [1]. Legg and Hutter also propose as a measure of universal
intelligence the expected reward over a Solomonoff distribution of environments
[2].
These have profound impacts on AGI design. First, AIXI is (provably) not
computable, which means there is no easy shortcut to AGI. Second, universal
intelligence is not computable because it requires testing in an infinite
number of environments. Since there is no other well accepted test of
intelligence above human level, it casts doubt on the main premise of the
singularity: that if humans can create agents with greater than human
intelligence, then so can they.
Prediction is central to intelligence, as I argue in [3]. Legg proved in [4]
that there is no elegant theory of prediction. Predicting all environments up
to a given level of Kolmogorov complexity requires a predictor with at least
the same level of complexity. Furthermore, above a small level of complexity,
such predictors cannot be proven because of Godel incompleteness. Prediction
must therefore be an experimental science.
There is currently no software or mathematical model of non-evolutionary
recursive self improvement, even for very restricted or simple definitions of
intelligence. Without a model you don't have friendly AI; you have accelerated
evolution with AIs competing for resources.
References
1. Hutter, Marcus (2003), "A Gentle Introduction to The Universal Algorithmic
Agent {AIXI}",
in Artificial General Intelligence, B. Goertzel and C. Pennachin eds.,
Springer. http://www.idsia.ch/~marcus/ai/aixigentle.htm
2. Legg, Shane, and Marcus Hutter (2006),
A Formal Measure of Machine Intelligence, Proc. Annual machine
learning conference of Belgium and The Netherlands (Benelearn-2006).
Ghent, 2006. http://www.vetta.org/documents/ui_benelearn.pdf
3. http://cs.fit.edu/~mmahoney/compression/rationale.html
4. Legg, Shane, (2006), Is There an Elegant Universal Theory of Prediction?,
Technical Report IDSIA-12-06, IDSIA / USI-SUPSI,
Dalle Molle Institute for Artificial Intelligence, Galleria 2, 6928 Manno,
Switzerland.
http://www.vetta.org/documents/IDSIA-12-06-1.pdf
-- Matt Mahoney, [EMAIL PROTECTED]
---
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51
Powered by Listbox: http://www.listbox.com
Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)
Eric Burton <[EMAIL PROTECTED]> wrote: >>These have profound impacts on AGI design. First, AIXI is (provably) not >>computable, >>which means there is no easy shortcut to AGI. Second, universal intelligence >>is not >>computable because it requires testing in an infinite number of environments. >>Since >>there is no other well accepted test of intelligence above human level, it >>casts doubt on >>the main premise of the singularity: that if humans can create agents with >>greater than >>human intelligence, then so can they. > >I don't know for sure that these statements logically follow from one >another. They don't. I cannot prove that there is no non-evolutionary model of recursive self improvement (RSI). Nor can I prove that there is. But it is a question we need to answer before an evolutionary model becomes technically feasible, because an evolutionary model is definitely unfriendly. >Higher intelligence bootstrapping itself has already been proven on >Earth. Presumably it can happen in a simulation space as well, right? If you mean the evolution of humans, that is not an example of RSI. One requirement of friendly AI is that an AI cannot alter its human-designed goals. (Another is that we get the goals right, which is unsolved). However, in an evolutionary environment, the parents do not get to choose the goals of their children. Evolution chooses goals that maximize reproductive fitness, regardless of what you want. I have challenged this list as well as the singularity and SL4 lists to come up with an example of a mathematical, software, biological, or physical example of RSI, or at least a plausible argument that one could be created, and nobody has. To qualify, an agent has to modify itself or create a more intelligent copy of itself according to an intelligence test chosen by the original. The following are not examples of RSI: 1. Evolution of life, including humans. 2. Emergence of language, culture, writing, communication technology, and computers. 3. A chess playing (or tic-tac-toe, or factoring, or SAT solving) program that makes modified copies of itself by randomly flipping bits in a compressed representation of its source code, and playing its copies in death matches. 4. Selective breeding of children for those that get higher grades in school. 5. Genetic engineering of humans for larger brains. 1 fails because evolution is smarter than all of human civilization if you measure intelligence in bits of memory. A model of evolution uses 10^37 bits (10^10 bits of DNA per cell x 10^14 cells in the human body x 10^10 humans x 10^3 ratio of biomass to human mass). Human civilization has at most 10^25 bits (10^15 synapses in the human brain x 10^10 humans). 2 fails because individual humans are not getting smarter with each generation, at least not nearly as fast as civilization is advancing. Rather, there are more humans, and we are getting better organized through specialization of tasks. Human brains are not much different than they were 10,000 years ago. 3 fails because there are no known classes of problems that are provably hard to solve but easy to verify. Tic-tac-toe and chess have bounded complexity. It has not been proven that factoring is harder than multiplication. We don't know that P != NP, and even if we did, many NP-complete problems have special cases that are easy to solve, and we don't know how to program the parent to avoid these cases through successive generations. 4 fails because there is no evidence that above a certain level (about IQ 200) that childhood intelligence correlates with adult success. The problem is that adults of average intelligence can't agree on how success should be measured*. 5 fails for the same reason. *For example, the average person recognizes Einstein as a genius not because they are awed by his theories of general relativity, but because other people have said so. If you just read his papers (without understanding their great insights) and knew that he never learned to drive a car, you might conclude differently. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] How Would You Design a Play Machine?
Kittens play with small moving objects because it teaches them to be better hunters. Play is not a goal in itself, but a subgoal that may or may not be a useful part of a successful AGI design. -- Matt Mahoney, [EMAIL PROTECTED] - Original Message From: Mike Tintner <[EMAIL PROTECTED]> To: agi@v2.listbox.com Sent: Monday, August 25, 2008 8:59:06 AM Subject: Re: [agi] How Would You Design a Play Machine? Brad, That's sad. The suggestion is for a mental exercise, not a full-scale project. And play is fundamental to the human mind-and-body - it characterises our more mental as well as more physical activities - drawing, designing, scripting, humming and singing scat in the bath, dreaming/daydreaming & much more. It is generally acknowledged by psychologists to be an essential dimension of creativity - which is the goal of AGI. It is also an essential dimension of animal behaviour and animal evolution. Many of the smartest companies have their play areas. But I'm not aware of any program or computer design for play - as distinct from elaborating systematically and methodically or "genetically" on themes - are you? In which case it would be good to think about one - it'll open your mind & give you new perspectives. This should be a group where people are not too frightened to play around with ideas. Brad:> Mike Tintner wrote: "...how would you design a play machine - a machine > that can play around as a child does?" > > I wouldn't. IMHO that's just another waste of time and effort (unless > it's being done purely for research purposes). It's a diversion of > intellectual and financial resources that those serious about building an > AGI any time in this century cannot afford. I firmly believe if we had > not set ourselves the goal of developing human-style intelligence > (embodied or not) fifty years ago, we would already have a working, > non-embodied AGI. > > Turing was wrong (or at least he was wrongly interpreted). Those who > extended his imitation test to humanoid, embodied AI were even more wrong. > We *do not need embodiment* to be able to build a powerful AGI that can be > of immense utility to humanity while also surpassing human intelligence in > many ways. To be sure, we want that AGI to be empathetic with human > intelligence, but we do not need to make it equivalent (i.e., "just like > us"). > > I don't want to give the impression that a non-Turing intelligence will be > easy to design and build. It will probably require at least another > twenty years of "two steps forward, one step back" effort. So, if we are > going to develop a non-human-like, non-embodied AGI within the first > quarter of this century, we are going to have to "just say no" to Turing > and start to use human intelligence as an inspiration, not a destination. > > Cheers, > > Brad > > > > Mike Tintner wrote: >> Just a v. rough, first thought. An essential requirement of an AGI is >> surely that it must be able to play - so how would you design a play >> machine - a machine that can play around as a child does? >> >> You can rewrite the brief as you choose, but my first thoughts are - it >> should be able to play with >> a) bricks >> b)plasticine >> c) handkerchiefs/ shawls >> d) toys [whose function it doesn't know] >> and >> e) draw. >> >> Something that should be soon obvious is that a robot will be vastly more >> flexible than a computer, but if you want to do it all on computer, fine. >> >> How will it play - manipulate things every which way? >> What will be the criteria of learning - of having done something >> interesting? >> How do infants, IOW, play? >> --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)
John, I have looked at your patent and various web pages. You list a lot of nice sounding ethical terms (honor, love, hope, peace, etc) but give no details on how to implement them. You have already admitted that you have no experimental results, haven't actually built anything, and have no other results such as refereed conference or journal papers describing your system. If I am wrong about this, please let me know. -- Matt Mahoney, [EMAIL PROTECTED] - Original Message From: John LaMuth <[EMAIL PROTECTED]> To: agi@v2.listbox.com Sent: Sunday, August 24, 2008 11:21:30 PM Subject: Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment) - Original Message - From: "Matt Mahoney" <[EMAIL PROTECTED]> To: Sent: Sunday, August 24, 2008 2:46 PM Subject: Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment) I have challenged this list as well as the singularity and SL4 lists to come up with an example of a mathematical, software, biological, or physical example of RSI, or at least a plausible argument that one could be created, and nobody has. To qualify, an agent has to modify itself or create a more intelligent copy of itself according to an intelligence test chosen by the original. The following are not examples of RSI: > > 1. Evolution of life, including humans. > 2. Emergence of language, culture, writing, communication technology, and computers. > -- Matt Mahoney, [EMAIL PROTECTED] > ### * Matt Where have you been for the last 2 months ?? I had been talking then about my 2 US Patents for ethical/friendly AI along lines of a recursive simulation targeting language (topic 2) above. This language agent employs feedback loops and LTM to increase comprehension and accuracy (and BTW - resolves the ethical safeguard problems for AI) ... No-one yet has proven me wrong ?? Howsabout YOU ??? More at www.angelfire.com/rnb/fairhaven/specs.html John LaMuth www.ethicalvalues.com agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Goedel machines (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment))
Abram Demski <[EMAIL PROTECTED]> wrote: >Matt, > >What is your opinion on Goedel machines? > > http://www.idsia.ch/~juergen/goedelmachine.html Thanks for the link. If I understand correctly, this is a form of bounded RSI, so it could not lead to a singularity. A Goedel machine is functionally equivalent to AIXI^tl in that it finds the optimal reinforcement learning solution given a fixed environment and utility function. The difference is that AIXI^tl does a brute force search of all machines up to length l for time t each, so it run in O(t 2^l) time. A Goedel machine achieves the same result more efficiently through a series of self improvments by proving that each proposed modification (including modifications to its own proof search code) is a actual improvement. It does this by using an instruction set such that it is impossible to construct incorrect proof verification code. What I am looking for is unbounded RSI capable of increasing intelligence. A Goedel machine doesn't do this because once it finds a solution, it stops. This is the same problem as a chess playing program that plays randomly modified copies of itself in death matches. At some point, it completely solves the chess problem and stops improving. Ideally we should use a scalable test for intelligence such as Legg and Hutter's universal intelligence, which measures expected accumulated reward over a Solomonoff distribution of environments (random programs). We can't compute this exactly because it requires testing an infinite number of environments, but we can approximate it to arbitrary precision by randomly sampling environments. RSI would require a series of increasingly complex test environments because otherwise there is an exact solution such that RSI would stop once found. For any environment with Kolmogorov complexity l, and agent can guess all environments up to length l. But this means that RSI cannot be implemented by a Turing machine because a parent with complexity l cannot test its children because it cannot create environments with complexity greater than l. RSI would be possible with a true source of randomness. A parent could create arbitrarily complex environments by flipping a coin. In practice, we usually ignore the difference between pseudo-random sources and true random sources. But in the context of Turing machines that can execute exponential complexity algorithms efficiently, we can't do this because the child could easily guess the parent's generator, which has low complexity. One could argue that the real universe does have true random sources, such as quantum mechanics. I am not convinced. The universe does have a definite quantum state, but it is not possible to know it because a memory within the universe cannot have more information than the universe. Therefore, any theory of physics must appear random. -- Matt Mahoney, [EMAIL PROTECTED] - Original Message From: Abram Demski <[EMAIL PROTECTED]> To: agi@v2.listbox.com Sent: Monday, August 25, 2008 3:30:59 PM Subject: Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment) Matt, What is your opinion on Goedel machines? http://www.idsia.ch/~juergen/goedelmachine.html --Abram On Sun, Aug 24, 2008 at 5:46 PM, Matt Mahoney <[EMAIL PROTECTED]> wrote: > Eric Burton <[EMAIL PROTECTED]> wrote: > > >>>These have profound impacts on AGI design. First, AIXI is (provably) not >>>computable, >>>which means there is no easy shortcut to AGI. Second, universal intelligence >>>is not >>>computable because it requires testing in an infinite number of >>>environments. Since >>>there is no other well accepted test of intelligence above human level, it >>>casts doubt on >>>the main premise of the singularity: that if humans can create agents with >>>greater than >>>human intelligence, then so can they. >> >>I don't know for sure that these statements logically follow from one >>another. > > They don't. I cannot prove that there is no non-evolutionary model of > recursive self improvement (RSI). Nor can I prove that there is. But it is a > question we need to answer before an evolutionary model becomes technically > feasible, because an evolutionary model is definitely unfriendly. > >>Higher intelligence bootstrapping itself has already been proven on >>Earth. Presumably it can happen in a simulation space as well, right? > > If you mean the evolution of humans, that is not an example of RSI. One > requirement of friendly AI is that an AI cannot alter its human-designed > goals. (Another is that we get the goals right, which is unsolved). However, > in an evolutionary environment, the parents do not get to choose the goals of > their
AGI goals (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment))
An AGI will not design its goals. It is up to humans to define the goals of an
AGI, so that it will do what we want it to do.
Unfortunately, this is a problem. We may or may not be successful in
programming the goals of AGI to satisfy human goals. If we are not successful,
then AGI will be useless at best and dangerous at worst. If we are successful,
then we are doomed because human goals evolved in a primitive environment to
maximize reproductive success and not in an environment where advanced
technology can give us whatever we want. AGI will allow us to connect our
brains to simulated worlds with magic genies, or worse, allow us to directly
reprogram our brains to alter our memories, goals, and thought processes. All
rational goal-seeking agents must have a mental state of maximum utility where
any thought or perception would be unpleasant because it would result in a
different state.
-- Matt Mahoney, [EMAIL PROTECTED]
- Original Message
From: Valentina Poletti <[EMAIL PROTECTED]>
To: agi@v2.listbox.com
Sent: Tuesday, August 26, 2008 11:34:56 AM
Subject: Re: Information theoretic approaches to AGI (was Re: [agi] The
Necessity of Embodiment)
Thanks very much for the info. I found those articles very interesting.
Actually though this is not quite what I had in mind with the term
information-theoretic approach. I wasn't very specific, my bad. What I am
looking for is a a theory behind the actual R itself. These approaches
(correnct me if I'm wrong) give an r-function for granted and work from that.
In real life that is not the case though. What I'm looking for is how the AGI
will create that function. Because the AGI is created by humans, some sort of
direction will be given by the humans creating them. What kind of direction, in
mathematical terms, is my question. In other words I'm looking for a way to
mathematically define how the AGI will mathematically define its goals.
Valentina
On 8/23/08, Matt Mahoney <[EMAIL PROTECTED]> wrote:
Valentina Poletti <[EMAIL PROTECTED]> wrote:
> I was wondering why no-one had brought up the information-theoretic aspect of
> this yet.
It has been studied. For example, Hutter proved that the optimal strategy of a
rational goal seeking agent in an unknown computable environment is AIXI: to
guess that the environment is simulated by the shortest program consistent with
observation so far [1]. Legg and Hutter also propose as a measure of universal
intelligence the expected reward over a Solomonoff distribution of environments
[2].
These have profound impacts on AGI design. First, AIXI is (provably) not
computable, which means there is no easy shortcut to AGI. Second, universal
intelligence is not computable because it requires testing in an infinite
number of environments. Since there is no other well accepted test of
intelligence above human level, it casts doubt on the main premise of the
singularity: that if humans can create agents with greater than human
intelligence, then so can they.
Prediction is central to intelligence, as I argue in [3]. Legg proved in [4]
that there is no elegant theory of prediction. Predicting all environments up
to a given level of Kolmogorov complexity requires a predictor with at least
the same level of complexity. Furthermore, above a small level of complexity,
such predictors cannot be proven because of Godel incompleteness. Prediction
must therefore be an experimental science.
There is currently no software or mathematical model of non-evolutionary
recursive self improvement, even for very restricted or simple definitions of
intelligence. Without a model you don't have friendly AI; you have accelerated
evolution with AIs competing for resources.
References
1. Hutter, Marcus (2003), "A Gentle Introduction to The Universal Algorithmic
Agent {AIXI}",
in Artificial General Intelligence, B. Goertzel and C. Pennachin eds.,
Springer. http://www.idsia.ch/~marcus/ai/aixigentle.htm
2. Legg, Shane, and Marcus Hutter (2006),
A Formal Measure of Machine Intelligence, Proc. Annual machine
learning conference of Belgium and The Netherlands (Benelearn-2006).
Ghent, 2006. http://www.vetta.org/documents/ui_benelearn.pdf
3. http://cs.fit.edu/~mmahoney/compression/rationale.html
4. Legg, Shane, (2006), Is There an Elegant Universal Theory of Prediction?,
Technical Report IDSIA-12-06, IDSIA / USI-SUPSI,
Dalle Molle Institute for Artificial Intelligence, Galleria 2, 6928 Manno,
Switzerland.
http://www.vetta.org/documents/IDSIA-12-06-1.pdf
-- Matt Mahoney, [EMAIL PROTECTED]
---
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/
Modify Your Subscription: https://www.listbox.com/member/?&;
Powered by Listbox: http://www.listbox.com
--
A true friend stabs you in the front. - O. Wilde
Einstein once thought he
Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)
John, are any of your peer-reviewed papers online? I can't seem to find them... -- Matt Mahoney, [EMAIL PROTECTED] - Original Message From: John LaMuth <[EMAIL PROTECTED]> To: agi@v2.listbox.com Sent: Tuesday, August 26, 2008 2:35:10 AM Subject: Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment) Matt Below is a sampling of my peer reviewed conference presentations on my background ethical theory ... This should elevate me above the common "crackpot" # Talks * Presentation of a paper at ISSS 2000 (International Society for Systems Sciences) Conference in Toronto, Canada on various aspects of the new science of Powerplay Politics. · Toward a Science of Consciousness: TUCSON April 8–12, 2002 Tucson Convention Center, Tucson, Arizona-sponsored by the Center for Consciousness Studies-University of Arizona (poster presentation). · John presented a poster at the 8th International Tsukaba Bioethics Conference at Tsukaba, Japan on Feb. 15 to 17, 2003. · John has presented his paper – “The Communicational Factors Underlying the Mental Disorders” at the 2006 Annual Conf. of the Western Psychological Association at Palm Springs, CA Honors * Honors Diploma for Research in Biological Sciences (June 1977) - Univ. of Calif. Irvine. * John is a member of the APA and the American Philosophical Association. LaMuth, J. E. (1977). The Development of the Forebrain as an Elementary Function of the Parameters of Input Specificity and Phylogenetic Age. J. U-grad Rsch: Bio. Sci. U. C. Irvine. (6): 274-294. LaMuth, J. E. (2000). A Holistic Model of Ethical Behavior Based Upon a Metaperspectival Hierarchy of the Traditional Groupings of Virtue, Values, & Ideals. Proceedings of the 44th Annual World Congress for the Int. Society for the Systems Sciences – Toronto. LaMuth, J. E. (2003). Inductive Inference Affective Language Analyzer Simulating AI. - US Patent # 6,587,846. LaMuth, J. E. (2004). Behavioral Foundations for the Behaviourome / Mind Mapping Project. Proceedings for the Eighth International Tsukuba Bioethics Roundtable,Tsukuba, Japan. LaMuth, J. E. (2005). A Diagnostic Classification of the Emotions: A Three-Digit Coding System for Affective Language. Lucerne Valley: Fairhaven. LaMuth, J. E. (2007). Inductive Inference Affective Language Analyzer Simulating Transitional AI. - US Patent # 7,236,963. ** Although I currently have no working model, I am collaborating on a working prototype. I was responding to your challenge for ...an example of a mathematical, software, biological, or physical example of RSI, or at least a plausible argument that one could be created I feel I have proposed a plausible argument, and considering the great stakes involved concerning ethical safeguards for AI, an avenue worthy of critique ... More on this in the last half of ) www.angelfire.com/rnb/fairhaven/specs.html John LaMuth www.ethicalvalues.com - Original Message - From: Matt Mahoney To: agi@v2.listbox.com Sent: Monday, August 25, 2008 7:30 AM Subject: Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment) John, I have looked at your patent and various web pages. You list a lot of nice sounding ethical terms (honor, love, hope, peace, etc) but give no details on how to implement them. You have already admitted that you have no experimental results, haven't actually built anything, and have no other results such as refereed conference or journal papers describing your system. If I am wrong about this, please let me know. -- Matt Mahoney, [EMAIL PROTECTED] - Original Message From: John LaMuth <[EMAIL PROTECTED]> To: agi@v2.listbox.com Sent: Sunday, August 24, 2008 11:21:30 PM Subject: Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment) ----- Original Message - From: "Matt Mahoney" <[EMAIL PROTECTED]> To: Sent: Sunday, August 24, 2008 2:46 PM Subject: Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment) I have challenged this list as well as the singularity and SL4 lists to come up with an example of a mathematical, software, biological, or physical example of RSI, or at least a plausible argument that one could be created, and nobody has. To qualify, an agent has to modify itself or create a more intelligent copy of itself according to an intelligence test chosen by the original. The following are not examples of RSI: > > 1. Evolution of life, including humans. &
Re: AGI goals (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment))
See also http://wireheading.com/ -- Matt Mahoney, [EMAIL PROTECTED] - Original Message From: BillK <[EMAIL PROTECTED]> To: agi@v2.listbox.com Sent: Wednesday, August 27, 2008 4:50:56 PM Subject: Re: AGI goals (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)) On Wed, Aug 27, 2008 at 8:43 PM, Abram Demski wrote: > By the way, where does this term "wireheading" come from? I assume > from context that it simply means self-stimulation. > Science Fiction novels. <http://en.wikipedia.org/wiki/Wirehead> In Larry Niven's Known Space stories, a wirehead is someone who has been fitted with an electronic brain implant (called a "droud" in the stories) to stimulate the pleasure centers of their brain. In 2006, The Guardian reported that trials of Deep brain stimulation with electric current, via wires inserted into the brain, had successfully lifted the mood of depression sufferers.[1] This is exactly the method used by wireheads in the earlier Niven stories (such as the 'Gil the Arm' story Death By Ectasy). In the Shaper/Mechanist stories of Bruce Sterling, "wirehead" is the Mechanist term for a human who has given up corporeal existence and become an infomorph. -- BillK --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?&; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: AGI goals (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment))
Mark Waser <[EMAIL PROTECTED]> wrote: >> All rational goal-seeking agents must have a mental state of maximum utility >> where any thought or perception would be unpleasant because it would result in a different state. >I'd love to see you attempt to prove the above statement. >What if there are several states with utility equal to or very close to the >maximum? Then you will be indifferent as to whether you stay in one state or move between them. >What if the utility of the state decreases the longer that you are in it >(something that is *very* true of human beings)? If you are aware of the passage of time, then you are not staying in the same state. >What if uniqueness raises the utility of any new state sufficient that there will always be states that are better than the current state (since experiencing uniqueness normally improves fitness through learning, etc)? Then you are not rational because your utility function does not define a total order. If you prefer A to B and B to C and C to A, as in the case you described, then you can be exploited. If you are rational and you have a finite number of states, then there is at least one state for which there is no better state. The human brain is certainly finite, and has at most 2^(10^15) states. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: Goedel machines (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment))
Abram Demski <[EMAIL PROTECTED]> wrote: >First, I do not think it is terribly difficult to define a Goedel >machine that does not halt. It interacts with its environment, and >there is some utility value attached to this interaction, and it >attempts to rewrite its code to maximize this utility. It's not that the machine halts, but that it makes no further improvements once the best solution is found. This might not be a practical concern if the environment is very complex. However, I doubt that a Goedel machine could even be built. Legg showed [1] that Goedel incompleteness is ubiquitous. To paraphrase, beyond some low level of complexity, you can't prove anything. Perhaps this is the reason we have not (AFAIK) built a software model, even for very simple sets of axioms. If we resort to probabilistic evidence of improvement rather than proofs, then it is no longer a Goedel machine, and I think we would need experimental verification of RSI. Random modifications of code are much more likely to be harmful than helpful, so we would need to show that improvements could be detected with a very low false positive rate. 1. http://www.vetta.org/documents/IDSIA-12-06-1.pdf -- Matt Mahoney, [EMAIL PROTECTED] - Original Message From: Abram Demski <[EMAIL PROTECTED]> To: agi@v2.listbox.com Sent: Wednesday, August 27, 2008 11:40:24 AM Subject: Re: Goedel machines (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)) Matt, Thanks for the reply. There are 3 reasons that I can think of for calling Goedel machines "bounded": 1. As you assert, once a solution is found, it stops. 2. It will be on a finite computer, so it will eventually reach the one best version of itself that it can reach. 3. It can only make provably correct steps, which is very limiting thanks to Godel's incompleteness theorem. I'll try to argue that each of these limits can be overcome in principle, and we'll see if the result satisfies your RSI criteria. First, I do not think it is terribly difficult to define a Goedel machine that does not halt. It interacts with its environment, and there is some utility value attached to this interaction, and it attempts to rewrite its code to maximize this utility. The second and third need to be tackled together, because the main reason that a Goedel machine can't improve its own hardware is because there is uncertainty involved, so it would never be provably better. There is always some chance of hardware malfunction. So, I think it is necessary to grant the possibility of modifications that are merely very probably correct. Once this is done, 2 and 3 fall fairly easily, assuming that the machine begins life with a good probabilistic learning system. That is a big assumption, but we can grant it for the moment I think? For the sake of concreteness, let's say that the utility value is some (probably very complex) attempt to logically describe Eliezer-style Friendliness, and that the probabilistic learning system is an approximation of AIXI (which the system will improve over time along with everything else). (These two choices don't reflect my personal tastes, they are just examples.) By tweaking the allowances the system makes, we might either have a slow self-improver that is, say, 99.999% probable to only improve itself in the next 100 years, or a faster self-improver that is 50% guaranteed. Does this satisfy your criteria? On Wed, Aug 27, 2008 at 9:14 AM, Matt Mahoney <[EMAIL PROTECTED]> wrote: > Abram Demski <[EMAIL PROTECTED]> wrote: >>Matt, >> >>What is your opinion on Goedel machines? >> >> http://www.idsia.ch/~juergen/goedelmachine.html > > > Thanks for the link. If I understand correctly, this is a form of bounded > RSI, so it could not lead to a singularity. A Goedel machine is functionally > equivalent to AIXI^tl in that it finds the optimal reinforcement learning > solution given a fixed environment and utility function. The difference is > that AIXI^tl does a brute force search of all machines up to length l for > time t each, so it run in O(t 2^l) time. A Goedel machine achieves the same > result more efficiently through a series of self improvments by proving that > each proposed modification (including modifications to its own proof search > code) is a actual improvement. It does this by using an instruction set such > that it is impossible to construct incorrect proof verification code. > > What I am looking for is unbounded RSI capable of increasing intelligence. A > Goedel machine doesn't do this because once it finds a solution, it stops. > This is the same problem as a chess playing program that plays randomly > modified copies of itself in death matches. At some point, it completely > solves the chess problem and sto
Re: AGI goals (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment))
Mark Waser <[EMAIL PROTECTED]> wrote: >>>What if the utility of the state decreases the longer that you are in it >>>(something that is *very* true of human >> beings)? >> If you are aware of the passage of time, then you are not staying in the >> same state. > >I have to laugh. So you agree that all your arguments don't apply to >anything that is aware of the passage of time? That makes them really >useful, doesn't it. No, the state of ultimate bliss that you, I, and all other rational, goal seeking agents seek is a mental state in which nothing perceptible happens. Without thought or sensation, you would be unaware of the passage of time, or of anything else. If you are aware of time then you are either not in this state yet, or are leaving it. You may say that is not what you want, but only because you are unaware of the possibilities of reprogramming your brain. It is like being opposed to drugs or wireheading. Once you experience it, you can't resist. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: AGI goals (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment))
Goals and motives are the same thing, in the sense that I mean them. We want the AGI to want to do what we want it to do. >Failure is an extreme danger, but it's not only failure to design safely >that's a danger. Failure to design a successful AGI at all could be >nearly as great a danger. Society has become too complex to be safely >managed by the current approaches...and things aren't getting any simpler. No, technology is the source of complexity, not the cure for it. But that is what we want. Life, health, happiness, freedom from work. AGI will cost $1 quadrillion to build, but we will build it because it is worth that much. And then it will kill us, not against our will, but because we want to live in simulated worlds with magic genies. -- Matt Mahoney, [EMAIL PROTECTED] - Original Message From: Charles Hixson <[EMAIL PROTECTED]> To: agi@v2.listbox.com Sent: Wednesday, August 27, 2008 7:16:53 PM Subject: Re: AGI goals (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)) Matt Mahoney wrote: > An AGI will not design its goals. It is up to humans to define the > goals of an AGI, so that it will do what we want it to do. Are you certain that this is the optimal approach? To me it seems more promising to design the motives, and to allow the AGI to design it's own goals to satisfy those motives. This provides less fine grained control over the AGI, but I feel that a fine-grained control would be counter-productive. To me the difficulty is designing the motives of the AGI in such a way that they will facilitate human life, when they must be implanted in an AGI that currently has no concept of an external universe, much less any particular classes of inhabitant therein. The only (partial) solution that I've been able to come up with so far (i.e., identify, not design) is based around imprinting. This is fine for the first generation (probably, if everything is done properly), but it's not clear that it would be fine for the second generation et seq. For this reason RSI is very important. It allows all succeeding generations to be derived from the first by cloning, which would preserve the initial imprints. > > Unfortunately, this is a problem. We may or may not be successful in > programming the goals of AGI to satisfy human goals. If we are not > successful, ... unpleasant because it would result in a different state. > > -- Matt Mahoney, [EMAIL PROTECTED] > Failure is an extreme danger, but it's not only failure to design safely that's a danger. Failure to design a successful AGI at all could be nearly as great a danger. Society has become too complex to be safely managed by the current approaches...and things aren't getting any simpler. --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: AGI goals (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment))
Valentina Poletti <[EMAIL PROTECTED]> wrote: > Got ya, thanks for the clarification. That brings up another question. Why do > we want to make an AGI? I'm glad somebody is finally asking the right question, instead of skipping over the specification to the design phase. It would avoid a lot of philosophical discussions that result from people having different ideas of what AGI should do. AGI could replace all human labor, worth about US $2 to $5 quadrillion over the next 30 years. We should expect the cost to be of this magnitude, given that having it sooner is better than waiting. I think AGI will be immensely complex, on the order of 10^18 bits, decentralized, competitive, with distributed ownership, like today's internet but smarter. It will converse with you fluently but know too much to pass the Turing test. We will be totally dependent on it. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: AGI goals (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment))
Nobody wants to enter a mental state where thinking and awareness are unpleasant, at least when I describe it that way. My point is that having everything you want is not the utopia that many people think it is. But it is where we are headed. -- Matt Mahoney, [EMAIL PROTECTED] - Original Message From: Mark Waser <[EMAIL PROTECTED]> To: agi@v2.listbox.com Sent: Thursday, August 28, 2008 9:18:05 AM Subject: Re: AGI goals (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)) > No, the state of ultimate bliss that you, I, and all other rational, goal > seeking agents seek Your second statement copied below not withstanding, I *don't* seek ultimate bliss. > You may say that is not what you want, but only because you are unaware of > the possibilities of reprogramming your brain. It is like being opposed to > drugs or wireheading. Once you experience it, you can't resist. It is not what I want *NOW*. It may be that once my brain has been altered by experiencing it, I may want it *THEN* but that has no relevance to what I want and seek now. These statements are just sloppy reasoning . . . . --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: Goedel machines (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment))
I'm not trying to win any arguments, but I am trying to solve the problem of whether RSI is possible at all. It is an important question because it profoundly affects the path that a singularity would take, and what precautions we need to design into AGI. Without RSI, then a singularity has to be a (very fast) evolutionary process in which agents compete for computing resources. In this scenario, friendliness is stable only to the extent that it contributes to fitness and fails when the AGI no longer needs us. If RSI is possible, then there is the additional threat of a fast takeoff of the kind described by Good and Vinge (and step 5 of the OpenCog roadmap). Friendliness and ethics are algorithmically complex functions that have to be hard coded into the first self-improving agent, and I have little confidence that this will happen. An unfriendly agent is much easier to build, so is likely to be built first. I looked at Legg's paper again, and I don't believe it rules out Goedel machines. Legg first proved that any program that predicts all infinite sequences up to Kolmogorov complexity n must also have complexity n, and then proved that except for very small n, that such predictors cannot be proven to work. This is a different context than a Goedel machine, which only has to learn a specific environment, not a set of environments. I don't know if Legg's proof would apply to RSI sequences of increasingly complex environments. -- Matt Mahoney, [EMAIL PROTECTED] - Original Message From: Abram Demski <[EMAIL PROTECTED]> To: agi@v2.listbox.com Sent: Thursday, August 28, 2008 11:42:10 AM Subject: Re: Goedel machines (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)) PS-- I have thought of a weak argument: If a fact is not probabilistically learnable, then it is hard to see how it has much significance for an AI design. A non-learnable fact won't reliably change the performance of the AI, since if it did, it would be learnable. Furthermore, even if there were *some* important nonlearnable facts, it still seems like significant self-improvements could be made using only probabilistically learned facts. And, since the amount of time spent testing cases is a huge factor, RSI will not stop except due to limited memory, even in a relatively boring environment (unless the AI makes a rational decision to stop using resources on RSI since it has found a solution that is probably optimal). On Thu, Aug 28, 2008 at 11:25 AM, Abram Demski <[EMAIL PROTECTED]> wrote: > Matt, > > Ok, you have me, I admit defeat. > > I could only continue my argument if I could pin down what sorts of > facts need to be learned with high probability for RSI, and show > somehow that this set does not include unlearnable facts. Learnable > facts form a larger set than provable facts, since for example we can > probabilistically declare that a program never halts if we run it for > a while and it doesn't. But there are certain facts that are not even > probabilistically learnable, so until I can show that none of these > are absolutely essential to RSI, I concede. > > --Abram Demski > > On Wed, Aug 27, 2008 at 6:48 PM, Matt Mahoney <[EMAIL PROTECTED]> wrote: >> Abram Demski <[EMAIL PROTECTED]> wrote: >>>First, I do not think it is terribly difficult to define a Goedel >>>machine that does not halt. It interacts with its environment, and >>>there is some utility value attached to this interaction, and it >>>attempts to rewrite its code to maximize this utility. >> >> It's not that the machine halts, but that it makes no further improvements >> once the best solution is found. This might not be a practical concern if >> the environment is very complex. >> >> However, I doubt that a Goedel machine could even be built. Legg showed [1] >> that Goedel incompleteness is ubiquitous. To paraphrase, beyond some low >> level of complexity, you can't prove anything. Perhaps this is the reason we >> have not (AFAIK) built a software model, even for very simple sets of axioms. >> >> If we resort to probabilistic evidence of improvement rather than proofs, >> then it is no longer a Goedel machine, and I think we would need >> experimental verification of RSI. Random modifications of code are much more >> likely to be harmful than helpful, so we would need to show that >> improvements could be detected with a very low false positive rate. >> >> 1. http://www.vetta.org/documents/IDSIA-12-06-1.pdf >> >> >> -- Matt Mahoney, [EMAIL PROTECTED] >> >> >> >> - Original Message >> From: Abram Demski <[EMAIL PROTECTED]> >> To: agi@v2.listbox.com >>
RSI (was Re: Goedel machines (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)))
Here is Vernor Vinge's original essay on the singularity. http://mindstalk.net/vinge/vinge-sing.html The premise is that if humans can create agents with above human intelligence, then so can they. What I am questioning is whether agents at any intelligence level can do this. I don't believe that agents at any level can recognize higher intelligence, and therefore cannot test their creations. We rely on competition in an external environment to make fitness decisions. The parent isn't intelligent enough to make the correct choice. -- Matt Mahoney, [EMAIL PROTECTED] - Original Message From: Mike Tintner <[EMAIL PROTECTED]> To: agi@v2.listbox.com Sent: Thursday, August 28, 2008 7:00:07 PM Subject: Re: Goedel machines (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)) Matt:If RSI is possible, then there is the additional threat of a fast takeoff of the kind described by Good and Vinge Can we have an example of just one or two subject areas or domains where a takeoff has been considered (by anyone) as possibly occurring, and what form such a takeoff might take? I hope the discussion of RSI is not entirely one of airy generalities, without any grounding in reality. --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: AGI goals (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment))
Group selection is not dead, just weaker than individual selection. Altruism in many species is evidence for its existence. http://en.wikipedia.org/wiki/Group_selection In any case, evolution of culture and ethics in humans is primarily memetic, not genetic. Taboos against nudity are nearly universal among cultures with language, yet unique to homo sapiens. You might believe that certain practices are intrinsically good or bad, not the result of group selection. Fine. That is how your beliefs are supposed to work. -- Matt Mahoney, [EMAIL PROTECTED] - Original Message From: Mark Waser <[EMAIL PROTECTED]> To: agi@v2.listbox.com Sent: Friday, August 29, 2008 1:13:43 PM Subject: Re: AGI goals (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)) "Group selection" (as used as the term of art in evolutionary biology) does not seem to be experimentally supported (and there have been a lot of recent experiments looking for such an effect). It would be nice if people could let the idea drop unless there is actually some proof for it other than "it seems to make sense that . . . . " - Original Message - From: "Eric Burton" <[EMAIL PROTECTED]> To: Sent: Friday, August 29, 2008 12:56 PM Subject: **SPAM** Re: AGI goals (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)) >I remember Richard Dawkins saying that group selection is a lie. Maybe > we shoud look past it now? It seems like a problem. > > On 8/29/08, Mark Waser <[EMAIL PROTECTED]> wrote: >>>> OK. How about this . . . . Ethics is that behavior that, >>>> when shown by you, >>>> makes me believe that I should facilitate your survival. >>>> Obviously, it is >>>> then to your (evolutionary) benefit to behave ethically. >>> >>> Ethics can't be explained simply by examining interactions between >>> individuals. It's an emergent dynamic that requires explanation at the >>> group level. It's a set of culture-wide rules and taboos - how did they >>> get there? >> >> I wasn't explaining ethics with that statement. I was identifying how >> "evolution operates in social groups in such a way that I can derive >> ethics" >> (in direct response to your question). >> >> Ethics is a system. The *definition of ethical behavior* for a given >> group >> is "an emergent dynamic that requires explanation at the group level" >> because it includes what the group believes and values -- but ethics (the >> system) does not require belief history (except insofar as it affects >> current belief). History, circumstances, and understanding what a >> culture >> has the rules and taboos that they have is certainly useful for deriving >> more effective rules and taboos -- but it doesn't alter the underlying >> system which is quite simple . . . . being perceived as helpful generally >> improves your survival chances, being perceived as harmful generally >> decreases your survival chances (unless you are able to overpower the >> effect). >> >>> Really? I must be out of date too then, since I agree with his >>> explanation >>> >>> of ethics. I haven't read Hauser yet though, so maybe you're right. >> >> The specific phrase you cited was "human collectives with certain taboos >> make the group as a whole more likely to persist". The correct term of >> art >> for this is "group selection" and it has pretty much *NOT* been supported >> by >> scientific evidence and has fallen out of favor. >> >> Matt also tends to conflate a number of ideas which should be separate >> which >> you seem to be doing as well. There need to be distinctions between >> ethical >> systems, ethical rules, cultural variables, and evaluations of ethical >> behavior within a specific cultural context (i.e. the results of the >> system >> given certain rules -- which at the first-level seem to be reasonably >> standard -- with certain cultural variables as input). Hauser's work >> identifies some of the common first-level rules and how cultural >> variables >> affect the results of those rules (and the derivation of secondary >> rules). >> It's good detailed, experiment-based stuff rather than the vague >> hand-waving >> that you're getting from armchair philosophers. >> >>> I fail to see how your above explanation is anything but an elaboration >>> of >>> >>> the idea that ethics is due to group s
Re: RSI (was Re: Goedel machines (was Re: Information theoretic approaches to AGI (was Re: [agi] The Necessity of Embodiment)))
It seems that the debate over recursive self improvement depends on what you mean by "improvement". If you define improvement as intelligence as defined by the Turing test, then RSI is not possible because the Turing test does not test for superhuman intelligence. If you mean improvement as more memory, faster clock speed, more network bandwidth, etc., then yes, I think it is reasonable to expect Moore's law to continue after we are all uploaded. If you mean improvement in the sense of competitive fitness, then yes, I expect evolution to continue, perhaps very rapidly if it is based on a computing substrate other than DNA. Whether you can call it "self" improvement or whether the result is desirable is debatable. We are, after all, pondering the extinction of Homo Sapiens and replacing it with some unknown species, perhaps gray goo. Will the nanobots look back at this as an improvement, the way we view the extinction of Homo Erectus? My question is whether RSI is mathematically possible in the context of universal intelligence, i.e. expected reward or prediction accuracy over a Solomonoff distribution of computable environments. I believe it is possible for Turing machines if and only if they have access to true random sources so that each generation can create successively more complex test environments to evaluate their offspring. But this is troubling because in practice we can construct pseudo-random sources that are nearly indistinguishable from truly random in polynomial time (but none that are *provably* so). -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Goedel machines ..PS
Mike Tintner wrote: >You may have noticed that AGI-ers are staggeringly resistant to learning new >domains. Remember you are dealing with human brains. You can only write into long term memory at a rate of 2 bits per second. :-) AGI spans just about every field of science, from ethics to quantum mechanics, child development to algorithmic information theory, genetics to economics. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] Preservation of goals in strongly self-modifying AI systems
In response to Ben's paper on goal preservation, I think that identifying attractors or fixed points requires that we identify sources of goal drift. Here are some: - Information loss - Software errors - Deliberate modification - Modification through learning - Evolution - Noise Information loss is caused by irreversible operations, for example, the assignment statement. The algorithmic complexity of a machine's state (such as a set of self improving processes) cannot increase over time without input. This suggests that the state of having no goals is an attractor. Software errors: an agent may have the goal of preserving its goals, but may be unsuccessful due to programming errors. Humans make errors, so there is no reason to believe that super intelligent beings will be different. Software verification reduces to the halting problem. This suggests that hard to detect bugs will accumulate. Deliberate modification: friendliness is algorithmically complex, so it will require human maintenance. Many situations will arise that were not planned for. For example, is it ethical to copy a human and kill the original? Is it ethical to turn off or simulate pain in an AI that has some human traits, but is not entirely human? Is it ethical to allow wireheading? This suggests a dynamic toward increasing algorithmic complexity (like our legal system). Modifications through learning: we could allow the AI to make ethical decisions for us on the premise that it is smarter than humans, and therefore will make more intelligent decisions than we could. This means we are also not intelligent enough to predict where this dynamic will go. Evolution: some goals may be harmful to the AI. Selective pressure will favor rapid reproduction and acquisition of computing resources. Noise results from copying errors. In living organisms there is an equilibrium between information gained through selection (about log n bits per birth/death, where n is the average number of children produced) and information loss through mutation that limits the algorithmic complexity of the genome of higher organisms (like humans) to around 10^7 bits. There may be other forces I missed. To make this concrete, imagine a simple self improving system, such as a data compressor that "wants" to output smaller files, or a linear regression program that "wants" to reduce RMS error. Describe an environment where the program could rewrite itself (or modify its copies). How would its goals drift, and where would they settle? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] Re: Preservation of goals in strongly self-modifying AI systems
I'm afraid I have to agree with Richard. "Goal" is not well defined if you have to point to a human to say what it means. Does a chess playing program have a goal of winning, or does it just execute a program? Does a calculator have a goal of solving arithmetic problems? Does autobliss ( http://www.mattmahoney.net/autobliss.txt ) have a goal of maximizing a reward signal? What property of a program distinguishes whether or not it has a goal? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Computation as an explanation of the universe (was Re: [agi] Recursive self-change: some definitions)
I think that computation is not so much a metaphor for understanding the universe as it is an explanation. If you enumerate all possible Turing machines, thus enumerating all possible laws of physics, then some of those universes will have the right conditions for the evolution of intelligent life. If neutrons were slightly heavier than they actually are (relative to protons), then stars could not sustain fusion. If they were slightly lighter, then they would be stable and we would have no elements. Because of gravity, the speed of light, Planck's constant, the quantization of electric charge, and the finite age of the universe, the universe has a finite length description, and is therefore computable. The Bekenstein bound of the Hubble radius is 2.91 x 10^122 bits. Any computer within a finite universe must have less memory than it, and therefore cannot simulate it except by using an approximate (probabilistic) model. One such model is quantum mechanics. For the same reason, an intelligent agent (which must be Turing computable if the universe is) cannot model itself, except probabilistically as an approximation. Thus, we cannot predict what we will think without actually thinking it. This property makes our own intelligence seem mysterious. An explanation is only useful if it makes predictions, and it does. If the universe were not Turing computable, then Solomonoff induction and AIXI as ideal models of prediction and intelligence would not be applicable to the real world. Yet we have Occam's Razor and find in practice that all successful machine learning algorithms use algorithmically simple hypothesis sets. -- Matt Mahoney, [EMAIL PROTECTED] --- On Wed, 9/3/08, Terren Suydam <[EMAIL PROTECTED]> wrote: From: Terren Suydam <[EMAIL PROTECTED]> Subject: Re: [agi] Recursive self-change: some definitions To: agi@v2.listbox.com Date: Wednesday, September 3, 2008, 4:17 PM Hi Ben, My own feeling is that computation is just the latest in a series of technical metaphors that we apply in service of understanding how the universe works. Like the others before it, it captures some valuable aspects and leaves out others. It leaves me wondering: what future metaphors will we apply to the universe, ourselves, etc., that will make computation-as-metaphor seem as quaint as the old clockworks analogies? I believe that computation is important in that it can help us simulate intelligence, but intelligence itself is not simply computation (or if it is, it's in a way that requires us to transcend our current notions of computation). Note that I'm not suggesting anything mystical or dualistic at all, just offering the possibility that we can find still greater metaphors for how intelligence works. Either way though, I'm very interested in the results of your work - at worst, it will shed some needed light on the subject. At best... well, you know that part. :-] Terren --- On Tue, 9/2/08, Ben Goertzel <[EMAIL PROTECTED]> wrote: From: Ben Goertzel <[EMAIL PROTECTED]> Subject: Re: [agi] Recursive self-change: some definitions To: agi@v2.listbox.com Date: Tuesday, September 2, 2008, 4:50 PM On Tue, Sep 2, 2008 at 4:43 PM, Eric Burton <[EMAIL PROTECTED]> wrote: I really see a number of algorithmic breakthroughs as necessary for the development of strong general AI I hear that a lot, yet I never hear any convincing arguments in that regard... So, hypothetically (and I hope not insultingly), I tend to view this as a kind of unconscious overestimation of the awesomeness of our own species ... we feel intuitively like we're doing SOMETHING so cool in our brains, it couldn't possibly be emulated or superseded by mere algorithms like the ones computer scientists have developed so far ;-) ben --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: Computation as an explanation of the universe (was Re: [agi] Recursive self-change: some definitions)
--- On Wed, 9/3/08, Abram Demski <[EMAIL PROTECTED]> wrote: > From: Abram Demski <[EMAIL PROTECTED]> > Subject: Re: Computation as an explanation of the universe (was Re: [agi] > Recursive self-change: some definitions) > To: agi@v2.listbox.com > Date: Wednesday, September 3, 2008, 7:35 PM > Matt, I have several objections. > > First, as I understand it, your statement about the > universe having a > finite description length only applies to the *observable* > universe, > not the universe as a whole. The hubble radius expands at > the speed of > light as more light reaches us, meaning that the observable > universe > has a longer description length every day. So it does not > seem very > relevant to say that the description length is finite. > > The universe as a whole (observable and not-observable) > *could* be > finite, but we don't know one way or the other so far > as I am aware. OK, then the observable universe has a finite description length. We don't need to describe anything else to model it, so by "universe" I mean only the observable part. > > Second, I do not agree with your reason for saying that > physics is > necessarily probabilistic. It seems possible to have a > completely > deterministic physics, which merely suffers from a lack of > information > and computation ability. Imagine if the universe happened > to follow > Newtonian physics, with atoms being little billiard balls. > The > situation is deterministic, if only we knew the starting > state of the > universe and had large enough computers to approximate the > differential equations to arbitrary accuracy. I am saying that the universe *is* deterministic. It has a definite quantum state, but we would need about 10^122 bits of memory to describe it. Since we can't do that, we have to resort to approximate models like quantum mechanics. I believe there is a simpler description. First, the description length is increasing with the square of the age of the universe, since it is proportional to area. So it must have been very small at one time. Second, the most efficient way to enumerate all possible universes would be to run each B-bit machine for 2^B steps, starting with B = 0, 1, 2... until intelligent life is found. For our universe, B ~ 407. You could reasonably argue that the algorithmic complexity of the free parameters of string theory and general relativity is of this magnitude. I believe that Wolfram also argued that the (observable) universe is a few lines of code. But even if we discover this program it does not mean we could model the universe deterministically. We would need a computer larger than the universe to do so. > Third, this is nitpicking, but I also am not sure about the > argument > that we cannot predict our thoughts. It seems formally > possible that a > system could predict itself. The system would need to be > compressible, > so that a model of itself could fit inside the whole. I > could be wrong > here, feel free to show me that I am. Anyway, the same > objection also > applies back to the necessity of probabilistic physics: is > it really > impossible for beings within a universe to have an accurate > compressed > model of the entire universe? (Similarly, if we have such a > model, > could we use it to run a simulation of the entire universe? > This seems > much less possible.) There is a simple argument using information theory. Every system S has a Kolmogorov complexity K(S), which is the smallest size that you can compress a description of S to. A model of S must also have complexity K(S). However, this leaves no space for S to model itself. In particular, if all of S's memory is used to describe its model, there is no memory left over to store any results of the simulation. > > --Abram > > > On Wed, Sep 3, 2008 at 6:45 PM, Matt Mahoney > <[EMAIL PROTECTED]> wrote: > > I think that computation is not so much a metaphor for > understanding the universe as it is an explanation. If you > enumerate all possible Turing machines, thus enumerating all > possible laws of physics, then some of those universes will > have the right conditions for the evolution of intelligent > life. If neutrons were slightly heavier than they actually > are (relative to protons), then stars could not sustain > fusion. If they were slightly lighter, then they would be > stable and we would have no elements. > > > > Because of gravity, the speed of light, Planck's > constant, the quantization of electric charge, and the > finite age of the universe, the universe has a finite length > description, and is therefore computable. The Bekenstein > bound of the Hubble radius is 2.91 x 10^122 bits. Any > compute
Re: [agi] What is Friendly AI?
--- On Wed, 9/3/08, Steve Richfield <[EMAIL PROTECTED]> wrote: >OK, lets take a concrete example: The Middle East situation, >and ask our infinitely intelligent AGI what to do about it. OK, lets take a concrete example of friendly AI, such as competitive message routing ( http://www.mattmahoney.net/agi.html ). CMR has an algorithmically complex definition of "friendly". The behavior of billions of peers (narrow-AI specialists) are controlled by their human owners who have an economic incentive to trade cooperatively and provide useful information. Nevertheless, the environment is hostile, so a large fraction (probably most) of CPU cycles and knowledge will probably be used to defend against attacks, primarily spam. CMR is friendly AGI because a lot of narrow-AI specialists that understand just enough natural language to do their jobs and know just a little about where to route other messages will result (I believe) in a system that is generally useful as a communication medium to humans. You would just enter any natural language message and it would get routed to anyone who cares, human or machine. So to answer your question, CMR would not solve the Middle East conflict. It is not designed to. That is for people to do. Forcing people to do anything is not friendly. CMR is friendly in the sense that a market is friendly. A market can sell weapons to both sides, but markets also reward cooperation. Countries that trade with each other have an incentive not to go to war. Likewise, the internet can be used to plan attacks and promote each sides' agenda, but also to make it easier for the two sides to communicate. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] What is Friendly AI?
--- On Wed, 9/3/08, Terren Suydam <[EMAIL PROTECTED]> wrote: > I'm talking about a situation where humans must interact > with the FAI without knowledge in advance about whether it > is Friendly or not. Is there a test we can devise to make > certain that it is? No. If an AI has godlike intelligence, then testing whether it is friendly would be like an ant proving that you won't step on it. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: Computation as an explanation of the universe (was Re: [agi] Recursive self-change: some definitions)
To clarify what I mean by "observable universe", I am including any part that could be observed in the future, and therefore must be modeled to make accurate predictions. For example, if our universe is computed by one of an enumeration of Turing machines, then the other enumerations are outside our observable universe. -- Matt Mahoney, [EMAIL PROTECTED] --- On Thu, 9/4/08, Abram Demski <[EMAIL PROTECTED]> wrote: > From: Abram Demski <[EMAIL PROTECTED]> > Subject: Re: Computation as an explanation of the universe (was Re: [agi] > Recursive self-change: some definitions) > To: agi@v2.listbox.com > Date: Thursday, September 4, 2008, 9:43 AM > > OK, then the observable universe has a finite > description length. We don't need to describe anything > else to model it, so by "universe" I mean only the > observable part. > > > > But, what good is it to only have finite description of the > observable > part, since new portions of the universe enter the > observable portion > continually? Physics cannot then be modeled as a computer > program, > because computer programs do not increase in Kolmogorov > complexity as > they run (except by a logarithmic term to count how long it > has been > running). > > > I am saying that the universe *is* deterministic. It > has a definite quantum state, but we would need about 10^122 > bits of memory to describe it. Since we can't do that, > we have to resort to approximate models like quantum > mechanics. > > > > Yes, I understood that you were suggesting a deterministic > universe. > What I'm saying is that it seems plausible for us to be > able to have > an accurate knowledge of that deterministic physics, > lacking only the > exact knowledge of particle locations et cetera. We would > be forced to > use probabilistic methods as you argue, but they would not > necessarily > be built into our physical theories; instead, our physical > theories > act as a deterministic function that is given probabilistic > input and > therefore yields probabilistic output. > > > I believe there is a simpler description. First, the > description length is increasing with the square of the age > of the universe, since it is proportional to area. So it > must have been very small at one time. Second, the most > efficient way to enumerate all possible universes would be > to run each B-bit machine for 2^B steps, starting with B = > 0, 1, 2... until intelligent life is found. For our > universe, B ~ 407. You could reasonably argue that the > algorithmic complexity of the free parameters of string > theory and general relativity is of this magnitude. I > believe that Wolfram also argued that the (observable) > universe is a few lines of code. > > > > I really do not understand your willingness to restrict > "universe" to > "observable universe". The description length of > the observable > universe was very small at one time because at that time > none of the > basic stuffs of the universe had yet interacted, so by > definition the > description length of the observable universe for each > basic entity is > just the description length of that entity. As time moves > forward, the > entities interact and the description lengths of their > observable > universes increase. Similarly, today, one might say that > the > observable universe for each person is slightly different, > and indeed > the universe observable from my right hand would be > slightly different > then the one observable from my left. They could have > differing > description lengths. > > In short, I think you really want to apply your argument to > the > "actual" universe, not merely observable > subsets... or if you don't, > you should, because otherwise it seems like a very strange > argument. > > > But even if we discover this program it does not mean > we could model the universe deterministically. We would need > a computer larger than the universe to do so. > > Agreed... partly thanks to your argument below. > > > There is a simple argument using information theory. > Every system S has a Kolmogorov complexity K(S), which is > the smallest size that you can compress a description of S > to. A model of S must also have complexity K(S). However, > this leaves no space for S to model itself. In particular, > if all of S's memory is used to describe its model, > there is no memory left over to store any results of the > simulation. > > Point conceded. > > > --Abram > > > --- > agi > Archives: https://www.listbox.com/member/
Real vs. simulated environments (was Re: [agi] draft for comment.. P.S.)
--- On Thu, 9/4/08, Valentina Poletti <[EMAIL PROTECTED]> wrote: >Ppl like Ben argue that the concept/engineering aspect of intelligence is >independent of the type of environment. That is, given you understand how >to make it in a virtual environment you can then tarnspose that concept >into a real environment more safely. > >Some other ppl on the other hand believe intelligence is a property of >humans only. So you have to simulate every detail about humans to get >that intelligence. I'd say that among the two approaches the first one >(Ben's) is safer and more realistic. The issue is not what is intelligence, but what do you want to create? In order for machines to do more work for us, they may need language and vision, which we associate with human intelligence. But building artificial humans is not necessarily useful. We already know how to create humans, and we are doing so at an unsustainable rate. I suggest that instead of the imitation game (Turing test) for AI, we should use a preference test. If you prefer to talk to a machine vs. a human, then the machine passes the test. Prediction is central to intelligence. If you can predict a text stream, then for any question Q and any answer A, you can compute the probability distribution P(A|Q) = P(QA)/P(Q). This passes the Turing test. More importantly, it allows you to output max_A P(QA), the most likely answer from a group of humans. This passes the preference test because a group is usually more accurate than any individual member. (It may fail a Turing test for giving too few wrong answers, a problem Turing was aware of in 1950 when he gave an example of a computer incorrectly answering an arithmetic problem). Text compression is equivalent to AI because we have already solved the coding problem. Given P(x) for string x, we know how to optimally and efficiently code x in log_2(1/P(x)) bits (e.g. arithmetic coding). Text compression has an advantage over the Turing or preference tests in that that incremental progress in modeling can be measured precisely and the test is repeatable and verifiable. If I want to test a text compressor, it is important to use real data (human generated text) rather than simulated data, i.e. text generated by a program. Otherwise, I know there is a concise code for the input data, which is the program that generated it. When you don't understand the source distribution (i.e. the human brain), the problem is much harder, and you have a legitimate test. I understand that Ben is developing AI for virtual worlds. This might produce interesting results, but I wouldn't call it AGI. The value of AGI is on the order of US $1 quadrillion. It is a global economic system running on a smarter internet. I believe that any attempt to develop AGI on a budget of $1 million or $1 billion or $1 trillion is just wishful thinking. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: Computation as an explanation of the universe (was Re: [agi] Recursive self-change: some definitions)
--- On Thu, 9/4/08, Abram Demski <[EMAIL PROTECTED]> wrote: > So, my only remaining objection is that while the universe > *could* be > computable, it seems unwise to me to totally rule out the > alternative. You're right. We cannot prove that the universe is computable. We have evidence like Occam's Razor (if the universe is computable, then algorithmically simple models are to be preferred), but that is not proof. At one time our models of physics were not computable. Then we discovered atoms, quantization of electric charge, general relativity (which bounds density and velocity), the big bang (history is finite) and quantum mechanics. Our models would still not be computable (requiring infinite description length) if any one of these events did not occur. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] draft for comment
--- On Wed, 9/3/08, Pei Wang <[EMAIL PROTECTED]> wrote: > TITLE: Embodiment: Who does not have a body? > > AUTHOR: Pei Wang > > ABSTRACT: In the context of AI, ``embodiment'' > should not be > interpreted as ``giving the system a body'', but as > ``adapting to the > system's experience''. Therefore, being a robot > is neither a > sufficient condition nor a necessary condition of being > embodied. What > really matters is the assumption about the environment for > which the > system is designed. > > URL: http://nars.wang.googlepages.com/wang.embodiment.pdf The paper seems to argue that embodiment applies to any system with inputs and outputs, and therefore all AI systems are embodied. However, there are important differences between symbolic systems like NARS and systems with external sensors such as robots and humans. The latter are analog, e.g. the light intensity of a particular point in the visual field, or the position of a joint in an arm. In humans, there is a tremendous amount of data reduction from the senses, from 137 million rods and cones in each eye each firing up to 300 pulses per second, down to 2 bits per second by the time our high level visual perceptions reach long term memory. AI systems have traditionally avoided this type of processing because they lacked the necessary CPU power. IMHO this has resulted in biologically implausible symbolic language models with only a small number of connections between concepts, rather than the tens of thousands of connections per neuron. Another aspect of embodiment (as the term is commonly used), is the false appearance of intelligence. We associate intelligence with humans, given that there are no other examples. So giving an AI a face or a robotic body modeled after a human can bias people to believe there is more intelligence than is actually present. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] open models, closed models, priors
In a closed model, every statement is either true or false. In an open model, every statement is either true or uncertain. In reality, all statements are uncertain, but we have a means to assign them probabilities (not necessarily accurate probabilities). A closed model is unrealistic, but an open model is even more unrealistic because you lack a means of assigning likelihoods to statements like "the sun will rise tomorrow" or "the world will end tomorrow". You absolutely must have a means of guessing probabilities to do anything at all in the real world. -- Matt Mahoney, [EMAIL PROTECTED] --- On Thu, 9/4/08, Abram Demski <[EMAIL PROTECTED]> wrote: > From: Abram Demski <[EMAIL PROTECTED]> > Subject: [agi] open models, closed models, priors > To: agi@v2.listbox.com > Date: Thursday, September 4, 2008, 2:19 PM > A closed model is one that is interpreted as representing > all truths > about that which is modeled. An open model is instead > interpreted as > making a specific set of assertions, and leaving the rest > undecided. > Formally, we might say that a closed model is interpreted > to include > all of the truths, so that any other statements are false. > This is > also known as the closed-world assumption. > > A typical example of an open model is a set of statements > in predicate > logic. This could be changed to a closed model simply by > applying the > closed-world assumption. A possibly more typical example of > a > closed-world model is a computer program that outputs the > data so far > (and predicts specific future output), as in Solomonoff > induction. > > These two types of model are very different! One important > difference > is that we can simply *add* to an open model if we need to > account for > new data, while we must always *modify* a closed model if > we want to > account for more information. > > The key difference I want to ask about here is: a > length-based > bayesian prior seems to apply well to closed models, but > not so well > to open models. > > First, such priors are generally supposed to apply to > entire joint > states; in other words, probability theory itself (and in > particular > bayesian learning) is built with an assumption of an > underlying space > of closed models, not open ones. > > Second, an open model always has room for additional stuff > somewhere > else in the universe, unobserved by the agent. This > suggests that, > made probabilistic, open models would generally predict > universes with > infinite description length. Whatever information was > known, there > would be an infinite number of chances for other unknown > things to be > out there; so it seems as if the probability of *something* > more being > there would converge to 1. (This is not, however, > mathematically > necessary.) If so, then taking that other thing into > account, the same > argument would still suggest something *else* was out > there, and so > on; in other words, a probabilistic open-model-learner > would seem to > predict a universe with an infinite description length. > This does not > make it easy to apply the description length principle. > > I am not arguing that open models are a necessity for AI, > but I am > curious if anyone has ideas of how to handle this. I know > that Pei > Wang suggests abandoning standard probability in order to > learn open > models, for example. > > --Abram Demski > --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] open models, closed models, priors
I mean that you have to assign likelihoods to beliefs, even if the numbers are wrong. Logic systems where every statement is true or false simply are too brittle to scale beyond toy problems. Everything in life is uncertain, including the degree of uncertainty. That's why we use terms like "probably", "maybe", etc. instead of numbers. -- Matt Mahoney, [EMAIL PROTECTED] --- On Thu, 9/4/08, Mike Tintner <[EMAIL PROTECTED]> wrote: > From: Mike Tintner <[EMAIL PROTECTED]> > Subject: Re: [agi] open models, closed models, priors > To: agi@v2.listbox.com > Date: Thursday, September 4, 2008, 3:23 PM > Matt:You absolutely must have a means of guessing > probabilities to do > anything at all in the real world > > Do you mean mathematically? Estimating chances as roughly, > even if > provisionally, 0.70? If so, manifestly, that is untrue. > What are your > chances that you will get lucky tonight? Will an inability > to guess the > probability stop you trying? Most of the time, arguably, > we have to and do, > act on the basis of truly vague magnitudes - a > mathematically horrendously > rough sense of probability. Or just: "what the heck - > what's the worst that > can happen? Let's do it. And let's just pray it > works out." How precise a > sense of the probabilities attending his current decisions > does even a > professionally mathematical man like Bernanke have? > > Only AGI's in a virtual world can live with cosy, > mathematically calculable > "uncertainty." Living in the real world is as > Kauffman points out to a great > extent living with *mystery*. What are the maths of > mystery? Do you think > Ben has the least realistic idea of the probabilities > affecting his AGI > projects? That's not how most creative projects get > done, or life gets > lived. Quadrillions, Matt, schmazillions. > > > > > --- > agi > Archives: https://www.listbox.com/member/archive/303/=now > RSS Feed: https://www.listbox.com/member/archive/rss/303/ > Modify Your Subscription: > https://www.listbox.com/member/?&; > Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] open models, closed models, priors
--- On Thu, 9/4/08, Bryan Bishop <[EMAIL PROTECTED]> wrote: > On Thursday 04 September 2008, Matt Mahoney wrote: > > A closed model is unrealistic, but an open model is > even more > > unrealistic because you lack a means of assigning > likelihoods to > > statements like "the sun will rise tomorrow" > or "the world will end > > tomorrow". You absolutely must have a means of > guessing probabilities > > to do anything at all in the real world. > > I don't assign or guess probabilities and I seem to get > things done. > What gives? Yes you do. Every time you make a decision, you are assigning a higher probability of a good outcome to your choice than to the alternative. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Language modeling (was Re: [agi] draft for comment)
--- On Thu, 9/4/08, Pei Wang <[EMAIL PROTECTED]> wrote: > I guess you still see NARS as using model-theoretic > semantics, so you > call it "symbolic" and contrast it with system > with sensors. This is > not correct --- see > http://nars.wang.googlepages.com/wang.semantics.pdf and > http://nars.wang.googlepages.com/wang.AI_Misconceptions.pdf I mean NARS is symbolic in the sense that you write statements in Narsese like "raven -> bird <0.97, 0.92>" (probability=0.97, confidence=0.92). I realize that the meanings of "raven" and "bird" are determined by their relations to other symbols in the knowledge base and that the probability and confidence change with experience. But in practice you are still going to write statements like this because it is the easiest way to build the knowledge base. You aren't going to specify the brightness of millions of pixels in a vision system in Narsese, and there is no mechanism I am aware of to collect this knowledge from a natural language text corpus. There is no mechanism to add new symbols to the knowledge base through experience. You have to explicitly add them. > You have made this point on "CPU power" several > times, and I'm still > not convinced that the bottleneck of AI is hardware > capacity. Also, > there is no reason to believe an AGI must be designed in a > "biologically plausible" way. Natural language has evolved to be learnable on a massively parallel network of slow computing elements. This should be apparent when we compare successful language models with unsuccessful ones. Artificial language models usually consist of tokenization, parsing, and semantic analysis phases. This does not work on natural language because artificial languages have precise specifications and natural languages do not. No two humans use exactly the same language, nor does the same human at two points in time. Rather, language is learnable by example, so that each message causes the language of the receiver to be a little more like that of the sender. Children learn semantics before syntax, which is the opposite order from which you would write an artificial language interpreter. An example of a successful language model is a search engine. We know that most of the meaning of a text document depends only on the words it contains, ignoring word order. A search engine matches the semantics of the query with the semantics of a document mostly by matching words, but also by matching semantically related words like "water" to "wet". Here is an example of a computationally intensive but biologically plausible language model. A semantic model is a word-word matrix A such that A_ij is the degree to which words i and j are related, which you can think of as the probability of finding i and j together in a sliding window over a huge text corpus. However, semantic relatedness is a fuzzy identity relation, meaning it is reflexive, commutative, and transitive. If i is related to j and j to k, then i is related to k. Deriving transitive relations in A, also known as latent semantic analysis, is performed by singular value decomposition, factoring A = USV where S is diagonal, then discarding the small terms of S, which has the effect of lossy compression. Typically, A has about 10^6 elements and we keep only a few hundred elements of S. Fortunately there is a parallel algorithm that incrementally updates the matrices as the system learns: a 3 layer neural network where S is the hidden layer (which can grow) and U and V are weight matrices. [1]. Traditional language processing has failed because the task of converting natural language statements like "ravens are birds" to formal language is itself an AI problem. It requires humans who have already learned what ravens are and how to form and recognize grammatically correct sentences so they understand all of the hundreds of ways to express the same statement. You have to have human level understand of the logic to realize that "ravens are coming" doesn't mean "ravens -> coming". If you solve the translation problem, then you must have already solved the natural language problem. You can't take a shortcut directly to the knowledge base, tempting as it might be. You have to learn the language first, going through all the childhood stages. I would have hoped we have learned a lesson from Cyc. 1. Gorrell, Genevieve (2006), "Generalized Hebbian Algorithm for Incremental Singular Value Decomposition in Natural Language Processing", Proceedings of EACL 2006, Trento, Italy. http://www.aclweb.org/anthology-new/E/E06/E06-1013.pdf -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: Language modeling (was Re: [agi] draft for comment)
--- On Fri, 9/5/08, Pei Wang <[EMAIL PROTECTED]> wrote: > NARS indeed can learn semantics before syntax --- see > http://nars.wang.googlepages.com/wang.roadmap.pdf Yes, I see this corrects many of the problems with Cyc and with traditional language models. I didn't see a description of a mechanism for learning new terms in your other paper. Clearly this could be added, although I believe it should be a statistical process. I am interested in determining the computational cost of language modeling. The evidence I have so far is that it is high. I believe the algorithmic complexity of a model is 10^9 bits. This is consistent with Turing's 1950 prediction that AI would require this much memory, with Landauer's estimate of human long term memory, and is about how much language a person processes by adulthood assuming an information content of 1 bit per character as Shannon estimated in 1950. This is why I use a 1 GB data set in my compression benchmark. However there is a 3 way tradeoff between CPU speed, memory, and model accuracy (as measured by compression ratio). I added two graphs to my benchmark at http://cs.fit.edu/~mmahoney/compression/text.html (below the main table) which shows this clearly. In particular the size-memory tradeoff is an almost perfectly straight line (with memory on a log scale) over tests of 104 compressors. These tests suggest to me that CPU and memory are indeed bottlenecks to language modeling. The best models in my tests use simple semantic and grammatical models, well below adult human level. The 3 top programs on the memory graph map words to tokens using dictionaries that group semantically and syntactically related words together, but only one (paq8hp12any) uses a semantic space of more than one dimension. All have large vocabularies, although not implausibly large for an educated person. Other top programs like nanozipltcb and WinRK use smaller dictionaries and strictly lexical models. Lesser programs model only at the n-gram level. I don't yet have an answer to my question, but I believe efficient human-level NLP will require hundreds of GB or perhaps 1 TB of memory. The slowest programs are already faster than real time, given that equivalent learning in humans would take over a decade. I think you could use existing hardware in a speed-memory tradeoff to get real time NLP, but it would not be practical for doing experiments where each source code change requires training the model from scratch. Model development typically requires thousands of tests. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: Language modeling (was Re: [agi] draft for comment)
--- On Fri, 9/5/08, Pei Wang <[EMAIL PROTECTED]> wrote: > Like to many existing AI works, my disagreement with you is > not that > much on the solution you proposed (I can see the value), > but on the > problem you specified as the goal of AI. For example, I > have no doubt > about the theoretical and practical values of compression, > but don't > think it has much to do with intelligence. In http://cs.fit.edu/~mmahoney/compression/rationale.html I explain why text compression is an AI problem. To summarize, if you know the probability distribution of text, then you can compute P(A|Q) for any question Q and answer A to pass the Turing test. Compression allows you to precisely measure the accuracy of your estimate of P. Compression (actually, word perplexity) has been used since the early 1990's to measure the quality of language models for speech recognition, since it correlates well with word error rate. The purpose of this work is not to solve general intelligence, such as the universal intelligence proposed by Legg and Hutter [1]. That is not computable, so you have to make some arbitrary choice with regard to test environments about what problems you are going to solve. I believe the goal of AGI should be to do useful work for humans, so I am making a not so arbitrary choice to solve a problem that is central to what most people regard as useful intelligence. I had hoped that my work would lead to an elegant theory of AI, but that hasn't been the case. Rather, the best compression programs were developed as a series of thousands of hacks and tweaks, e.g. change a 4 to a 5 because it gives 0.002% better compression on the benchmark. The result is an opaque mess. I guess I should have seen it coming, since it is predicted by information theory (e.g. [2]). Nevertheless the architectures of the best text compressors are consistent with cognitive development models, i.e. phoneme (or letter) sequences -> lexical -> semantics -> syntax, which are themselves consistent with layered neural architectures. I already described a neural semantic model in my last post. I also did work supporting Hutchens and Alder showing that lexical models can be learned from n-gram statistics, consistent with the observation that babies learn the rules for segmenting continuous speech before they learn any words [3]. I agree it should also be clear that semantics is learned before grammar, contrary to the way artificial languages are processed. Grammar requires semantics, but not the other way around. Search engines work using semantics only. Yet we cannot parse sentences like "I ate pizza with Bob", "I ate pizza with pepperoni", "I ate pizza with chopsticks", without semantics. My benchmark does not prove that there aren't better language models, but it is strong evidence. It represents the work of about 100 researchers who have tried and failed to find more accurate, faster, or less memory intensive models. The resource requirements seem to increase as we go up the chain from n-grams to grammar, contrary to symbolic approaches. This is my argument why I think AI is bound by lack of hardware, not lack of theory. 1. Legg, Shane, and Marcus Hutter (2006), A Formal Measure of Machine Intelligence, Proc. Annual machine learning conference of Belgium and The Netherlands (Benelearn-2006). Ghent, 2006. http://www.vetta.org/documents/ui_benelearn.pdf 2. Legg, Shane, (2006), Is There an Elegant Universal Theory of Prediction?, Technical Report IDSIA-12-06, IDSIA / USI-SUPSI, Dalle Molle Institute for Artificial Intelligence, Galleria 2, 6928 Manno, Switzerland. http://www.vetta.org/documents/IDSIA-12-06-1.pdf 3. M. Mahoney (2000), A Note on Lexical Acquisition in Text without Spaces, http://cs.fit.edu/~mmahoney/dissertation/lex1.html -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
AI isn't cheap (was Re: Real vs. simulated environments (was Re: [agi] draft for comment.. P.S.))
--- On Fri, 9/5/08, Steve Richfield <[EMAIL PROTECTED]> wrote: >I think that a billion or so, divided up into small pieces to fund EVERY >disparate approach to see where the "low hanging fruit" is, would go a >LONG way in guiding subsequent billions. I doubt that it would take a >trillion to succeed. Sorry, the low hanging fruit was all picked by the early 1960's. By then we had neural networks [1,6,7,11,12], natural language processing and language translation [2], models of human decision making [3], automatic theorem proving [4,8,10], natural language databases [5], game playing programs [9,13], optical character recognition [14], handwriting and speech recognition [15], and important theoretical work [16,17,18]. Since then we have had mostly just incremental improvements. Big companies like Google and Microsoft have strong incentives to develop AI and have billions to spend. Maybe the problem really is hard. References 1. Ashby, W. Ross (1960), Design for a Brain, 2’nd Ed., London: Wiley. Describes a 4 neuron electromechanical neural network. 2. Borko, Harold (1967), Automated Language Processing, The State of the Art, New York: Wiley. Cites 72 NLP systems prior to 1965, and the 1959-61 U.S. government Russian-English translation project. 3. Feldman, Julian (1961), "Simulation of Behavior in the Binary Choice Experiment", Proceedings of the Western Joint Computer Conference 19:133-144 4. Gelernter, H. (1959), "Realization of a Geometry-Theorem Proving Machine", Proceedings of an International Conference on Information Processing, Paris: UNESCO House, pp. 273-282. 5. Green, Bert F. Jr., Alice K. Wolf, Carol Chomsky, and Kenneth Laughery (1961), "Baseball: An Automatic Question Answerer", Proceedings of the Western Joint Computer Conference, 19:219-224. 6. Hebb, D. O. (1949), The Organization of Behavior, New York: Wiley. Proposed the first model of learning in neurons: when two neurons fire simultaneously, the synapse between them becomes stimulating. 7. McCulloch, Warren S., and Walter Pitts (1943), "A logical calculus of the ideas immanent in nervous activity", Buletin of Mathematical Biophysics (5) pp. 115-133. 8. Newell, Allen, J. C. Shaw, H. A. Simon (1957), "Empirical Explorations with the Logic Theory Machine: A Case Study in Heuristics", Proceedings of the Western Joint Computer Conference, 15:218-239. 9. Newell, Allen, J. C. Shaw, and H. A. Simon (1958), "Chess-Playing Programs and the Problem of Complexity", IBM Journal of Research and Development, 2:320-335. 10. Newell, Allen, H. A. Simon (1961), "GPS: A Program that Simulates Human Thought", Lernende Automaten, Munich: R. Oldenbourg KG. 11. Rochester, N., J. J. Holland, L. H. Haibt, and Wl L. Duda (1956), "Tests on a cell assembly theory of the action of the brain, using a large digital computer", IRE Transactions on Information Theory IT-2: pp. 80-93. 12. Rosenblatt, F. (1958), "The perceptron: a probabilistic model for information storage and organization in the brain", Psychological Review (65) pp. 386-408. 13. Samuel, A. L. (1959), "Some Studies in Machine Learning using the Game of Checkers", IBM Journal of Research and Development, 3:211-229. 14. Selfridge, Oliver G., Ulric Neisser (1960), "Pattern Recognition by Machine", Scientific American, Aug., 203:60-68. 15. Uhr, Leonard, Charles Vossler (1963) "A Pattern-Recognition Program that Generates, Evaluates, and Adjusts its own Operators", Computers and Thought, E. A. Feigenbaum and J. Feldman eds, New York: McGraw Hill, pp. 251-268. 16. Turing, A. M., (1950) "Computing Machinery and Intelligence", Mind, 59:433-460. 17. Shannon, Claude, and Warren Weaver (1949), The Mathematical Theory of Communication, Urbana: University of Illinois Press. 18. Minsky, Marvin (1961), "Steps toward Artificial Intelligence", Proceedings of the Institute of Radio Engineers, 49:8-30. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: Language modeling (was Re: [agi] draft for comment)
--- On Fri, 9/5/08, Pei Wang <[EMAIL PROTECTED]> wrote: > Thanks for taking the time to explain your ideas in detail. > As I said, > our different opinions on how to do AI come from our very > different > understanding of "intelligence". I don't take > "passing Turing Test" as > my research goal (as explained in > http://nars.wang.googlepages.com/wang.logic_intelligence.pdf > and > http://nars.wang.googlepages.com/wang.AI_Definitions.pdf). > I disagree > with Hutter's approach, not because his SOLUTION is not > computable, > but because his PROBLEM is too idealized and simplified to > be relevant > to the actual problems of AI. I don't advocate the Turing test as the ideal test of intelligence. Turing himself was aware of the problem when he gave an example of a computer answering an arithmetic problem incorrectly in his famous 1950 paper: Q: Please write me a sonnet on the subject of the Forth Bridge. A: Count me out on this one. I never could write poetry. Q: Add 34957 to 70764. A: (Pause about 30 seconds and then give as answer) 105621. Q: Do you play chess? A: Yes. Q: I have K at my K1, and no other pieces. You have only K at K6 and R at R1. It is your move. What do you play? A: (After a pause of 15 seconds) R-R8 mate. I prefer a "preference test", which a machine passes if you prefer to talk to it over a human. Such a machine would be too fast and make too few errors to pass a Turing test. For example, if you had to add two large numbers, I think you would prefer to use a calculator than ask someone. You could, I suppose, measure intelligence as the fraction of questions for which the machine gives the preferred answer, which would be 1/4 in Turing's example. If you know the probability distribution P of text, and therefore know the distribution P(A|Q) for any question Q and answer A, then to pass the Turing test you would randomly choose answers from this distribution. But to pass the preference test for all Q, you would choose A that maximizes P(A|Q) because the most probable answer is usually the correct one. Text compression measures progress toward either test. I believe that compression measures your definition of intelligence, i.e. adaptation given insufficient knowledge and resources. In my benchmark, there are two parts: the size of the decompression program, which measures the initial knowledge, and the compressed size, which measures prediction errors that occur as the system adapts. Programs must also meet practical time and memory constraints to be listed in most benchmarks. Compression is also consistent with Legg and Hutter's universal intelligence, i.e. expected reward of an AIXI universal agent in an environment simulated by a random program. Suppose you have a compression oracle that inputs any string x and outputs the shortest program that outputs a string with prefix x. Then this reduces the (uncomputable) AIXI problem to using the oracle to guess which environment is consistent with the interaction so far, and figuring out which future outputs by the agent will maximize reward. Of course universal intelligence is also not testable because it requires an infinite number of environments. Instead, we have to choose a practical data set. I use Wikipedia text, which has fewer errors than average text, but I believe that is consistent with my goal of passing the preference test. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
RE: Language modeling (was Re: [agi] draft for comment)
--- On Sat, 9/6/08, John G. Rose <[EMAIL PROTECTED]> wrote: > Compression in itself has the overriding goal of reducing > storage bits. Not the way I use it. The goal is to predict what the environment will do next. Lossless compression is a way of measuring how well we are doing. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: AI isn't cheap (was Re: Real vs. simulated environments (was Re: [agi] draft for comment.. P.S.))
Steve, where are you getting your cost estimate for AGI? Is it a gut feeling,
or something like the common management practice of "I can afford $X so it will
cost $X"?
My estimate of $10^15 is based on the value of the world economy, US $66
trillion per year and increasing 5% annually over the next 30 years, which is
how long it will take for the internet to grow to the computational power of
10^10 human brains (at 10^15 bits and 10^16 OPS each) at the current rate of
growth, doubling every couple of years. Even if you disagree with these numbers
by a factor of 1000, it only moves the time to AGI by a few years, so the cost
estimate hardly changes.
And even if the hardware is free, you still have to program or teach about
10^16 to 10^17 bits of knowledge, assuming 10^9 bits of knowledge per brain [1]
and 1% to 10% of this is not known by anyone else. Software and training costs
are not affected by Moore's law. Even if we assume human level language
understanding and perfect sharing of knowledge, the training cost will be 1% to
10% of your working life to train the AGI to do your job.
Also, we have made *some* progress toward AGI since 1965, but it is mainly a
better understanding of why it is so hard, e.g.
- We know that general intelligence is not computable [2] or provable [3].
There is no "neat" theory.
- From Cyc, we know that coding common sense is more than a 20 year effort.
Lenat doesn't know how much more, but guesses it is maybe between 0.1% and 10%
finished.
- Google is the closest we have to AI after a half trillion dollar effort.
1. Landauer, Tom (1986), “How much do
people remember? Some estimates of the quantity of learned
information in long term memory”, Cognitive Science (10) pp.
477-493.
2. Hutter, Marcus (2003), "A Gentle
Introduction to The Universal Algorithmic Agent {AIXI}", in
Artificial General Intelligence, B. Goertzel and C. Pennachin
eds., Springer. http://www.idsia.ch/~marcus/ai/aixigentle.htm
3. Legg, Shane, (2006), "Is There an
Elegant Universal Theory of Prediction?", Technical Report
IDSIA-12-06, IDSIA / USI-SUPSI, Dalle Molle Institute for Artificial
Intelligence, Galleria 2, 6928 Manno, Switzerland.
http://www.vetta.org/documents/IDSIA-12-06-1.pdf
-- Matt Mahoney, [EMAIL PROTECTED]
--- On Sat, 9/6/08, Steve Richfield <[EMAIL PROTECTED]> wrote:
From: Steve Richfield <[EMAIL PROTECTED]>
Subject: Re: AI isn't cheap (was Re: Real vs. simulated environments (was Re:
[agi] draft for comment.. P.S.))
To: agi@v2.listbox.com
Date: Saturday, September 6, 2008, 2:58 PM
Matt,
I heartily disagree with your view as expressed here, and as stated to my by
heads of CS departments and other "high ranking" CS PhDs, nearly (but not
quite) all of whom have lost the "fire in the belly" that we all once had for
CS/AGI.
I DO agree that CS is like every other technological endeavor, in that almost
everything that can be done as a PhD thesis has already been done. but there is
a HUGE gap between a PhD thesis scale project and what that same person can do
with another few more millions and a couple more years, especially if allowed
to ignore the naysayers.
The reply is a even more complex than your well documented statement, but I'll
take my best shot at it, time permitting. Here, the angel is in the details.
On 9/5/08, Matt Mahoney <[EMAIL PROTECTED]> wrote:
--- On Fri, 9/5/08, Steve Richfield <[EMAIL PROTECTED]> wrote:
>I think that a billion or so, divided up into small pieces to fund EVERY
>disparate approach to see where the "low hanging fruit" is, would go a
>LONG way in guiding subsequent billions. I doubt that it would take a
>trillion to succeed.
Sorry, the low hanging fruit was all picked by the early 1960's. By then we had
neural networks [1,6,7,11,12],
... but we STILL do not have any sort of useful unsupervised NN, the equivalent
of which seems to be needed for any good AGI. Note my recent postings about a
potential "theory of everything" that would most directly hit unsupervised NN,
providing not only a good way of operating these, but possibly the provably
best way of operating.
natural language processing and language translation [2],
My Dr. Eliza is right there and showing that useful "understanding" out of
precise context is almost certainly impossible. I regularly meet with the folks
working on the Russian translator project, and rest assured, things are STILL
advancing fairly rapidly. Here, there is continuing funding, and I expect that
the Russian translator will eventually succeed (they already claim success).
models of human decision making [3],
These are curious but I believe them to be an emergent properties of processes
that we don't understand at all, so they have no value other than for testi
Re: Language modeling (was Re: [agi] draft for comment)
--- On Sat, 9/6/08, Pei Wang <[EMAIL PROTECTED]> wrote: > As for "compression", yes every intelligent > system needs to 'compress' > its experience in the sense of "keeping the essence > but using less > space". However, it is clearly not loseless. It is > even not what we > usually call "loosy compression", because what to > keep and in what > form is highly context-sensitive. Consequently, this > process is not > reversible --- no decompression, though the result can be > applied in > various ways. Therefore I prefer not to call it compression > to avoid > confusing this process with the technical sense of > "compression", > which is reversible, at least approximately. I think you misunderstand my use of compression. The goal is modeling or prediction. Given a string, predict the next symbol. I use compression to estimate how accurate the model is. It is easy to show that if your model is accurate, then when you connect your model to an ideal coder (such as an arithmetic coder), then compression will be optimal. You could actually skip the coding step, but it is cheap, so I use it so that there is no question of making a mistake in the measurement. If a bug in the coder produces a too small output, then the decompression step won't reproduce the original file. In fact, many speech recognition experiments do skip the coding step in their tests and merely calculate what the compressed size would be. (More precisely, they calculate word perplexity, which is equivalent). The goal of speech recognition is to find the text y that maximizes P(y|x) for utterance x. It is common to factor the model using Bayes law: P(y|x) = P(x|y)P(y)/P(x). We can drop P(x) since it is constant, leaving the acoustic model P(x|y) and language model P(y) to evaluate. We know from experiments that compression tests on P(y) correlate well with word error rates for the overall system. Internally, all lossless compressors use lossy compression or data reduction to make predictions. Most commonly, a context is truncated and possibly hashed before looking up the statistics for the next symbol. The top lossless compressors in my benchmark use more sophisticated forms of data reduction, such as mapping upper and lower case letters together, or mapping groups of semantically or syntactically related words to the same context. As a test, lossless compression is only appropriate for text. For other hard AI problems such as vision, art, and music, incompressible noise would overwhelm the human-perceptible signal. Theoretically you could compress video to 2 bits per second (the rate of human long term memory) by encoding it as a script. The decompressor would read the script and create a new movie. The proper test would be lossy compression, but this requires human judgment to evaluate how well the reconstructed data matches the original. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
RE: Language modeling (was Re: [agi] draft for comment)
--- On Sun, 9/7/08, John G. Rose <[EMAIL PROTECTED]> wrote: > From: John G. Rose <[EMAIL PROTECTED]> > Subject: RE: Language modeling (was Re: [agi] draft for comment) > To: agi@v2.listbox.com > Date: Sunday, September 7, 2008, 9:15 AM > > From: Matt Mahoney [mailto:[EMAIL PROTECTED] > > > > --- On Sat, 9/6/08, John G. Rose > <[EMAIL PROTECTED]> wrote: > > > > > Compression in itself has the overriding goal of > reducing > > > storage bits. > > > > Not the way I use it. The goal is to predict what the > environment will > > do next. Lossless compression is a way of measuring > how well we are > > doing. > > > > Predicting the environment in order to determine which data > to pack where, > thus achieving higher compression ratio. Or compression as > an integral part > of prediction? Some types of prediction are inherently > compressed I suppose. Predicting the environment to maximize reward. Hutter proved that universal intelligence is a compression problem. The optimal behavior of an AIXI agent is to guess the shortest program consistent with observation so far. That's algorithmic compression. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] open models, closed models, priors
--- On Sun, 9/7/08, Bryan Bishop <[EMAIL PROTECTED]> wrote: > On Thursday 04 September 2008, Matt Mahoney wrote: > > Yes you do. Every time you make a decision, you are > assigning a > > higher probability of a good outcome to your choice > than to the > > alternative. > > You'll have to prove to me that I make > "decisions", whatever that means. Depends on what you mean by "I". -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
[agi] Re: AI isn't cheap
--- On Sun, 9/7/08, Steve Richfield <[EMAIL PROTECTED]> wrote: >1. I believe that there is some VERY fertile but untilled ground, which >if it is half as good as it looks, could yield AGI a LOT cheaper than >other higher estimates. Of course if I am wrong, I would probably accept >your numbers. > >2. I believe that AGI will take VERY different (cheaper and more >valuable) forms than do other members on this forum. > >Each of the above effects are worth several orders of magnitude in effort. You are just speculating. The fact is that thousands of very intelligent people have been trying to solve AI for the last 50 years, and most of them shared your optimism. Perhaps it would be more fruitful to estimate the cost of automating the global economy. I explained my estimate of 10^25 bits of memory, 10^26 OPS, 10^17 bits of software and 10^15 dollars. >You really should see my Dr. Eliza demo. Perhaps you missed my comments in April. http://www.listbox.com/member/archive/303/2008/04/search/ZWxpemE/sort/time_rev/page/2/entry/5:53/20080414221142:407C652C-0A91-11DD-B3D2-6D4E66D9244B/ In any case, what does Dr. Eliza do that hasn't been done 30 years ago? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] open models, closed models, priors
--- On Sun, 9/7/08, Bryan Bishop <[EMAIL PROTECTED]> wrote: > > Depends on what you mean by "I". > > You started it - your first message had that dependency on > identity. :-) OK then. You decided to reply to my email, vs. not replying. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] Philosophy of General Intelligence
Mike, or anyone else, perhaps you can solve the technical problem of art. The entertainment sector is a significant fraction of the economy. There is a lot of money to be made if you have a machine that can produce good music or entertaining movies as well as the best artists. Most AI researchers have ignored this field, either because they don't consider it worthy of solving, or because they believe it is impossible. I believe both views are wrong. For one thing, the brain is a computer and obviously the brain can do it. Suppose you write a program that inputs jokes or cartoons and outputs whether or not they are funny. Then there is an iterative process by which you can create funny jokes or cartoons. Write a program that inputs a movie and outputs a rating of 1 to 5 stars. Then you have an iterative process for creating good movies. Write a program that recognizes good music, and you have an iterative process for creating good music. I believe the field of artificial art has been neglected because it is rare to find good artists who are also good programmers. Artists can't teach or explain how to recognize or create good art; they can only give examples. Even Donald Knuth (The Art of Computer Programming) can't explain the technique for finding beautiful algorithms, although he has created plenty of them. That is the one algorithm he doesn't know. I suspect that all AI problems, such as language, vision, and art, are of similar difficulty in terms of both hardware and software, because they are all executed on the same wetware. Rather than discuss whether the problem should be solved or can't be solved, I welcome any insights toward a solution. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] Re: AI isn't cheap
--- On Mon, 9/8/08, Steve Richfield <[EMAIL PROTECTED]> wrote: On 9/7/08, Matt Mahoney <[EMAIL PROTECTED]> wrote: >>The fact is that thousands of very intelligent people have been trying >>to solve AI for the last 50 years, and most of them shared your optimism. >Unfortunately, their positions as students and professors at various >universities have forced almost all of them into politically correct >paths, substantially all of which lead nowhere, for otherwise they would >have succeeded long ago. The few mavericks who aren't stuck in a >university (like those on this forum) all lack funding. Google is actively pursuing AI and has money to spend. If you have seen some of their talks, you know they are pursuing some basic and novel research. >>Perhaps it would be more fruitful to estimate the cost of automating the >>global economy. I explained my estimate of 10^25 bits of memory, 10^26 >>OPS, 10^17 bits of software and 10^15 dollars. You want to replicate the work currently done by 10^10 human brains. A brain has 10^15 synapses. A neuron axon has an information rate of 10 bits per second. As I said, you can argue about these numbers but it doesn't matter much. An order of magnitude error only changes the time to AGI by a few years at the current rate of Moore's Law. Software is not subject to Moore's Law so its cost will eventually dominate. A human brain has about 10^9 bits of knowledge, of which probably 10^7 to 10^8 bits are unique to each individual. That makes 10^17 to 10^18 bits that have to be extracted from human brains and communicated to the AGI. This could be done in code or formal language, although most of it will probably be done in natural language once this capability is developed. Since we don't know which parts of our knowledge is shared, the most practical approach is to dump all of it and let the AGI remove the redundancies. This will require a substantial fraction of each person's life time, so it has to be done in non obtrusive ways, such as recording all of your email and conversations (which, of course, all the major free services already do). The cost estimate of $10^15 comes by estimating the world GDP ($66 trillion per year in 2006, increasing 5% annually) from now until we have the hardware to support AGI. We have the option to have AGI sooner by paying more. Simple economics suggests we will pay up to what it is worth. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
[agi] Artificial humor
A model of artificial humor, a machine that tells jokes, or at least inputs jokes and outputs whether or not they are funny. Identify associations of the form (A ~ B) and (B ~ C) in the audience language model where (A ~ C) is believed to be false or unlikely through other associations. Test whether the joke activates A, B, and C by association to induce the association (A ~ C). This approach differs from pattern recognition and machine learning techniques used in other text classification tasks such as spam detection or information retrieval: a joke is only funny the first time you hear it. That's because once you form the association (A ~ C), it is added to the language model and you no longer have the prerequisites for the joke. Example 1: Q. Why did the chicken cross the road? A. To get to the other side. (I know, not funny, but pretend you haven't heard it). We have: A ~ B: Chickens have legs and can walk. B ~ C: People walk across the road for a reason. A ~ C: Chickens have human-like motivations. Example 2 requires a longer associative chain: (A comment about Sarah Palin) A vice president who likes hunting. What could go wrong? It invokes the false conclusion: (Sarah Palin ~ hunting accident) by inductive reasoning: (Sarah Palin ~ vice president ~ Dick Cheney ~ hunting accident) and (Sarah Palin ~ hunting ~ hunting accident). Note that all of the preconditions must be present for the joke to work. For example, the joke would not be funny if told about Joe Biden (doesn't hunt), George W. Bush (not vice president), or if you were unaware of Dick Cheney's hunting accident or that he was vice president. In order for a language model to detect the joke as funny, it would have to know that you know all four of these facts and also know that you haven't heard the joke before. Humor detection obviously requires a sophisticated language model and knowledge of popular culture, current events, and what jokes have been told before. Since entertainment is a big sector of the economy, an AGI needs all human knowledge, not just knowledge that is work related. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] Re: AI isn't cheap
(Top posting because Yahoo won't quote HTML email) Steve, Some of Google's tech talks on AI are here: http://www.google.com/search?hl=en&q=google+video+techtalks+ai&btnG=Search Google has an interest in AI because search is an AI problem, especially if you are searching for images or video. Also, their advertising model could use some help. I often go to data compression sites where Google is advertising compression socks, compression springs, air compressors, etc. I'm sure you've seen the problem. >>Software is not subject to Moore's Law so its cost will eventually dominate. >Here I could write a book and more. It could and should obey Moore's law, but history >and common practice has gone in other directions. Since you have experience writing sophisticated software on very limited hardware, perhaps you can enlighten us on how to exponentially reduce the cost of software instead of just talking about it. Maybe you can write AGI, or the next version of Windows, in one day. You might encounter a few obstacles, e.g. 1. Software testing is not computable (the halting problem reduces to it). 2. The cost of software is O(n log n). This is because you need O(log n) levels of abstraction to keep the interconnectivity of the software below the threshold of stability to chaos, above which it is not maintainable (where each software change introduces more bugs than it fixes). Abstraction levels are things like symbolic names, functions, classes, namespaces, libraries, and client-server protocols. 3. Increasing the computational power of a computer by n only increases its usefulness by log n. Useful algorithms tend to have a power law distribution over computational requirements. >>A human brain has about 10^9 bits of knowledge, of which probably 10^7 to 10^8 bits are unique to each individual. That makes 10^17 to 10^18 bits that have to be extracted from human brains and communicated to the AGI. This could be done in code or formal language, although most of it will probably be done in natural language once this capability is developed. >It would be MUCH easier and cheaper to just scan it out with something like a >scanning >UV fluorescent microscope. No it would not. Assuming we had the technology to copy brains (which we don't and you don't), then you have created a machine with human motives. You would still have to pay it to work. Do you really think you understand the brain well enough to reprogram it to want to work? >Further, I see the interest in AGIs on this forum as a sort of religious quest, that is >absurd to even consider outside of Western religions No, it is about the money. The AGIs that actually get built will be the ones that can make money for their owners. If an AGI can do anything that a human can do, then that would include work. Currently that's worth $66 trillion per year. -- Matt Mahoney, [EMAIL PROTECTED] --- On Tue, 9/9/08, Steve Richfield <[EMAIL PROTECTED]> wrote: From: Steve Richfield <[EMAIL PROTECTED]> Subject: Re: [agi] Re: AI isn't cheap To: agi@v2.listbox.com Date: Tuesday, September 9, 2008, 2:10 PM Matt, On 9/9/08, Matt Mahoney <[EMAIL PROTECTED]> wrote: --- On Mon, 9/8/08, Steve Richfield <[EMAIL PROTECTED]> wrote: On 9/7/08, Matt Mahoney <[EMAIL PROTECTED]> wrote: >>The fact is that thousands of very intelligent people have been trying >>to solve AI for the last 50 years, and most of them shared your optimism. >Unfortunately, their positions as students and professors at various >universities have forced almost all of them into politically correct >paths, substantially all of which lead nowhere, for otherwise they would >have succeeded long ago. The few mavericks who aren't stuck in a >university (like those on this forum) all lack funding. Google is actively pursuing AI and has money to spend. Maybe I am a couple of years out of date here, but the last time I looked, they were narrowly interested in search capabilities and not at all interested in linking up fragments from around the Internet, filling in missing metadata, problem solving, and the other sorts of things that are in my own area of interest. I attempted to interest them in my approaches, but got blown off apparently because they thought that my efforts were in a different direction than their interests. Have I missed something? If you have seen some of their talks, I haven't. Are any of them available somewhere? you know they are pursuing some basic and novel research. Outside of searching? >>Perhaps it would be more fruitful to estimate the cost of automating the >>global economy. I explained my estimate of 10^25 bits of memory, 10^26 >>OPS, 10^17 bits of software and 10^15 dollars. You want to replicate the work currently done by 10^10 human brains. A brain has 10^
Re: [agi] Artificial humor
--- On Wed, 9/10/08, Mike Tintner <[EMAIL PROTECTED]> wrote: > 4.To have a sense of humour, as I more or less indicated, > you have to be > able to identify with the "funny guy" making the > error - and that is an > *embodied* identification. The humour that gets the > biggest, most physical > laughs and even has you falling on the floor, usually > involves the biggest, > most physical errors - e.g. slapstick. There are no plans > that I know of, to > have computers "falling about." No, the computer's task is to recognize humor, not to experience it. You only have to model the part of the brain that sends the signal to your pleasure center. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] Artificial humor
I think artificial humor has gotten little attention because humor (along with art and emotion) is mostly a right-brain activity, while science, math, and language are mostly left-brained. It should be no surprise that since most AI researches are left-brained, their interest is in studying problems that the left brain solves. Studying humor would be like me trying to write a Russian-Chinese translator without knowing either language. It could be done, but I would have to study how other people think without introspecting on my own mind. It seems little research has been done in spite of the huge economic potential for AI. For example, we know that most of what we laugh at is ordinary conversation rather than jokes, that some animals laugh, and that infants laugh at 3.5 to 4 months (before learning language). It is not clear why laughter (the involuntary response) or the desire to laugh evolved. How does it increases fitness? http://men.webmd.com/features/why-do-we-laugh http://www.livescience.com/animals/050331_laughter_ancient.html Nevertheless, the brain computes it, so there is no reason in principle why a computer could not. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] Ability to improve ones own efficiency as a measure of intelligence
-- Matt Mahoney, [EMAIL PROTECTED]
--- On Wed, 9/10/08, Rene de Visser <[EMAIL PROTECTED]> wrote:
> Hello,
>
> I propose the following
>
> Define
>
> _a = (output - resource_usage) / resource_usage ;
> a measure of
> effeciency
>
> _b = d_a / dt; i.e. the derivative of _a with respect
> to time
>
> where output and resource_usage are appropriately
> calculated depending on
> context (scarcity/need).
This leads to programs like:
main(){while(1) printf("paperclip\n");}
which outputs lots of text without using many resources, and therefore is
intelligent.
Unfortunately we cannot define what output is useful above our own intelligence
level. We have to rely on external tests by systems that are more complex than
a single human brain. I know of two such systems. One is the economy, which
measures intelligence in dollars. The other is evolution, which measures
intelligence in number of descendants.
> We could then consider an AGI as a system that given a
> series of function
> definitions f(n) from a particular domain D and the ability
> to evaluate
> these functions over a set of parameters _x to give an
> output _y,
>
> is a system that improves its effeciency in calculating _x
> for a given _y
> and f(n), both for n fixed and for related f(n). (i.e. an
> inverse problem)
>
> i.e. _y = eval f(n) _x ;the application of a
> particular function f(n)
> from D to _x.
>
> With this framework, measuring resource usage is a
> necessary part of
> measuring intelligence, as is measuring change in
> efficiency over time
> (learning).
>
> What existing work already exists in this direction?
Legg and Hutter have proposed a definition of universal intelligence as the
expected accumulated reward in an environment simulated by random programs [1].
In practice we have to randomly sample from the infinite set of possible
environments. Alternatively, the Turing test [2] is widely accepted, and is
equivalent to text compression [3]. But the Turing test only defines human
level intelligence, nothing higher. Universal intelligence avoids this problem,
but we don't know which environments are useful. If we knew that, we would
already be that intelligent, or at least have an iterative process to get there.
I have asked this list as well as the singularity and SL4 lists whether there
are any non-evolutionary models (mathematical, software, physical, or
biological) for recursive self improvement (RSI), i.e. where the parent and not
the environment decides what the goal is and measures progress toward it. But
as far as I know, the answer is no.
Your example is one of a type that I have proposed, in which an agent chooses a
mathematical puzzle that is hard to solve but easy to verify, generates a
random puzzle, then makes randomly modified copies of itself and competes to
the death with the copies to solve the puzzle first. There are many problems
which we believe to be hard to solve but easy to check, such as subset-sum,
chess, factoring, cryptanalysis, and data compression, but none that are
provably hard.
Also, you can't just proceed on the assumption that a problem is hard. You have
to prove that a problem is hard when scaled to arbitrarily high levels of
intelligence (as measured by ability to solve it). For example, you might
choose subset-sum (find a set of integers that add to 0) without waiting for a
proof that P != NP and say that you have RSI if it is true. But that is not
enough. You also have to choose puzzles in such a way as to avoid shortcuts.
For example you have to avoid sets that contain both x and -x, and sets with
all positive numbers (provably no solution). More intelligent agents could
exploit more subtle shortcuts, which means that agents have to get smarter to
generate puzzles that avoid them. However, *that* aspect of intelligence is not
the one being optimized.
The question of RSI is obviously important to friendly AI. If RSI is possible,
then the problem is how to make the seed AI friendly and make its goal stable
through improvement, which is a very hard problem. If RSI is not possible, then
we know that friendliness is not just hard, it is impossible; intelligence will
be a strictly evolutionary process.
References
1. Legg, Shane, and Marcus Hutter (2006), A Formal Measure of Machine
Intelligence, Proc. Annual machine learning conference of Belgium and The
Netherlands (Benelearn-2006). Ghent, 2006.
http://www.vetta.org/documents/ui_benelearn.pdf
2. Turing, A. M., (1950) Computing Machinery and Intelligence, Mind, 59:433-460.
3. http://cs.fit.edu/~mmahoney/compression/rationale.html
---
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51
Powered by Listbox: http://www.listbox.com
Re: [agi] Re: AI isn't cheap
I suppose in order to justify my cost estimate I need to define more precisely what I mean by AGI. I mean the cost of building an automated economy in which people don't have to work. This is not the same as automating what people currently do. Fifty years ago we might have imagined a future with robot gas station attendants and robot sales clerks. Nobody imagined self serve gas or shopping on the internet. But the exact form of the technology does not matter. People will invest money if there is an expected payoff higher than market driven interest rates. These numbers are known. AGI is worth $10^15 no matter how you build it. An alternative goal of AGI is uploading, which I believe will cost considerably less. How much would you pay to have a machine that duplicates your memories, goals, and behavior well enough to convince everyone else that it is you, and have that machine turned on after you die? Whether such a machine is "you" (does your consciousness transfer?) is an irrelevant philosophical issue. It is not important. What is important is the percentage of people who believe it is true and are therefore willing to pay to upload. However, once we develop the technology to scan brains and simulate them, there should be no need to develop custom software or training for each individual as there is for building an economy. The cost will be determined by Moore's Law. (This does not solve the economic issues. You still have to pay uploads to work, or to write the software to automate the economy). > > Software is not subject to Moore's Law so its cost > will eventually > > dominate. > > So creating software creating software may be a high payoff > subtask. If it is possible. However, there is currently no model for recursive self improvement. The major cost of "write a program to solve X" is the cost of describing X. When you give humans a programming task, they already know most of X without you specifying the details. To tell a machine, you either have to specify X in such detail that it is equivalent to writing the program, or you have to have a machine that knows everything that humans know, which is AGI. > > A human brain has about 10^9 bits of knowledge, of > which probably > > 10^7 to 10^8 bits are unique to each individual. > > How much of this uniqueness is little more than variations > on a much > smaller number of themes and/or irrelevant to the task? Good question. Everything you have learned through language is already known to somebody else. However, the fact that you learned X from Y is known only to you and possibly Y. Some fraction of nonverbally acquired knowledge is unique to you also. What fraction is relevant? Perhaps very little if AGI means new ways of solving problems rather than duplicating the work we now do. For other tasks such as entertainment, advertising, or surveillance, everything you know is relevant. > Google to the best of my knowledge is pursuing a some areas > of narrow > AI. I do not believe they are remotely after AGI. Google has only $10^11 to spend, not $10^15. -- Matt Mahoney, [EMAIL PROTECTED] --- On Thu, 9/11/08, Samantha Atkins <[EMAIL PROTECTED]> wrote: > From: Samantha Atkins <[EMAIL PROTECTED]> > Subject: Re: [agi] Re: AI isn't cheap > To: agi@v2.listbox.com > Date: Thursday, September 11, 2008, 3:19 AM > On Sep 9, 2008, at 7:54 AM, Matt Mahoney wrote: > > > --- On Mon, 9/8/08, Steve Richfield > <[EMAIL PROTECTED]> wrote: > > On 9/7/08, Matt Mahoney <[EMAIL PROTECTED]> > wrote: > > > >>> The fact is that thousands of very intelligent > people have been > >>> trying > >>> to solve AI for the last 50 years, and most of > them shared your > >>> optimism. > > > >> Unfortunately, their positions as students and > professors at various > >> universities have forced almost all of them into > politically correct > >> paths, substantially all of which lead nowhere, > for otherwise they > >> would > >> have succeeded long ago. The few mavericks who > aren't stuck in a > >> university (like those on this forum) all lack > funding. > > > > Google is actively pursuing AI and has money to spend. > If you have > > seen some of their talks, you know they are pursuing > some basic and > > novel research. > > Google to the best of my knowledge is pursuing a some areas > of narrow > AI. I do not believe they are remotely after AGI. > > > > > > > >>> Perhaps it would be more fruitful to estimate > the cost of > >>> automating the > >>> global economy. I explained my estimate of > 10^25 bits of memory, > &
Re: [agi] Artificial humor
Mike, your argument would be on firmer ground if you could distinguish between when a computer "understands" something and when it just reacts as if it understands. What is the test? Otherwise, you could always claim that a machine doesn't understand anything because only humans can do that. -- Matt Mahoney, [EMAIL PROTECTED] --- On Thu, 9/11/08, Mike Tintner <[EMAIL PROTECTED]> wrote: > From: Mike Tintner <[EMAIL PROTECTED]> > Subject: Re: [agi] Artificial humor > To: agi@v2.listbox.com > Date: Thursday, September 11, 2008, 1:31 PM > Jiri, > > Clearly a limited 3d functionality is possible for a > program such as you > describe - as for SHRDLU. But what we're surely > concerned with here is > generality. So fine start with a restricted world of say > different kinds of > kid's blocks and similar. But then the program must be > able to tell what is > "in" what or outside, what is behind/over etc. - > and also what is moving > towards or away from an object, ( it surely should be a > "mobile" program) - > and be able to move objects. My assumption is that even a > relatively simple > such general program wouldn't work - (I obviously > haven't thought about this > in any detail). It would be interesting to have the details > about how SHRDLU > broke down. > > Also - re BillK's useful intro. of DARPA - do those > vehicles work by GPS? > > > Mike, > > > > Imagine a simple 3D scene with 2 different-size > spheres. A simple > > program allows you to change positions of the spheres > and it can > > answer question "Is the smaller sphere inside the > bigger sphere?" > > [Yes|Partly|No]. I can write such program in no time. > Sure, it's > > extremely simple, but it deals with 3D, it > demonstrates certain level > > of 3D understanding without embodyment and there is no > need to pass > > the orientation parameter to the query function. Note > that the > > orientation is just a parameter. It Doesn't > represent a "body" and it > > can be added. Of course understanding all the > real-world 3D concepts > > would take a lot more code and data than when playing > with 3D > > toy-worlds, but in principle, it's possible to > understand 3D without > > having a body. > > > > Jiri > > > > On Thu, Sep 11, 2008 at 11:24 AM, Mike Tintner > <[EMAIL PROTECTED]> > > wrote: > >> Jiri, > >> > >> Quick answer because in rush. Notice your > "if" ... Which programs > >> actually > >> do understand any *general* concepts of > orientation? SHRDLU I will gladly > >> bet, didn't...and neither do any others. > >> > >> The v. word "orientation" indicates the > reality that every picture has a > >> point of view, and refers to an observer. And > there is no physical way > >> around that. > >> > >> You have been seduced by an illusion - the > illusion of the flat, printed > >> page, existing in a timeless space. And you have > accepted implicitly that > >> there really is such a world - > "flatland" - where geometry and > >> geometrical > >> operations take place, utterly independent of you > the viewer and > >> puppeteer, > >> and the solid world of real objects to which they > refer. It demonstrably > >> isn't true. > >> > >> Remove your eyes from the page and walk around in > the world - your room, > >> say. Hey, it's not flat...and neither are any > of the objects in it. > >> Triangular objects in the world are different from > triangles on the page, > >> fundamentally different. > >> > >> But it is so difficult to shed yourself of this > illusion. You need to > >> look > >> at the history of culture and realise that the > imposition on the world/ > >> environment of first geometrical figures, and > then, more than a thousand > >> years later, the fixed point of view and > projective geometry, were - > >> and > >> remain - a SUPREME TRIUMPH OF THE HUMAN > IMAGINATION. They don't exist, > >> Jiri. They're just one of many possible > frameworks (albeit v useful) to > >> impose on the physical world. Nomadic tribes > couldn't conceive of squares > >> and enclosed spaces. Future generations will > invent new frameworks. > >> > >> Simple example of how persuasive the illusion is. > I didn't understand > >>
Re: [agi] Artificial humor
Mike Tintner <[EMAIL PROTECTED]> wrote: >To "understand" is to "REALISE" what [on earth, or >in the [real] world] is being talked about. Nice dodge. How do you distinguish between when a computer realizes something and when it just reacts as if it realizes it? Yeah, I know. Turing dodged the question too. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=111637683-c8fa51 Powered by Listbox: http://www.listbox.com
Re: [agi] Artificial humor... P.S
--- On Thu, 9/11/08, Mike Tintner <[EMAIL PROTECTED]> wrote: > "To understand/realise" is to be distinguished > from (I would argue) "to comprehend" statements. How long are we going to go round and round with this? How do you know if a machine "comprehends" something? Turing explained why he ducked the question in 1950. Because you really can't tell. http://www.loebner.net/Prizef/TuringArticle.html -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] Non-evolutionary models of recursive self-improvement (was: Ability to improve ones own efficiency as a measure of intelligence)
--- On Fri, 9/12/08, Bryan Bishop <[EMAIL PROTECTED]> wrote: > On Wednesday 10 September 2008, Matt Mahoney wrote: > > I have asked this list as well as the singularity and > SL4 lists > > whether there are any non-evolutionary models > (mathematical, > > software, physical, or biological) for recursive self > improvement > > (RSI), i.e. where the parent and not the environment > decides what the > > goal is and measures progress toward it. But as far as > I know, the > > answer is no. > > Have considered resource constraint situations where > parents kill their > young? The runt of the litter or, sometimes, others - like > when a lion > takes over a pride. Mostly in the non-human, non-Chinese > portions of > the animal kingdom. (I refer to current events re: > China's population > constraints on female offspring, of course.) There are two problems with this approach. First, if your child is smarter than you, how would you know? Second, this approach favors parents who don't kill their children. How do you prevent this trait from evolving? > Secondly, I'm still wondering about the representations > of goals in the > brain. So far, there has been no study showing the > neurobiological > basis of 'goal' in the human brain. As far as we > know, it's folk > psychology anyway, and it might not be 'true', > since there's no hard > physical evidence of the existence of goals. I'm > talking about > bottom-up existence, not top-down (top being > "us", humans and our > social contexts and such). You can define an algorithm as goal-oriented if it can be described as having a utility function U(x): X -> R (any input, real-valued output) and an iterative search over x in X such that U(x) increases over time. Whether a program has a goal depends on how you describe it. For example, linear regression has the goal of finding m and b such the straight line equation (y = mx + b) minimizes RMS error given a set of (x,y) points, but only if you solve it by iteratively adjusting m and b and evaluating the error, rather than use the conventional closed form solution. The human brain is easiest to describe as having a utility function described by Maslow's hierarchy of needs. Or you could describe it as a state table with 2^(10^15) inputs. > Does RSI have to be measured with respect to goals? Can you > prove to me > that there exists no non-goal oriented improvement > methodology? No, it is a philosophical question. What do you mean by "improvement"? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] Artificial humor... P.S
--- On Fri, 9/12/08, Mike Tintner <[EMAIL PROTECTED]> wrote: > Matt, > > What are you being so tetchy about? The issue is what it > takes for any > agent, human or machine.to understand information . How are you going to understand the issues behind programming a computer for human intelligence if you have never programmed a computer? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] Artificial humor... P.S
Mike, I understand what "understand" means. It is easy to describe what it
means to another human. But to a computer you have to define it at the level of
moving bits between registers. If you have never written software, you won't
understand the problem.
So does the following program understand?
main(){printf("Ah, now I understand!");}
You need a precise test. That is what Turing did.
-- Matt Mahoney, [EMAIL PROTECTED]
--- On Sat, 9/13/08, Mike Tintner <[EMAIL PROTECTED]> wrote:
> From: Mike Tintner <[EMAIL PROTECTED]>
> Subject: Re: [agi] Artificial humor... P.S
> To: agi@v2.listbox.com
> Date: Saturday, September 13, 2008, 12:18 AM
> >> Matt: How are you going to understand the issues
> behind programming a
> >> computer for human intelligence if you have never
> programmed a computer?
>
> Matt,
>
> We simply have a big difference of opinion. I'm saying
> there is no way a
> computer [or agent, period] can understand language if it
> can't basically
> identify/*see* (and sense) the real objects - (and
> therefore doesn't know
> what) - it's talking about. Hence people say when they
> understand at last -
> "ah now I see.." "now I see what you're
> talking about.." "now I get the
> picture."
>
> The issue of what faculties are needed to understand
> language (and be
> intelligent) is not, *in the first instance,* a matter of
> programming. I
> suggest you may have been v. uncharacteristically short in
> this exchange,
> because you may not like the starkness of the message. It
> is stark, but I
> believe it's the truth.
>
>
>
>
> ---
> agi
> Archives: https://www.listbox.com/member/archive/303/=now
> RSS Feed: https://www.listbox.com/member/archive/rss/303/
> Modify Your Subscription:
> https://www.listbox.com/member/?&;
> Powered by Listbox: http://www.listbox.com
---
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69
Powered by Listbox: http://www.listbox.com
[agi] A model for RSI
I have produced a mathematical model for recursive self improvement and would appreciate any comments before I publish this. http://www.mattmahoney.net/rsi.pdf In the paper, I try to give a sensible yet precise definition of what it means for a program to have a goal. Then I describe infinite sequences of programs that improve with respect to reaching a goal within fixed time bounds, and finally I give an example (in C) of a program that outputs the next program in this sequence. Although it is my long sought goal to prove or disprove RSI, it doesn't entirely resolve the question because the rate of knowledge gain is O(log n) and I prove that is the best you can do given fixed goals. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] A model for RSI
Pei, thanks for the comments. I posted an updated version of the paper to http://www.mattmahoney.net/rsi.pdf I also found a bug in my RSI program and posted a revised version (which is also easier to read). My original program did not depend on input t, so according to my definition, it did not have a goal. The new program has a simple goal of printing big numbers, which get bigger as the time bound t increases. It also outputs an improved copy of itself such that for any input, the output is larger by 1. It is a 13 line C program such that the n'th generation outputs t+n. I added a section to the paper explaining why I used a batch mode model of an intelligent agent when normally we use an interactive model (such as yours and AIXI). I distinguish between self-improvement and learning, such that self improvement means the program rewrites its software to better achieve some goal. It has to do this without any outside help beyond what it initially knows. If it updates itself based on new information, then that's learning. For batch mode testing, a utility function is sufficient to define a goal. > *. "AIXI has insufficient knowledge (none initially) > ..." > > But it assumes a reward signal, which contains sufficient > knowledge to > evaluate behaviors. What if the reward signal is wrong? The reward signal (controlled by the environment) is by definition what the agent tries to maximize. It can't be "wrong". -- Matt Mahoney, [EMAIL PROTECTED] --- On Sun, 9/14/08, Pei Wang <[EMAIL PROTECTED]> wrote: > From: Pei Wang <[EMAIL PROTECTED]> > Subject: Re: [agi] A model for RSI > To: agi@v2.listbox.com > Date: Sunday, September 14, 2008, 1:58 PM > Matt, > > Thanks for the paper. Some random comments: > > *. "If RSI is possible, then it is critical that the > initial goals of > the first iteration of agents (seed AI) are friendly to > humans and > that the goals not drift through successive > iterations." > > As I commented on Ben's paper recently, here the > implicit assumption > is that the initial goals fully determines the goal > structure, which I > don't think is correct. If you think otherwise, you > should argue for > it, or at least make it explicit. > > *. "Turing [5] defined AI as the ability of a machine > to fool a human > into believing it was another human." > > No he didn't. Turing proposed the imitation game as a > sufficient > condition for intelligence, and he made it clear that it > may not be a > necessary condition by saying "May not machines carry > out something > which ought to be described as thinking but which is very > different > from what a man does? This objection is a very strong one, > but at > least we can say that if, nevertheless, a machine can be > constructed > to play the imitation game satisfactorily, we need not be > troubled by > this objection." > > *. "This would solve the general intelligence problem > once and for > all, except for the fact that the strategy is not > computable." > > Not only that. Other exceptions include the situations > where the > definition doesn't apply, such as in systems where > goals change over > time, where no immediate and reliable reward signals are > given, etc., > not to mention the unrealistic assumption on infinite > resources. > > *. "AIXI has insufficient knowledge (none initially) > ..." > > But it assumes a reward signal, which contains sufficient > knowledge to > evaluate behaviors. What if the reward signal is wrong? > > *. "Hutter also proved that in the case of space bound > l and time bound t ..." > > That is not the same thing as "insufficient > resources". > > *. "We define a goal as a function G: N → R mapping > natural numbers > ... to real numbers." > > I'm sure you can build systems with such a goal, though > call it a > "definition of goal" seems too strong --- are you > claiming that all > the "goals" in the AGI context can be put into > this format? On the > other hand, are all N → R functions goals? If not, how to > distinguish > them? > > *. "running P longer will eventually produce a better > result and never > produce a worse result afterwards" > > This is true for certain goals, though not for all. Some > goals ask for > keeping some parameter (such as body temperature) at a > certain value, > which cannot be covered by your definition using > monotonically > increasing function. > > *. "Define an improving sequence with respect to G as > an infinite > sequence of programs P1, P2,
Re: [agi] uncertain logic criteria
--- On Wed, 9/17/08, Abram Demski <[EMAIL PROTECTED]> wrote: > Most people on this list should know about at least 3 > uncertain logics > claiming to be AGI-grade (or close): > > --Pie Wang's NARS > --Ben Goertzel's PLN > --YKY's recent hybrid logic proposal > > It seems worthwhile to stop and take a look at what > criteria such > logics should be judged by. So, I'm wondering: what > features would > people on this list like to see? How about testing in the applications where they would actually be used, perhaps on a small scale. For example, how would these logics be used in a language translation program, where the problem is to convert a natural language sentence into its structured representation and convert it back in another language. How easy is it to populate the database with the gigabyte or so of common sense knowledge needed to provide the context in which natural language statements are interpreted? (Cyc proved it is very hard). For a lot of the problems where we actually use structured data, a relational database works pretty well. However it is nice to see proposals that deal with inconsistencies in the database better than just reporting an error. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] Re: [OpenCog] Proprietary_Open_Source
2008/9/17 JDLaw <[EMAIL PROTECTED]>: > IMHO to all, > There is an important morality discussion about how sentient life will > be treated that has not received its proper treatment in your > discussion groups. I have seen glimpses of this topic, but no real > action proposals. How would you feel if you created this wonderful > child (computer intelligence) in this standard GNU model and then > people began to exploit and torture your own child? You can do this now if you wish. I wrote a program called autobliss (see http://www.mattmahoney.net/autobliss.txt ), a 2-input logic gate that is trained by reinforcement learning. A teacher selects random 2-bit inputs, then rewards the student if it gives the correct output or punishes it if the output is incorrect. You can choose the level of simulated pleasure or pain given during each training session. The program protects against excessive simulated torture by killing the student first, but you can easily modify the software to remove this protection and then choose punishment regardless of which output the student gives. The program is released under GPL so you can legally do this and then distribute it in an @home type screensaver so that millions of PC's use all their spare CPU cycles to inflict excruciating simulated pain on millions of copies. Or maybe you can precisely define what makes a program "sentient" as opposed to just property. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Two goals of AGI (was Re: [agi] Re: [OpenCog] Re: Proprietary_Open_Source)
--- On Thu, 9/18/08, Bob Mottram <[EMAIL PROTECTED]> wrote: > > And this is the problem. Although some people have > the goal of making > > an artificial person with all the richness and nuance > of a sentient > > creature with thoughts and feelings and yada yada > yada.. some of us > > are just interested in making more intelligent systems > to do automated > > tasks. For some reason people think we're going > to do this by making > > an artificial person and then enslaving them.. > that's not going to > > happen because its just not necessary. > > > In this case what you're doing is really narrow AI, not > AGI. Lets distinguish between the two major goals of AGI. The first is to automate the economy. The second is to become immortal through uploading. The first goal does not require any major breakthroughs in AI theory, just lots of work. If you have a lot of narrow AI and an infrastructure for routing natural language messages to the right experts, then you have AGI. I described one protocol (competitive message routing, or CMR) to make this happen at http://www.mattmahoney.net/agi.html but the reality will probably be more complex, using many protocols to achieve the same result. Regardless of the exact form, we can estimate its cost. The human labor now required to run the global economy was worth US $66 trillion in 2006 and is increasing at 5% per year. At current interest rates, the value of an automated economy is about $1 quadrillion. We should expect to pay this much, because there is a tradeoff between having it sooner and waiting until the cost of hardware drops. This huge cost requires a competitive system with distributed ownership in which information has negative value and resource owners compete for attention and reputation by providing quality data. CMR, like any distributed knowledge base, is hostile: we will probably spend as many CPU cycles and human labor filtering spam and attacks as detecting useful features in language and video. The second goal of AGI is uploading and intelligence augmentation. It requires advances in modeling, scanning, and programming human brains and bodies. You are programmed by evolution to fear death, so creating a copy of you that others cannot distinguish from you that will be turned on after you die has value to you. Whether the copy is really "you" and contains your consciousness is an unimportant philosophical question. If you see your dead friends brought back to life with all of their memories and behavior intact (as far as you can tell), you will probably consider it a worthwhile investment. Brain scanning is probably not required. By the time we have the technology to create artificial generic humans, surveillance will probably be so cheap and pervasive that creating a convincing copy of you could be done just by accessing public information about you. This would include all of your communication through computers (email, website accesses, phone calls, TV), and all of your travel and activities in public places captured on video. Uploads will have goals independent of their owners because their owners have died. They will also have opportunities not available to human brains. They could add CPU power, memory, I/O, and bandwidth. Or they could reprogram their brains, to live in simulated Utopian worlds, modify their own goals to want what they already have, or enter euphoric states. Natural selection will favor the former over the latter. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: Two goals of AGI (was Re: [agi] Re: [OpenCog] Re: Proprietary_Open_Source)
--- On Thu, 9/18/08, Ben Goertzel <[EMAIL PROTECTED]> wrote: >>Lets distinguish between the two major goals of AGI. The first is to >>automate the economy. The second is to become immortal through uploading. > >Peculiarly, you are leaving out what to me is by far the most important and >interesting goal: > >The creation of beings far more intelligent than humans yet benevolent toward >humans That's what I mean by an automated economy. Google is already more intelligent than any human at certain tasks. So is a calculator. Both are benevolent. They differ in the fraction of our tasks that they can do for us. When that fraction is 100%, that's AGI. >>The first goal does not require any major breakthroughs in AI theory, just >>lots of work. If you have a lot of narrow AI and an infrastructure for >>routing natural language messages to the right experts, then you have AGI. >Then you have a hybrid human/artificial intelligence, which does not fully >automate the economy, but only partially does so -- it still relies on human >experts. If humans are to remain in control of AGI, then we have to make informed, top level decisions. You can call this work if you want. But if we abdicate all thinking to machines, then where does that leave us? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: Two goals of AGI (was Re: [agi] Re: [OpenCog] Re: Proprietary_Open_Source)
--- On Thu, 9/18/08, Vladimir Nesov <[EMAIL PROTECTED]> wrote: > And to boot, both of you don't really know what you want. What we want has been programmed into our brains by the process of evolution. I am not pretending the outcome will be good. Once we have the technology to have everything we want, or to want what we have, then a more intelligent species will take over. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: Two goals of AGI (was Re: [agi] Re: [OpenCog] Re: Proprietary_Open_Source)
--- On Thu, 9/18/08, Trent Waddington <[EMAIL PROTECTED]> wrote: > On Fri, Sep 19, 2008 at 3:36 AM, Matt Mahoney > <[EMAIL PROTECTED]> wrote: > > Lets distinguish between the two major goals of AGI. > The first is to automate the economy. The second is to > become immortal through uploading. > > Umm, who's goals are these? Who said they are > "the [..] goals of > AGI"? I'm pretty sure that what I want AGI for is > going to be > different to what you want AGI for as to what anyone else > wants AGI > for.. and any similarities are just superficial. So, I guess I should say, "the two commercial applications of AGI". I realize people are working on AGI today as pure research, to better understand the brain, to better understand how to solve hard problems, and so on. I think eventually this knowledge will be applied for profit. Perhaps there are some applications I haven't thought of? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] Case-by-case Problem Solving (draft)
--- On Thu, 9/18/08, Pei Wang <[EMAIL PROTECTED]> wrote: > URL: http://nars.wang.googlepages.com/wang.CaseByCase.pdf I think it would be interesting if you had some experimental results. Could CPS now solve a problem like "sort [3 2 4 1]" in its current state? If not, how much knowledge does it need, and how long would it run? How long would it take to program its knowledge base? Would CPS then use its experience to help it solve similar problems like "sort [4 2 4 3]"? Could you give an example of a problem that CPS can now solve? What is your opinion on using CPS to solve hard problems, like factoring 1000 digit numbers, or finding strings x and y such that x != y and MD5(x) = MD5(y). Do you think that CPS could find clever solutions such as the collisions found by Wang and Yu? If so, what resources would be required? MD5 cryptographic one way hash standard: http://www.ietf.org/rfc/rfc1321.txt Attack on MD5: http://web.archive.org/web/20070604205756/http://www.infosec.sdu.edu.cn/paper/md5-attack.pdf -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: Two goals of AGI (was Re: [agi] Re: [OpenCog] Re: Proprietary_Open_Source)
--- On Thu, 9/18/08, Ben Goertzel <[EMAIL PROTECTED]> wrote: >I believe there is a qualitative difference btw AGI and narrow-AI, so that no >tractably small collection of computationally-feasible narrow-AI's (like >Google etc.) are going to achieve general intelligence at the human level or >anywhere near. I think you need an AGI architecture & approach that is >fundamentally different from narrow-AI approaches... Well, yes, and that difference is a distributed index, which has yet to be built. Also, what do you mean by "human level intelligence"? What test do you use? My calculator already surpasses human level intelligence depending on the tests I give it. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: Two goals of AGI (was Re: [agi] Re: [OpenCog] Re: Proprietary_Open_Source)
--- On Thu, 9/18/08, Trent Waddington <[EMAIL PROTECTED]> wrote: > On Fri, Sep 19, 2008 at 7:54 AM, Matt Mahoney > <[EMAIL PROTECTED]> wrote: > > Perhaps there are some applications I haven't > thought of? > > Bahahaha.. Gee, ya think? So perhaps you could name some applications of AGI that don't fall into the categories of (1) doing work or (2) augmenting your brain? A third one occurred to me: launching a self improving or evolving AGI to consume all available resources, i.e. an intelligent worm or self replicating nanobots. This really isn't a useful application, but I'm sure somebody, somewhere, might think it would be really cool to see if it would launch a singularity and/or wipe out all DNA based life. Oh, I'm sure the first person to try it would take precautions like inserting a self destruct mechanism that activates after some number of replications. (The 1988 Morris worm had software intended to slow its spread, but it had a bug). Or maybe they will be like the scientists who believed that the idea of a chain reaction in U-235 was preposterous... (Thankfully, the scientists who actually built the first atomic pile took some precautions, such as standing by with an axe to cut a rope suspending a cadmium control rod in case things got out of hand. They got lucky because of an unanticipated phenomena in which a small number of nuclei had delayed fission, which made the chain reaction much easier to control). -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] Case-by-case Problem Solving (draft)
Actually, CPS doesn't mean solving problems without algorithms. CPS is itself an algorithm, as described on pages 7-8 of Pei's paper. However, as I mentioned, I would be more convinced if there were some experimental results showing that it actually worked. -- Matt Mahoney, [EMAIL PROTECTED] --- On Thu, 9/18/08, Mike Tintner <[EMAIL PROTECTED]> wrote: From: Mike Tintner <[EMAIL PROTECTED]> Subject: Re: [agi] Case-by-case Problem Solving (draft) To: agi@v2.listbox.com Date: Thursday, September 18, 2008, 8:51 PM Ben, I'm only saying that CPS seems to be loosely equivalent to wicked, ill-structured problem-solving, (the reference to convergent/divergent (or crystallised vs fluid) etc is merely to point out a common distinction in psychology between two kinds of intelligence that Pei wasn't aware of in the past - which is actually loosely equivalent to the distinction between narrow AI and general AI problemsolving). In the end, what Pei is/isn't aware of in terms of general knowledge, doesn't matter much - don't you think that his attempt to do without algorithms IS v. important? And don't you think any such attempt would be better off referring explicitly to the literature on wicked, ill-structured problems? I don't think that pointing all this out is silly - this (a non-algorithmic approach to CPS/wicked/whatever) is by far the most important thing currently being discussed here - and potentially, if properly developed, revolutionary.. Worth getting excited about, no? (It would also be helpful BTW to discuss the "wicked" literature because it actually has abundant examples of wicked problems - and those, you must admit, are rather hard to come by here ). Ben: TITLE: Case-by-case Problem Solving (draft) AUTHOR: Pei Wang But you seem to be reinventing the term for wheel. There is an extensive literature, including AI stuff, on "wicked, ill-structured" problems, (and even "nonprogrammed decisionmaking" which won't, I suggest, be replaced by "case-by-case PS". These are well-established terms. You similarly seemed to be unaware of the v. common distinction between convergent & divergent problem-solving. Mike, I have to say I find this mode of discussion fairly silly.. Pei has a rather comprehensive knowledge of AI and a strong knowledge of cog-sci as well. It is obviously not the case that he is unaware of these terms and ideas you are referring to. Obviously, what he means by "case-by-case problem solving" is NOT the same as "nonprogrammed decisionmaking" nor "divergent problem-solving." In his paper, he is presenting a point of view, not seeking to compare this point of view to the whole corpus of literature and ideas that he has absorbed during his lifetime. I happen not to fully agree with Pei's thinking on these topics (though I like much of it), but I know Pei well enough to know that those. places where his thinking diverges from mine, are *not* due to ignorance of the literature on his part... agi | Archives | Modify Your Subscription --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: Two goals of AGI (was Re: [agi] Re: [OpenCog] Re: Proprietary_Open_Source)
--- On Thu, 9/18/08, Ben Goertzel <[EMAIL PROTECTED]> wrote: >>Well, yes, and that difference is a distributed index, which has yet to be >>built. >I extremely strongly disagree with the prior sentence ... I do not think that >a distributed index is a sufficient architecture for powerful AGI at the human >level, beyond, or anywhere near... Well, keep in mind that I am not trying to build a human-like AGI with its own goals. I am designing a distributed system with billions of owners, each of whom has their own interests and (conflicting) goals. To the user, the AGI is like a smarter internet. It would differ from Google in that any message you post is instantly available to anyone who cares (human or machine). There is no distinction between queries and documents. Posting a message could initiate an interactive conversation, or result in related messages posted later being sent to you. A peer needs two types of knowledge. It knows about some specialized topic, and it also knows which other peers are experts on related topics. For simple peers, "related" just means they share the same words, and a peer is simply a cache of messages posted and received recently by its owner. In my CMR proposal, messages are stamped with the ID and time of origin as well as any peers they were routed through. This cached header information constitutes knowledge about related peers. When a peer receives a message, it compares the words in it to cached messages and routes a copy to the peers listed in the headers of those messages. Peers have their own policies regarding their areas of specialization, which can be as simple as giving the cache priority to messages originating from its owner. There is no provision to delete messages from the network once they are posted. Each peer would have its own deletion policy. The environment is competitive and hostile. Peers compete for reputation and attention by providing quality information, which allows them to charge more for routing targeted ads. Peers are responsible for authenticating their sources, and risk blacklisting if they route too much spam. Peers thus have an incentive to be intelligent, for example, using better language models such as a stemmer, thesaurus, and parser to better identify related messages, or providing specialized services that understand a narrow subset of natural language, the way Google calculator understands questions like "how many gallons in 50 cubic feet?" So yeah, it is a little different than narrow AI. As to why I'm not building it, it's because I estimate it will cost $1 quadrillion. Google controls about 1/1000 of the computing power of the internet. I am talking about building something 1000 times bigger. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: Two goals of AGI (was Re: [agi] Re: [OpenCog] Re: Proprietary_Open_Source)
--- On Thu, 9/18/08, John LaMuth <[EMAIL PROTECTED]> wrote: > You have completely left out the human element or > friendly-type appeal > > How about a AGI personal assistant / tutor / PR interface > > Everyone should have one > > The market would be virtually unlimited ... That falls under the category of (1) doing work. -- Matt Mahoney, [EMAIL PROTECTED] > - Original Message - > From: "Matt Mahoney" <[EMAIL PROTECTED]> > To: > Sent: Thursday, September 18, 2008 6:34 PM > Subject: Re: Two goals of AGI (was Re: [agi] Re: [OpenCog] > Re: > Proprietary_Open_Source) > > > > --- On Thu, 9/18/08, Trent Waddington > <[EMAIL PROTECTED]> wrote: > > > >> On Fri, Sep 19, 2008 at 7:54 AM, Matt Mahoney > >> <[EMAIL PROTECTED]> wrote: > > > >> > Perhaps there are some applications I > haven't > >> thought of? > >> > >> Bahahaha.. Gee, ya think? > > > > So perhaps you could name some applications of AGI > that don't fall into > > the categories of (1) doing work or (2) augmenting > your brain? --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: Two goals of AGI (was Re: [agi] Re: [OpenCog] Re: Proprietary_Open_Source)
--- On Thu, 9/18/08, Trent Waddington <[EMAIL PROTECTED]> wrote: > On Fri, Sep 19, 2008 at 11:34 AM, Matt Mahoney > <[EMAIL PROTECTED]> wrote: > > So perhaps you could name some applications of AGI > that don't fall into the categories of (1) doing work or > (2) augmenting your brain? > > Perhaps you could list some uses of a computer that > don't fall into > the category of (1) computation (2) communication. Do you > see how > pointless reasoning at this level of abstraction is? No it is not. (and besides, there is (3) storage). We can usefully think of the primary uses of computers going through different phases, e.g. 1950-1970 - computation (numerical calculation) 1970-1990 - storage (databases) 1990-2010 - communication (internet) 2010-2030 - profit-oriented AI (automating the economy) 2030-2050 - brain augmentation and uploading > And to get back to the original topic of conversation, > putting > restrictions on the use of supposedly open source code, the > effects of > those restrictions can no more be predicted than the > potential > applications of the technology. Which, I think, is a > rational piler > of the need for freedom.. you don't know better, so who > are you to put > these restrictions on others? I don't advocate any such thing, even if it were practical. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: Two goals of AGI (was Re: [agi] Re: [OpenCog] Re: Proprietary_Open_Source)
--- On Thu, 9/18/08, Ben Goertzel <[EMAIL PROTECTED]> wrote: Sorry for being unclear. The two categories of AI that I refer to are the near term "smart internet" automated economy and longer term "artificial human" or transhuman phases. In the smart internet phase, individuals with competing goals own parts of the AGI (peers) and the message routing infrastructure provides a market that satisfies human goals efficiently. Peers work to satisfy the goals of their owners. Later, the network will be populated with intelligent peers that have their own goals independent of their (former) owners. Just as the computation, storage, and communication eras of computing lack sharp boundaries, so will the automated economy and transhuman eras. Early on, people will add peers that try to appear human for various reasons, and with various degrees of success. These peers will know a lot about one person (such as its owner) and go to the net for more general knowledge about people. This becomes easier as computers get faster and surveillance becomes more pervasive. Basically, your CMR client knows everything you ever typed into a computer. People may program their peers to become autonomous and emulate their owners after they die. They might work, earn money, and pay for hosting. Later, peers may buy robotic bodies as the technology becomes available. About intelligence testing, early AGI would pass an IQ test or Turing test by routing questions to the appropriate experts. Later, transhumans could do the same, only they might choose not to take your silly test. >>So perhaps you could name some applications of AGI that don't fall into the >>categories of (1) doing work or (2) augmenting your brain? > >3) learning as much as possible Early AGI would do so because it is the most effective strategy to meet the goals of its owners. Later, transhumans would learn because they want to learn. They would want to learn because this is a basic human goal which was copied into them. Humans want to learn because intelligence requires both the ability to learn and the desire to learn. Humans are intelligent because it increases evolutionary fitness. >4) proving as many theorems as possible Early AGI would route your theorem to theorem proving experts, rank the results, and use the results to improve future rankings and future routing of similar questions. Later, transhumans could just ask the net. >5) figuring out how to improve human life as much as possible Early AGI will make the market more efficient, which improves the lives of everyone who uses it. Later, transhumans will have their own ideas what "improve" means. That is where AGI becomes dangerous. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: Two goals of AGI (was Re: [agi] Re: [OpenCog] Re: Proprietary_Open_Source)
--- On Thu, 9/18/08, John LaMuth <[EMAIL PROTECTED]> wrote: > I always advocated a clear seperation between work and PLAY > > Here the appeal would be amusement / entertainment - not > any specified work > goal > > Have my PR - AI call your PR - AI !! > > and Show Me the $$$ !! As more of the economy is automated, we will spend a greater fraction of our time and money on entertainment. Automatically generating music, movies, art, and artificial worlds are hard AI problems. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] Where the Future of AGI Lies
--- On Fri, 9/19/08, Ben Goertzel <[EMAIL PROTECTED]> wrote: Mike, Google has had basically no impact on the AGI thinking of myself or 95% of the other serious AGI researchers I know... Which is rather curious, because Google is the closest we have to AI at the moment. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] Where the Future of AGI Lies
--- On Fri, 9/19/08, Jiri Jelinek <[EMAIL PROTECTED]> wrote: > On Fri, Sep 19, 2008 at 10:46 AM, Matt Mahoney > <[EMAIL PROTECTED]> wrote: > >Google is the closest we have to AI at the moment. > > Matt, > > There is a difference between being good at > a) finding problem-related info/pages, and > b) finding functional solutions (through reasoning), > especially when > all the needed data is available. > > Google cannot handle even trivial answer-embedded > questions. Q: how many fluid ounces in a cubic mile? Google: 1 cubic mile = 1.40942995 × 10^14 US fluid ounces Q: who is the tallest U.S. president? Google: Abraham Lincoln at six feet four inches. (along with other text) Current AI (or AGI) research tends to emphasize reasoning ability rather than natural language understanding or rating the reliability of information from different sources, as if these things were not hard or important. Reasoning requires far less computation, as was demonstrated in the early 1960's. Current models that deal with uncertainty have not addressed the hard problems. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] Where the Future of AGI Lies
Quick test. Q: What world leader lost 2 fingers playing with grenades as a boy? powerset.com: doesn't know. cognition.com (wiki): doesn't know. google.com: the second link leads to a scanned page of a book giving the answer as Boris Yeltsin. -- Matt Mahoney, [EMAIL PROTECTED] --- On Fri, 9/19/08, BillK <[EMAIL PROTECTED]> wrote: > From: BillK <[EMAIL PROTECTED]> > Subject: Re: [agi] Where the Future of AGI Lies > To: agi@v2.listbox.com > Date: Friday, September 19, 2008, 11:34 AM > On Fri, Sep 19, 2008 at 3:15 PM, Jiri Jelinek wrote: > > There is a difference between being good at > > a) finding problem-related info/pages, and > > b) finding functional solutions (through reasoning), > especially when > > all the needed data is available. > > > > Google cannot handle even trivial answer-embedded > questions. > > > > Last I heard Peter Norvig was saying that Google had no > interest in > putting a natural language front-end on Google. > <http://slashdot.org/article.pl?sid=07/12/18/1530209> > > But other companies are interested. The main two are: > Powerset <http://www.powerset.com/> > and > Cognition <http://www.cognition.com/> > > A new startup Eeggi is also interesting. > <http://www.eeggi.com/> > > > BillK > --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] Where the Future of AGI Lies
--- On Fri, 9/19/08, BillK <[EMAIL PROTECTED]> wrote: > On Fri, Sep 19, 2008 at 6:13 PM, Matt Mahoney wrote: > > Quick test. > > > > Q: What world leader lost 2 fingers playing with > grenades as a boy? > > powerset.com: doesn't know. > > cognition.com (wiki): doesn't know. > > > > google.com: the second link leads to a scanned page of > a book giving the answer as Boris Yeltsin. > > > > > At present, they are not trying to compete with the Google > search engine. :) > > They only search in Wikipedia, using this as a basis to > test their > Natural Language front-end. > > A better test would be to ask a complex question, where you > know the > answer is in Wikipedia, and see if they answer the question > better > than a Google search of Wikipedia only. >From http://en.wikipedia.org/wiki/Yeltsin "Boris Yeltsin studied at Pushkin High School in Berezniki in Perm Krai. He was fond of sports (in particular skiing, gymnastics, volleyball, track and field, boxing and wrestling) despite losing the thumb and index finger of his left hand when he and some friends sneaked into a Red Army supply depot, stole several grenades, and tried to dissect them.[5]" But to be fair, Google didn't find it either. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
The brain does not implement formal logic (was Re: [agi] Where the Future of AGI Lies)
--- On Fri, 9/19/08, Jiri Jelinek <[EMAIL PROTECTED]> wrote: > Try "What's the color of Dan Brown's black coat?" What's the excuse > for a general problem solver to fail in this case? NLP? It > then should use a formal language or so. Google uses relatively good > search algorithms but decent general problem solving IMO requires > very different algorithms/design. So, what formal language model can solve this problem? First order logic? Uncertain logic (probability and confidence)? Logic augmented with notions of specialization, time, cause and effect, etc. There seems to be a lot of effort to implement reasoning in knowledge representation systems, even though it has little to do with how we actually think. We focus on problems like: All men are mortal. Socrates is a man. Therefore ___? The assumed solution is to convert it to a formal representation and apply the rules of logic: For all x: man(x) -> mortal(x) man(Socrates) => mortal(Socrates) which has 3 steps: convert English to a formal representation (hard AI), solve the problem (easy), and convert back to English (hard AI). Sorry, that is not a solution. Consider how you learned to convert natural language to formal logic. You were given lots of examples and induced a pattern: Frogs are green = for all x: frog(x) -> green(x). Fish are animals = for all x: fish(x) -> animal(x). ... Y are Z: for all x: Y(x) -> Z(x). along with many other patterns. (Of course, this requires learning semantics first, so you don't confuse examples like "they are coming"). But if you can learn these types of patterns then with no additional effort you can learn patterns that directly solve the problem... Frogs are green. Kermit is a frog. Therefore Kermit is green. Fish are animals. A minnow is a fish. Therefore a minnow is an animal. ... Y are Z. X is a Y. Therefore X is a Z. ... Men are mortal. Socrates is a man. Therefore Socrates is mortal. without ever going to a formal representation. People who haven't studied logic or its notation can certainly learn to do this type of reasoning. So perhaps someone can explain why we need formal knowledge representations to reason in AI. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: The brain does not implement formal logic (was Re: [agi] Where the Future of AGI Lies)
--- On Fri, 9/19/08, Jan Klauck <[EMAIL PROTECTED]> wrote: > Formal logic doesn't scale up very well in humans. That's why this > kind of reasoning is so unpopular. Our capacities are that > small and we connect to other human entities for a kind of > distributed problem solving. Logic is just a tool for us to > communicate and reason systematically about problems we would > mess up otherwise. Exactly. That is why I am critical of probabilistic or uncertain logic. Humans are not very good at logic and arithmetic problems requiring long sequences of steps, but duplicating these defects in machines does not help. It does not solve the problem of translating natural language into formal language and back. When we need to solve such a problem, we use pencil and paper, or a calculator, or we write a program. The problem for AI is to convert natural language to formal language or a program and back. The formal reasoning we already know how to do. Even though a language model is probabilistic, probabilistic logic is not a good fit. For example, in NARS we have deduction (P->Q, Q->R) => (P->R), induction (P->Q, P->R) => (Q->R), and abduction (P->R, Q->R) => (P->Q). Induction and abduction are not strictly true, of course, but in a probabilistic logic we can assign them partial truth values. For language modeling, we can simplify the logic. If we accept the "converse" rule (P->Q) => (Q->P) as partially true (if rain predicts clouds, then clouds may predict rain), then we can derive induction and abduction from deduction and converse. For induction, (P->Q, P->R) => (Q->P, P->R) => (Q->R). Abduction is similar. Allowing converse, the statement (P->Q) is really a fuzzy equivalence or association (P ~ Q), e.g. (rain ~ clouds). A language model is a set of associations between concepts. Language learning consists of two operations carried out on a massively parallel scale: forming associations and forming new concepts by clustering in context space. An example of the latter is: the dog is the cat is the house is ... the (noun) is So if we read "the glorp is" we learn that "glorp" is a noun. Likewise, we learn something of its meaning from its more distant context, e.g. "the glorp is eating my flowers". We do this by the transitive property of association, e.g. (glorp ~ eating flowers ~ rabbit). This is not to say NARS or other systems are wrong, but rather that they have more capability than we need to solve reasoning in AI. Whether the extra capability helps or not is something that requires experimental verification. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: The brain does not implement formal logic (was Re: [agi] Where the Future of AGI Lies)
--- On Sat, 9/20/08, Ben Goertzel <[EMAIL PROTECTED]> wrote: >If formal reasoning were a solved problem in AI, then we would have >theorem-provers that could prove deep, complex theorems unassisted. We >don't. This indicates that formal reasoning is NOT a solved problem, because >no one has yet gotten "history guided adaptive inference control" to really >work well. Which is IMHO because formal reasoning guidance ultimately >requires the same kind of analogical, contextual commonsense reasoning as >guidance of reasoning about everyday life... I mean that formal reasoning is solved in the sense of executing algorithms, once we can state the problems in that form. I know that some problems in CS are hard. I think that the intuition that mathematicians use to prove theorems is a language modeling problem. >Also, you did not address my prior point that Hebbian learning at the neural >level is strikingly similar to formal logic... I agree that neural networks can model formal logic. However, I don't think that formal logic is a good way to model neural networks. Language learning consists of learning associations between concepts (possibly time-delayed, enabling prediction) and learning new concepts by clustering in context space. Both of these operations can be done efficiently and in parallel with neural networks. They can't be done efficiently with logic. There is experimental evidence to back up this view. The top two compressors in my large text benchmark use dictionaries in which semantically related words are grouped together and the groups are used as context. In the second place program (paq8hp12any), the grouping was done mostly manually. In the top program (durilca4linux), the grouping was done by clustering in context space. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: The brain does not implement formal logic (was Re: [agi] Where the Future of AGI Lies)
-- Matt Mahoney, [EMAIL PROTECTED] --- On Sat, 9/20/08, Ben Goertzel <[EMAIL PROTECTED]> wrote: >It seems a big stretch to me to call theorem-proving guidance a "language >modeling problem" ... one may be able to make sense of this statement, but >only by treating the concept of language VERY abstractly, differently from the >commonsense use of the word... I mean that for search problems such as theorem proving, solving differential equations, or integration, you look for "similar" problems in the sense of natural language modeling, i.e. related words or similar grammatical structures. We think about symbols in formal languages in fundamentally the same way we think about words and sentences. We learn to associate "x > y" with "y < x" by the same process that we learn to associate "x is over y" with "y is under x". As a formal language, the representation in our brains is inefficient, so we use pencil and paper or computers for anything that requires a long sequence of steps. But it is just what we need for heuristics to guide a theorem prover. To prove a theorem, you study "similar" theorems and try "similar" steps, where "similar" means they share the same or related terms and grammatical structures. Math textbooks contain lots of proofs, not because we wouldn't otherwise believe the theorems, but because they teach us to come up with our own proofs. >But neverthless, I don't think that the current best-of-breed text processing >approaches have much to teach us about AGI. > >To pursue an overused metaphor, to me that's sort of like trying to understand >flight by carefully studying the most effective high-jumpers. OK, you might >learn something, but you're not getting at the crux of the problem... A more appropriate metaphor is that text compression is the altimeter by which we measure progress. --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: The brain does not implement formal logic (was Re: [agi] Where the Future of AGI Lies)
--- On Sat, 9/20/08, Mike Tintner <[EMAIL PROTECTED]> wrote: > Matt: A more appropriate metaphor is that text compression > is the altimeter > by which we measure progress. (1) > > Matt, > > Now that sentence is a good example of general intelligence > - forming a new > connection between domains - altimeters and progress. > > Can you explain how you could have arrived at it by > > A)logic ( incl. Nars or PLN or any other kind) > B)mathematics > > or how you would *understand* it by any means of > > C) text compression, > D) predictive analysis of sentences/texts in Google. > > Can you explain how any of the rational systems, currently > being discussed > here, can be applied to any problem of general intelligence > whatsoever? Certainly. A metaphor is a type of analogy: "intelligence is to compression as flight is to ___?" The general form is "A is to B as C is to X", and solve for X. Roughly, the solution is X = B + C - A where A, B, C, and X are vectors in semantic space, i.e. rows of a matrix M such that M[i,j] is the probability of words i and j appearing near each other (e.g. in the same paragraph or document) in a large text corpus. Variations of this technique very nearly equal human performance on the analogy section of the college SAT exam. http://aclweb.org/aclwiki/index.php?title=SAT_Analogy_Questions The leading contender is latent relational analysis (LRA), which means applying the above equation to a matrix M that has been compressed using singular value decomposition (SVD). SVD consists of factoring the matrix M = USV, where U and V are orthonormal and S is diagonal (the eigenvalues), then tossing out all but the largest elements of S. This allows U and V to be reduced from, say, 2 x 2 to 2 x 200. SVD in effect applies the transitive property of semantic relatedness, the notion that if A is near B and B is near C, then A is near C. Gorrell gives an efficient algorithm for computing the SVD using a neural network. In this example, the network would be 2 x 200 x 2 where U and V are the weight matrices and the retained elements of S are the hidden units. Hidden units are added as training proceeds, such that the size of the weight matrices is approximately the size of the text corpus read so far. http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.60.7961 To answer your question, this is most like (D), predictive analysis of text, and would be a useful technique for text compression. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: The brain does not implement formal logic (was Re: [agi] Where the Future of AGI Lies)
--- On Sat, 9/20/08, Ben Goertzel <[EMAIL PROTECTED]> wrote: >>A more appropriate metaphor is that text compression is the altimeter by >>which we measure progress. >An extremely major problem with this idea is that, according to this >"altimeter", gzip is vastly more intelligent than a chimpanzee or a two year >old child. > >I guess this shows there is something profoundly wrong with the idea... No it doesn't. It is not gzip that is intelligent. It is the model that gzip uses to predict text. Shannon showed in 1950 that humans can predict successive characters in text such that each character conveys on average about 1 bit per character. It is a level not yet achieved by any text compressor, although we are close. (More precise tests of human prediction are needed). Don't confuse ability to compress text with intelligence. Rather, the size of the output is a measure of the intelligence of the model. A compressor does two things that no human brain can do: repeat the exact sequence of predictions during decompression, and ideally encode the predicted symbols in log(1/p) bits (e.g. arithmetic coding). These capabilities are easily implemented in computers and are independent of the predictive power of the model, which is what we measure. I addressed this in http://cs.fit.edu/~mmahoney/compression/rationale.html Now if you want to compare gzip, a chimpanzee, and a 2 year old child using language prediction as your IQ test, then I would say that gzip falls in the middle. A chimpanzee has no language model, so it is lowest. A 2 year old child can identify word boundaries in continuous speech, can semantically associate a few hundred words, and recognize grammatically correct phrases of 2 or 3 words. This is beyond the capability of gzip's model (substituting text for speech), but not of some of the top compressors. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] NARS probability
--- On Sat, 9/20/08, Pei Wang <[EMAIL PROTECTED]> wrote: > Think about a concrete example: if from one source the > system gets > P(A-->B) = 0.9, and P(P(A-->B) = 0.9) = 0.5, while > from another source > P(A-->B) = 0.2, and P(P(A-->B) = 0.2) = 0.7, then > what will be the > conclusion when the two sources are considered together? This is a common problem in text prediction. In general, there is no right answer. You have to determine experimentally what works best. You compute the probability using some method, run it on some test data, and measure the accuracy of your predictions. To give a more concrete example, suppose that A is some context (the last n bytes of text), and B is the event that the next bit is a 1. We get different predictions for different orders (different values of n) which we need to combine. In PAQ1-PAQ3, I count zeros and ones in context A. Call these counts c0 and c1. Then I let P(A-->B) = c1/(c0+c1) and let the confidence (what you call P(P(A-->B)) be c0+c1. To combine them I add up the c0's and c1's and compute SUM c1 / (SUM c0 + SUM c1). I also discovered experimentally that the prediction is more accurate if the counts are weighted by n^2. For example the order 19 context: "the cat caught a mo_" is a better predictor of the next symbol than the order 2 context: "mo_" even though the latter has probably collected more statistics, and therefore has a higher confidence. In PAQ4-PAQ6 I adjust the weights dynamically using gradient descent of coding cost in weight space to reduce prediction error. This can be improved further by using multiple weight tables indexed by a low order context. In PAQ7-PAQ8 I dynamically map each bit history (truncated c0,c1 plus the last bit) to a probability p_i using a table that is adjusted to reduce prediction error when the bit is observed. Then the predictions p_i are combined using a neural network: p = squash(SUM w_i stretch(p_i)) where squash(x) = 1/(1 + exp(-x)) bounds the output to (0, 1), and stretch(x) = ln(x / (1 - x)) is the inverse of squash. (This implicitly gives greater confidence to probabilities near 0 or 1). When actual bit b is observed, the weights are adjusted to reduce the prediction error b - p by gradient descent of coding cost in weight space as follows: w_i := w_i + L stretch(p_i) (b - p) where L is the learning rate, typically 0.001 to 0.005. Again, we can improve this by using multiple weight tables indexed by a low order context. Or you can use multiple neural networks indexed by different order contexts and combine them by linear averaging or another neural network. In PAQ9 I use chains of 2 input neural networks, where one input is the previous prediction from the next lower context order and the other input is fixed. The weight table is selected by the bit history in the next higher context. This method is still experimental. It works well for simple n-gram models but worse than PAQ8 when there are large numbers of approximately equally good predictions to combine, such as when semantic (cat ~ mouse) and other contexts are added. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: The brain does not implement formal logic (was Re: [agi] Where the Future of AGI Lies)
--- On Sat, 9/20/08, Pei Wang <[EMAIL PROTECTED]> wrote: > Matt, > > I really hope NARS can be simplified, but until you give me the > details, such as how to calculate the truth value in your "converse" > rule, I cannot see how you can do the same things with a simpler > design. You're right. Given P(A), P(B), and P(A->B) = P(B|A), you could derive P(A|B) using Bayes law. But you can't assume this knowledge is available. > For your original claim that "The brain does not > implement formal > logic", my brief answers are: > > (1) So what? Who said AI must duplicate the brain? Just > because we cannot image another possibility? It doesn't. The problem is that none of the probabilistic logic proposals I have seen address the problem of converting natural language to formal statements. I see this as a language modeling problem that can be addressed using the two fundamental language learning processes, which are learning to associate time-delayed concepts and learning new concepts by clustering in context space. Arithmetic and logic can be solved directly in the language model by learning the rules to convert to formal statements and learning the rules for manipulating the statements as grammar rules, e.g. "I had $5 and spent $2" -> "5 - 2" -> "3". But a better model would deviate from the human model and use an exact formal logic system (such as calculator) when long sequences of steps or lots of variables are required. My vision of AI is more like a language model that knows how to write programs and has a built in computer. Neither component requires probabilistic logic. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: The brain does not implement formal logic (was Re: [agi] Where the Future of AGI Lies)
--- On Sun, 9/21/08, Ben Goertzel <[EMAIL PROTECTED]> wrote: >Hmmm I am pretty strongly skeptical of intelligence tests that do not >measure the actual functionality of an AI system, but rather measure the >theoretical capability of the structures or processes or data inside the >system... > >The only useful way I know how to define intelligence is **functionally**, in >terms of what a system can actually do ... > >A 2 year old cannot get itself to pay attention to predicting language for >more than a few minutes, so in a functional sense, it is a much stupider >language predictor than gzip ... Intelligence is not a point on a line. A calculator could be more intelligent than any human, depending on what you want it to do. Text compression measures the capability of a language model, which is an important, unsolved problem in AI. (Vision is another). I'm not building AGI. (That is a $1 quadrillion problem). I'm studying algorithms for learning language. Text compression is a useful tool for measuring progress (although not for vision). -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
Re: [agi] Re: AGI for a quadrillion
--- On Sun, 9/21/08, Vladimir Nesov <[EMAIL PROTECTED]> wrote: > > I'm not building AGI. (That is a $1 quadrillion > problem). > > > > How do you estimate your confidence in this assertion that > developing > AGI (singularity capable) requires this insane effort (odds > of the bet > you'd take for it)? This is an easily falsifiable > statement, if a > small group implements AGI, you'll be proven wrong. As I explained my estimate earlier, that is the cost of automating the economy, which AGI could presumably do (by doing all of our work for us). There is a tradeoff between spending more to have AGI sooner or spending less and waiting. The optimal point depends on current interest rates, and can be found by dividing the world GDP (US $66 trillion in 2006) by this rate. My estimate is independent of technology. However, one possibility is that AGI is unsolved because it requires too much computing power, so it will be solved by Moore's Law. (A human brain sized neural network requires 10^15 bits of memory and 10^16 OPS). But even if the hardware were free, the cost of software and/or training is not dropping exponentially. An organization is most efficient when its members specialize, which means each member has to learn to do its job. You have to choose between a direct cost in training or an indirect cost in additional mistakes. Some of the training knowledge is already on the internet but most is still in our heads. Transferring your knowledge to a machine will take a long time. The speed of human language is the same for input and output. Another possibility is that we will discover some low cost shortcut to AGI. Recursive self improvement is one example, but I showed that this won't work. (See http://www.mattmahoney.net/rsi.pdf ). So far no small group (or even a large group like Google) has produced AGI, in spite of efforts in this direction since the 1950's. In fact, there has been very little theoretical or practical progress since 1965. It seems like if there was a simple way to do it, we would have figured it out by now. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69 Powered by Listbox: http://www.listbox.com
