AW: AW: Defining AGI (was Re: AW: [agi] META: A possible re-focusing of this list)
In my opinion, the domain of software development is far too ambitious for the first AGI. Software development is not a closed domain. The AGI will need at least knowledge about the domain of the problems for which the AGI shall write a program. The English interface is nice but today it is just a dream. An English interface is not needed for a proof of concept for first AGI. So why to make the problem harder as it already is? The domain of mathematics is closed but can be extended by adding more and more definitions and axioms which are very compact. The interface could be very simple. And thus you can mainly concentrate to build the kernel AGI algorithm. -Matthias Trent Waddington [mailto:[EMAIL PROTECTED] wrote Yes, I'd want it to have an English interface.. because I'd also expect it to be able to read comments and commit messages in the revision control and the developer mailing lists, etc. Open Source programmers (and testers!) are basically disembodied but they get along just fine. I'd also expect it to be able to see windows and icons and all those other things that are part of software these days. I wouldn't expect it to be able to test a program and say whether it was working correctly or not if it couldn't even see it running and interact with it. Of course, if you're testing command line apps you could get away with a much simpler sensor. Trent --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
RE: AW: Defining AGI (was Re: AW: [agi] META: A possible re-focusing of this list)
From: Dr. Matthias Heger [mailto:[EMAIL PROTECTED] In my opinion, the domain of software development is far too ambitious for the first AGI. Software development is not a closed domain. The AGI will need at least knowledge about the domain of the problems for which the AGI shall write a program. The English interface is nice but today it is just a dream. An English interface is not needed for a proof of concept for first AGI. So why to make the problem harder as it already is? English is just the gel that the knowledge is embedded in. Sorting out that format is bang for the buck. And it is just symbology as math is symbology, or representation. The domain of mathematics is closed but can be extended by adding more and more definitions and axioms which are very compact. The interface could be very simple. And thus you can mainly concentrate to build the kernel AGI algorithm. Mathematics effectively is just another gel that the knowledge is stored in. It's a representation of (some other wierd physics stuff that I won't bring up). I think I can say that, that math is just an instantiation of something other. Unless the actual math symbology is the math and not what it represents. Either way, all will be represented in binary, for software or an electronics based AGI. How can you get away from the coupling of math and software? Unless there is some REAL special sauce like some analog based hyperbrid. Loosemore would say that the coupling breaks somehow at complexity regions. And I think that the representation of reality has to include those. John --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: AW: Defining AGI (was Re: AW: [agi] META: A possible re-focusing of this list)
Matt, I know this is a late response, but your statement there is odd. The halting problem is not algorithmically describable at all. It *does* have a simple description, but not an algorithmic one (ie, not one that can be completely captured by axioms). That is the whole point! -Abram On Wed, Oct 15, 2008 at 8:35 PM, Matt Mahoney [EMAIL PROTECTED] wrote: --- On Wed, 10/15/08, Dr. Matthias Heger [EMAIL PROTECTED] wrote: Text compression would be AGI-complete but I think it is still too big. The problem is the source of knowledge. If you restrict to mathematical expressions then the amount of data necessary to teach the AGI is probably much smaller. In fact AGI could teach itself using a current theorem prover. Goedel and Turing showed that theorem proving is equivalent to solving the halting problem. So a simple measure of intelligence might be to count the number of programs that can be decided. But where does that get us? Either way (as as set of axioms, or a description of a universal Turing machine), the problem is algorithmically simple to describe. Therefore (by AIXI) any solution will be algorithmically simple too. If you defined AGI this way, what would be your approach? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
AW: Defining AGI (was Re: AW: [agi] META: A possible re-focusing of this list)
Text compression would be AGI-complete but I think it is still too big. The problem is the source of knowledge. If you restrict to mathematical expressions then the amount of data necessary to teach the AGI is probably much smaller. In fact AGI could teach itself using a current theorem prover. -Matthias Matt Mahoney wrote I have argued that text compression is just such a problem. Compressing natural language dialogs implies passing the Turing test. Compressing text containing mathematical expressions implies solving those expressions. Compression also allows for precise measurements of progress. Text compression is not completely general. It tests language, but not vision or embodiment. Image compression is a poor test for vision because any progress in modeling high level visual features is overwhelmed by incompressible low level noise. Text does not have this problem. Hutter proved that compression is a completely general solution in the AIXI model. (The best predictor of the environment is the shortest model that is consistent with observation so far). However, this may not be very useful as a test, because it would require testing over a random distribution of environmental models rather than problems of interest to people such as language. -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: AW: Defining AGI (was Re: AW: [agi] META: A possible re-focusing of this list)
--- On Wed, 10/15/08, Dr. Matthias Heger [EMAIL PROTECTED] wrote: Text compression would be AGI-complete but I think it is still too big. The problem is the source of knowledge. If you restrict to mathematical expressions then the amount of data necessary to teach the AGI is probably much smaller. In fact AGI could teach itself using a current theorem prover. Goedel and Turing showed that theorem proving is equivalent to solving the halting problem. So a simple measure of intelligence might be to count the number of programs that can be decided. But where does that get us? Either way (as as set of axioms, or a description of a universal Turing machine), the problem is algorithmically simple to describe. Therefore (by AIXI) any solution will be algorithmically simple too. If you defined AGI this way, what would be your approach? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: AW: Defining AGI (was Re: AW: [agi] META: A possible re-focusing of this list)
On Thu, Oct 16, 2008 at 1:35 AM, Matt Mahoney [EMAIL PROTECTED] wrote: Goedel and Turing showed that theorem proving is equivalent to solving the halting problem. So a simple measure of intelligence might be to count the number of programs that can be decided. But where does that get us? Either way (as as set of axioms, or a description of a universal Turing machine), the problem is algorithmically simple to describe. Therefore (by AIXI) any solution will be algorithmically simple too. This doesn't follow. Per Chaitin, you can't prove a 20 pound theorem with 10 pounds of axioms, even if the formal system being discussed is algorithmically simple. More to the point, a program that can prove interesting theorems before the sun burns out, will be much more complex than the axiom system. --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
AW: AW: Defining AGI (was Re: AW: [agi] META: A possible re-focusing of this list)
My intention is not to define intelligence. I choose mathematics just as a test domain for first AGI algorithms. The reasons: 1. The domain is well understood. 2. The domain has regularities. Therefore a high intelligent algorithm has a chance to outperform less intelligent algorithms 3. The domain can be modeled easily by software. 4. The domain is non-trivial. Current algorithms fail for hard problems in this domain because of the exponential growing complexity. 5. The domain allows a comparison with performance of human intelligence. To decide, whether you have an AGI or not, you also have to evaluate the proofs and not only the fact whether it could prove something or not. For example, the formula 1+2+3+...+n = 0.5*n*(n+1) can be proven by seeing a pattern: 1 + 2+ 3 + ... + n-2 + n-1 + n + n + (n-1) + (n-2) + .. + 3+ 2+ 1 = (n+1) + (n+1) + (n+1) + ... +(n+1) + (n+1) + (n+1) = n*(n+1) AGI will differ from AI by often using such pattern based proofs. An AGI based theorem prover represents expressions by patterns. When it comes to prove a certain formula, patterns of known expressions and rules become active or inactive. -Matthias Matt Mahoney wrote: Goedel and Turing showed that theorem proving is equivalent to solving the halting problem. So a simple measure of intelligence might be to count the number of programs that can be decided. But where does that get us? Either way (as as set of axioms, or a description of a universal Turing machine), the problem is algorithmically simple to describe. Therefore (by AIXI) any solution will be algorithmically simple too. If you defined AGI this way, what would be your approach? -- Matt Mahoney, [EMAIL PROTECTED] --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?; Powered by Listbox: http://www.listbox.com --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: AW: Defining AGI (was Re: AW: [agi] META: A possible re-focusing of this list)
On Thu, Oct 16, 2008 at 12:50 PM, Dr. Matthias Heger [EMAIL PROTECTED] wrote: The reasons: 1. The domain is well understood. 2. The domain has regularities. Therefore a high intelligent algorithm has a chance to outperform less intelligent algorithms 3. The domain can be modeled easily by software. 4. The domain is non-trivial. Current algorithms fail for hard problems in this domain because of the exponential growing complexity. 5. The domain allows a comparison with performance of human intelligence. For all the same reasons I advocate program testing, writing and debugging as a good experimental domain for AGI. It also has the added reason: 6. You can make a big fat bucket load of money doing it. Which is not the case for mathematics. And: 7. The AGI can help you test, write and debug itself. And typically when I bring this up I get the blank stares as people figure out how much work making such an AGI would be. Except from some people who just think that programming is something only an advanced AGI could do and advocate that the way to can get from a proto-AGI to an advanced AGI is via embodied learning in a simulated environment. Because, apparently, knowing things about space and motion (aka playing fetch) is important when programming.. of which I'm not terribly convinced. I believe the same has been said about doing mathematical proofs. Oh yeah, and there's some others who think it wouldn't be so hard, so long as we're talking about a programming language that is easy for an AGI to manipulate. *cough*LISP*cough* They'd have us believe that it would be significantly harder to develop an AGI that can do work that is actually marketable.. Trent --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com
Re: AW: Defining AGI (was Re: AW: [agi] META: A possible re-focusing of this list)
On Thu, Oct 16, 2008 at 2:05 PM, charles griffiths [EMAIL PROTECTED] wrote: You have a point, but how would you propose giving specifications to the AGI-programmer? Teach it English? Always have it modify working programs with the same objectives (e.g., reduce runtime/memory, avoid crashing on any input) in mind, like some kind of AGI-compiler? Yes, I'd want it to have an English interface.. because I'd also expect it to be able to read comments and commit messages in the revision control and the developer mailing lists, etc. Open Source programmers (and testers!) are basically disembodied but they get along just fine. I'd also expect it to be able to see windows and icons and all those other things that are part of software these days. I wouldn't expect it to be able to test a program and say whether it was working correctly or not if it couldn't even see it running and interact with it. Of course, if you're testing command line apps you could get away with a much simpler sensor. Trent --- agi Archives: https://www.listbox.com/member/archive/303/=now RSS Feed: https://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: https://www.listbox.com/member/?member_id=8660244id_secret=117534816-b15a34 Powered by Listbox: http://www.listbox.com