John, As you say the hardware is just going to get better and better. In five years the PC's of most of the people on this list will probably have at least 8 cores and 16 gig of ram.
But even with a current 32 bit PC with say 4G of Ram you should be able to build an AGI that would be a meaningful proof of concept. Lets say 3G is for representation, at say 60 bytes per atom (less than my usual 100 bytes/atom because using 32bit pointers), that would allow you roughly 50Million atoms. Over 1 million seconds (very roughly two weeks 24/7) that would allow an average of 50 atoms a second of representation. Of course your short term memory would record at a much higher frequency, and over time more and more of your representation would go into models rather than episodic recording. But as this happened the vocabulary of patterns would grow and thus one atom, on average would be able to represent more. But it seems to me such an AGI should be able to have meaningful world knowledge about certain simple worlds, or certain simple subparts of the world. For example, it should be able to have a pretty good model for the world of many early video games, such as pong and perhaps even pac-man (Its been so long since I've seen pac-man I don't know how complex it is, but I am assuming 50 million atoms, many of which, over time, would represent complex patterns, would be able to catch most of the meaningful generalizations of pac-man including its control mechanisms and the results they occur). Is I said in an earlier email, if we want AGI-at-Home to catch on it would be valuable to think of some sort of application that would either inspire through importance or entice by usefulness or amusement to cause people let it use a substantial part of their machine cycles. You mention an interest in intelligent indexing. Of course, hierarchical memory provides a fairly good from of intelligent indexing, in the sense that it automatically promotes indexing through learned combinations of indicies, and can be easily made to have probabilistic and importance weights on its index links to more efficiency allocate index activations. How does your intelligent indexing work? Ed Porter -----Original Message----- From: John G. Rose [mailto:[EMAIL PROTECTED] Sent: Tuesday, December 04, 2007 2:17 PM To: agi@v2.listbox.com Subject: RE: Hacker intelligence level [WAS Re: [agi] Funding AGI research] > From: Ed Porter [mailto:[EMAIL PROTECTED] > John, > > I am sure there is interesting stuff that can be done. It would be > interesting just to see what sort of an agi could be made on a PC. Yes it would be interesting to see what could be done on a small cluster of modern server grade computers. I like to think about the newer Penryn 45nm, SSE4, quadcore quadproc servers with lots of FB DDR3 800mhz RAM running 64 bit OS (sorry I prefer coding in Windows) using standard gigabit Ethernet quad NICs, with solid state drives, and 15,000 RPM SAS for the slower stuff, and a take maybe 10 of these servers. There HAS to be enough resource there to get some small prototype going. And look at next year's 8 core Nehalem procs coming out... Interserver messaging should make heavy use of IP multicasting. Then another messaging channel with the new USB 3.0... Supposedly USB 3.0 is 4.8 gigabits. > I would be interested in you Ideas for how to make a powerful AGI > without a > vast amount of interconnect. The major schemes I know about for > reducting > interconnect involve allocating what interconnect you have to the links > with > the highest probability or importance, varying those measures of > probability > and importance in a contest specific way, and being guided by prior > similar > experiences. Well I actually don't have the theory far enough to calculate interconnect metrics. But I try to minimize that through storage structure. What gets stored, how it gets stored, where it's stored, how systems are modeled, what a model is, what a system of models are, how systems of models are stored,.. don't store dupes, store diffs... mixing code and data, collapsing data into code, what is code and what is data? Basically a lot of intelligent indexing, like real intelligent indexing... I'm working on using CA's as universal symbolistic indexors and generators - IOW exploring a theory of uncalculated precalcs for computational complexity indexing using CA's in order to control uncertainty and manage complexity... Lots of addicting brain candy stuff... John ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?& ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=8660244&id_secret=71938578-e534ed