PEI'S SELF-DEFEATING LOOP WILL PROBABLY BE BROKEN WITHIN 3 TO 8 YEARS Over then next 3 - 8 years there probably will arise from the level of AI and AGI projects being funded an ever increasing appreciation and proof of the power, generality, and potential of AGI --- enough so that funding of large AGI projects will start.
This will be aided by the growing increase in the number of people who believe in AGI --- the increasing organization of the AGI community, as represented by the recent AGI 2008 conference and the planed AGI 2009 --- the growing knowledge and tools, such as OpenCog tools, for building AGI projects --- and, perhaps most importantly, the rapidly dropping cost and rapidly rising power of hardware. Today a PC with 4G or RAM should be able to demonstrate some important pieces of the AGI problem. For under $40k --- money many funded projects can afford --- you can buy a 4 processor quad-core server with 256GB DRAM and many TBs of disk. With that you should be able to demonstrate even more of AGI's potential. But in about 5 years things should really start changing with the arrival of the many-core, many-layer chips and the operating systems for efficiently using them (with the help of people like Sam Adams). We should begin to see chips with 256 cores --- connected by a high-speed on chip mesh network --- with each core having fat thru-wafer paths to memory totaling --- for the whole chip --- roughly 8GBytes of DRAM --- plus a GByte of embedded DRAM for L2 cache. All this will fit in one multi-layer "chip." Each such chip will have --- say --- 1TByte of external bandwidth in the form of 128 different 8GByte/sec channels to memory or other processors. I'm guessing such chip will probably sell for a very fat premium at introduction --- say $5K --- but will be under $2K within two years. In 5 years a TByte of DRAM should cost about $10K. This means by 2013 you could have a system with such a chip and 1TB of DRAM for under $20K --- cheap enough for most teams of AI grad students to share one. Such a system would crank roughly 1TOpp/sec and enable 4G random accesses in its 1TB of DRAM/sec. That allows a fair amount of connectionism to be computed --- more than 1000 times as much as current PCs. Such a >$20K system should be a powerful enough testbed to develop and test much of the basic architecture for artificial minds. And they should be cheap enough that within the next six or seven years hundreds of AGI teams could test and perfect their approaches on them. Such testbeds could come in a range of different sizes --- including one with 256 such 256 core chips and 50TBytes of DRAM. This would include 2TB of on-chip DRAM and 256G of L2 cache. Such hardware could provide 100 TOpps --- 512G random accesses/sec in the 50TB of DRAM --- and a theoretical cross-sectional bandwidth between chips of about 512G global 64byte messages/sec (with a much higher total number of interchip messages/sec, since a many messages would be between near by chips). This computational, representational, and communication capacity is very possibly enough to create human level AGI. In 7 years such hardware could cost roughly $1m in parts, which means it probably could be profitable be sold for under $2m if it was being sold in any quantity. And understand these system would be good for general scientific, data base, virtual reality, and advanced web-based programming as well --- so they should sell in quantity --- and their price should come down rapidly. At these price every major university and research lab could afford several of them. Within 10 years from today there could be thousands such roughly human level machines. The combination of increasingly interest in AGI, increasing understand of it, increasingly powerful AGI tools --- and more powerful hardware --- all will combine to make it highly likely that within 3 to 8 years there will have been enough progress that it will become obvious that there is tremendous strategic and economic value for investing in it big time --- and then the big money will come in the hundreds of millions or billions of dollars. So I think Pei's self defeating "strange loop of AGI funding" holds in the short term --- but also that in less than a decade this self-defeating loop will --- itself --- be defeated. -----Original Message----- From: Pei Wang [mailto:[EMAIL PROTECTED] Sent: Friday, April 18, 2008 1:02 PM To: [email protected] Subject: [agi] The Strange Loop of AGI Funding: now logically proved! PREMISES: (1) AGI is one of the most complicated problem in the history of science, and therefore requires substantial funding for it to happen. (2) Since all previous attempts failed, investors and funding agencies have enough reason to wait until a recognizable breakthrough to put their money in. (3) Since the people who have the money are usually not AGI researchers (so won't read papers and books), a breakthrough becomes recognizable to them only by impressive demos. (4) If the system is really general-purpose, then if it can give an impressive demo on one problem, it should be able to solve all kinds of problems to roughly the same level. (5) If a system already can solve all kinds of problems, then the research has mostly finished, and won't need funding anymore. CONCLUSION: AGI research will get funding when and only when the funding is no longer needed anymore. Q.E.D. :-( Pei ------------------------------------------- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?&; Powered by Listbox: http://www.listbox.com ------------------------------------------- agi Archives: http://www.listbox.com/member/archive/303/=now RSS Feed: http://www.listbox.com/member/archive/rss/303/ Modify Your Subscription: http://www.listbox.com/member/?member_id=8660244&id_secret=101455710-f059c4 Powered by Listbox: http://www.listbox.com
<<attachment: winmail.dat>>
