--- On Sat, 12/6/08, Steve Richfield <[EMAIL PROTECTED]> wrote: > Internet AGIs are the technology of the future, and always will be. There > will NEVER EVER in a million years be a thinking Internet silicon > intelligence that will be able to solve substantial real-world problems based > only on what exists on the Internet. I think that my prior email was pretty > much a closed-form proof of that. However, there are MUCH simpler methods > that work TODAY, given the metadata that is presently missing from the > Internet.
The internet has about 10^14 to 10^15 bits of knowledge as searchable text. AGI requires 10^17 to 10^18 bits. If we assume that the internet doubles every 1.5 to 2 years with Moore's Law, then we should have enough knowledge in 15-20 years. However, much of this new knowledge is video, so we also need to solve vision and speech along with language. > While VERY interesting, your proposal appears to leave the following > important questions unanswered: > 1. How is it an AGI? I suppose this is a matter of definitions. It looks to > me more like a protocol. AGI means automating the economy so we don't have to work. It means not just solving the language and vision problems, but also training the equivalent of 10^10 humans to make money for us. After hardware costs come down, custom training for specialized roles will be the major expense. I proposed surveillance as the cheapest way for AGI to learn what we want. A cheaper alternative might be brain scanning, but we have not yet developed the technology. (It will be worth US$1 quadrillion if you can do it). Or another way to answer your question, AGI is a lot of dumb specialists plus an infrastructure to route messages to the right experts. > 2. As I explained earlier on this thread, all human-human languages have > severe semantic limitations, such that (applying this to your porposal), only > very rarely will there ever exist an answer that PRECISELY answers a > question, so some sort of "acceptable error" must go into the equation. In > the example you used in your paper, Jupiter is NOT the largest planet that is > known, as the astronomers have identified larger planets in other solar > systems. There may be a good solution to this, e.g. provide the 3 best > answers that are semantically disjoint. People communicate in natural language 100 to 1000 times faster than any artificial language, in spite of its supposed limitations. Remember that the limiting cost is transferring knowledge from human brains to AGI, 10^17 to 10^18 bits at 2 bits per second per person. As for Jupiter, any question you ask is going to get more than one answer. This is not a new problem. http://www.google.com/search?q=what+is+the+largest+planet%3F In my proposal, peers compete for reputation and have a financial incentive to provide useful information to avoid being blocked or ignored in an economy where information has negative value. This is why it is important for an AGI protocol to provide for secure authentication. > 3. Your paper addresses question answering, which as I have explained here > in the past, is a much lower form of art than is problem solving, where you > simply state an unsatisfactory situation and let the computer figure out why > things are as they are and how to improve them. Problem solving pre-dates AGI by decades. We know how to solve problems in many narrow domains. The problem I address is finding the right experts. -- 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=120640061-aded06 Powered by Listbox: http://www.listbox.com
