My perspective on grounding is partially summarized here www.goertzel.org/papers/*PostEmbodied*AI_June7.htm
-- Ben G On Mon, Aug 4, 2008 at 6:33 PM, Matt Mahoney <[EMAIL PROTECTED]> wrote: > My opinion on grounding is that it depends on the application. I have > argued in > http://cs.fit.edu/~mmahoney/compression/rationale.html<http://cs.fit.edu/%7Emmahoney/compression/rationale.html>that > text compression is at least as hard as passing the Turing test. > (Predicting text requires lots of real-world knowledge). This is > (non-grounded) AI as defined by Turing. However, some may argue this is not > AGI because a language model cannot see or control a robot. > > This is an engineering question. Is it easier to teach a system that the > sky is blue using pure text based I/O or by adding vision and embodiment? If > it is text based, is it easier to use statistics (e.g. "blue sky" returns > more Google hits than "red sky"), or is it easier to explicitly encode the > knowledge? > > An engineering problem starts with a specification, not a design. Do we > want text based AI or do we want a robot? What is the problem we are trying > to solve? > > AGI has the potential to replace nearly all human labor worldwide, which is > valued at US $2 to $5 quadrillion over the next 30 years. (Currently $66 > trillion per year and increasing 5% annually). We should expect the cost of > AGI to be of this magnitude. It is what we are willing to pay to get it now. > Note this is a different problem than building artificial human brains. We > do need to solve the language and vision problems in order for machines to > do certain types of work that currently must be done by humans. This does > not mean building robots that look or act like humans. > > I believe that the problem can be solved using faster hardware and > otherwise mature technology, consisting of lots of narrow-AI specialists, a > protocol that links them together (routing messages to the right experts), > and an economy that rewards the most useful experts and routers. In > http://www.mattmahoney.net/agi.html I proposed a P2P protocol called > competitive message routing (CMR). In CMR, each peer understands a subset of > natural language relevant to its domain of expertise and learns about > experts in related topics in a hostile environment. Peers are administered > by humans who have an incentive to gain the trust of other peers and provide > useful services in exchange for not having their outgoing messages blocked. > I believe the resulting network will be a powerful and useful intelligence > that nobody would mistake for human. It would be intelligent in the sense > that a calculator or Google is intelligent, but far more useful. > > An alternative approach is recursive self improvement (RSI). If humans can > produce superhuman AGI, then so can those agents, launching a singularity. I > don't believe that will happen because I don't believe that for any level > and reasonable definition of intelligence of the agent's choosing that it is > possible for it to produce a more intelligent agent. We currently lack > software and mathematical models of RSI, even in very restricted or simple > environments. There are no known problems that are provably hard to solve > but easy to verify (e.g. factoring, NP-complete problems, or cryptographic > puzzles) that an AI could use to test its children. Furthermore, humans and > other animals do not recognize higher intelligence than themselves, for > example, we can measure an IQ of 200 in children but not adults. However, I > have no proof that RSI is impossible either. > > Even without RSI, CMR is not without risks. Agents compete for resources > and could still make modified copies of themselves. This is an evolutionary > environment where fitness does not equal usefulness or friendliness. It > could still produce a singularity in the sense of greatly accelerated > evolution where humans are no longer the dominant intelligence. We would > observe the CMR protocol quickly changing from natural language to something > incomprehensible, and we would no longer know what our computers were doing. > I don't have a good solution to this problem. > > -- 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 > -- Ben Goertzel, PhD CEO, Novamente LLC and Biomind LLC Director of Research, SIAI [EMAIL PROTECTED] "Nothing will ever be attempted if all possible objections must be first overcome " - Dr Samuel Johnson ------------------------------------------- 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
