Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!bcm!dimacs.rutgers.edu!aramis.rutgers.edu!athos.rutgers.edu!nanotech From: mmt@client2.DRETOR (Martin Taylor) Newsgroups: sci.nanotech Subject: Re: Is this stuff for real? Keywords: reality nanotech questions Message-ID: Date: 8 Mar 91 21:35:18 GMT Sender: nanotech@athos.rutgers.edu Organization: Defence and Civil Institute of Environmental Medicine Lines: 41 Approved: nanotech@aramis.rutgers.edu I think that much of the argument as to whether nanobots might be likely to mutate into grey goo hinges on a difference of opinion about the underlying structure of a successful nanobot. One designed to do "exactly what we want (we hope) rather than relying on chance" is likely to be built, shall we say, symbolically. and rely on rules with truth values near 1 or 0. The design space is large, and viable designs few and far between. JoSH's arguments apply pretty well to such machines (even though theoretically the chance of an error leading to a new viable design can never be reduced to zero). But it is unlikely that we will know "exactly what we want" the machine to do, and even if we did, we would probably want it to do something quite similar if it was confronted with circumstances very like those we envisaged in our designing. A machine with these desirable abilities would be in a design space that (at least locally) was rather dense with viable machines, and the probability of a mutation leading to a viable design could be appreciably different from zero. If a mutated machine propagated its design better than the original did, then it has at least made the first step toward grey goo. I'm not sure that the argument is as clear-cut as either the worriers or JoSH make it out to be, but I am sure that it is better to err on the side of prudence, and think very carefully about all the trade-offs between behavioural flexibility (topological neighbourhoods likely to contain viable points), design rigidity (enpty neighbourhoods but probably ineffective machine), and mutability. I know that there is not a LOGICAL connection here, but there is a probable linkage if designs are not well thought out. For example, with what we know now, behavioural flexibility is likely to be attained through the use of distributed representations for the perceptual-behavioural knowledge and the incorporation of trainability. But a design of this kind which replicated itself would be very likely to produce a working descendant if some mutation altered the form of the network. It would just do something a little different. I think here we have a situation much closer to that of natural evolution than is envisaged by the "clean design" school. -- Martin Taylor (mmt@ben.dciem.dnd.ca ...!uunet!dciem!mmt) (416) 635-2048 To be a fundamentalist takes considerable flexibility of mind.