Re: note on AI
- Posted by gertie at PELL.NET Jan 31, 2001
- 626 views
On 31 Jan 2001, at 11:54, Mike The Spike wrote: > Ok Ok Allright... > I haven't commented on this whole AI discussion > eventhough I am an AI programmer by statute, but I > will now... By statute? > The are ways to do what yall describe, that are much, > much more simple than you might think. > What you need to do 'Imparatic AI', or Learning AI, > doesn't have anything to do with dynamic code > generation. It does if you want the Ai to be able to deal with things you didn't think of before. Take walking for instance. Program in how to walk across your living room. No problem. Then there are these instances of walking: up a mountain, across a ceiling while carrying a table (do not spill the milk in the saucer), down a tree, with broken mechanical systems, underwater on a spinning oil tanker propeller while doing maintenance, across the International Space station with a new jetpack, carrying a cubic yard of water at 60mph thru the forest to fight a forest fire, etc.. > That's what, (sorry Kitty!), lamers that don't know > Jack Shit nor John Diddly Squat about AI think they > should do. People who have not had the time to consider these things in depth usually use the old ways. > Basically, to implement dynamic learning AI, > you don't have to dynamically rewrite your program to > be smarter. > You should set up an infinite state machine (well... > infinite untill memory is used up). Or an off-program interpreter, or a dos driver in the dos redirector interrrupt,, or heck, executing vars with the existing interpreter, so it's easier to get at the main program's variables. > And dynamically generate new states allow with > specifiers that describe the way that state is > handled. But executing a variable, or instituting a new state machine, or changing the weights in a neural net,, pretty much the same thing. It's like what i said about dynamically plugging in new TSRs in dos. <shrug> > Then a single 'learning' algorithm reads new state's > specification and handles a state the way it is > specified. > That's about it. > No need for l33t H4x0r dynacode support, just a smart > algorithm. > > But the reality is, that by implementing AI using a > learning mechanism, isn't going to make for a very > smart AI. While this approach doesn't make for smart humans, it should work fine for any Ai, given the access to the net and the speed of several puters, accessing terabytes of memory online/onharddrive. > You need to combine an AI that understands a basic set > of input *and* is able to learn new stuff. > Otherwise you will have an AI about as smart as a > Tamagotchi, that you'd have to teach new stuff *you* > allready know for about a month before it even > resembles something intelligent. A month? then why has millions of dollars and centuries of people-hours not done it yet? > The best way to implement a super-inteligent learning > AI, is an AI client that frequently downloads new code > for itself from the net. Like an extra set of words it > understands or a plugin that allows it to understand a > new language. Ah, back to exec'ing new code. And plugging it into the existing program too. Sounds like what i said last week. Kat ____________________________________________________________ T O P I C A -- Learn More. Surf Less. Newsletters, Tips and Discussions on Topics You Choose. http://www.topica.com/partner/tag01