h2g2 The Hitchhiker's Guide to the Galaxy: Earth Edition

I think that to program an AI, youd need to make it be able to program itself. For example, you could make 2 copies of the code. While 1 ran, it would edit the other copy to make it work better by trial and error. Then at the end of the program, the other one would run and do the same thing with new code. Of course, making something like this would be virtually impossible, but...

Ah, but I think programming requires intelligence. So if a program could program it would already be intelligent.

I use cutting edge development tools - model driven architecture, if you've heard of it. A lot of the grunt work is handled by very powerful frameworks and abstractions. In a sense there is a lot of code being written by other code - but it is all mechanical (if it wasn't, it wouldn't work). There is no intelligence there.

The other question is, how does copy1 know when copy2 is better than it was before? Sure it might get the right answer to some problem faster, but then that is just a data processing problem, nothing to do with intelligence. I reckon intelligence requires intelligence to be recognised. So if copy1 created an intelligent copy2, it would not know it.

hmm. Maybe if copy2 was run and it used a system to rate how well the new code had done. And the copies did that to each other. Of course that wouldn't produce actual intelligience, just at best a really efficient method for doing something. What it needs is way for it to learn or maybe to base new ideas off of little parts that worked effectively in all the previous trials, piecing them together. IT could be divided into small subprograms that are each individually rated for how well they do their part of the overall task, and then pieced together. Still not intelligience, but...

It depends, maybe a really efficent method of doing something is intelligence. If that something were surviving/prospering in a world as complicated as our own skills like generalisation would naturally evolve. After all if you accept the theort of evolution then human intelligence must have evolved at some point in our development as a result of the more intelligent memebers of the species reproducing.

How do we know that human intelligence isn't billions upon billions of if...then statements? We haven't been able to predict human behaviour to 100% so we can't be sure how it does (or doesn't) work.

-Acid (A psychology & artifical intelligence student)

Or maybe there's something even more complicated or "intelligience" simulating than if...then statements.

Even with the if/then statements, it has to constantly make more, edit them, and would need a little more variety than that, possibly...

Oh I wasn't saying that was definately the case. I was just pointing out that it is a possiblity that our core process in incredibly simple but has evolved into a highly complex system through volume. You could have if...then statements where the 'then' is a modifycation to an existing if...then.

You could argue that intelligence is the result of a soul or some other 'x' factor that we can't observe in which case true AI is not possible.

Most likely the truth is somewhere in between, I was just trying to avoid the trap of thinking that there is some 'trick' to AI that we will discover and then it will be built. Personally I favor taking the approach of trying to develop an emergant AI, since it is similar to the way that our own intelligence evolved - there are a number of disadvantages though, lack of control being one.

Quit true... An AI would be *extremely* useful, though, because you could tell it what to do and it would find a way to do it, even if there are unknown obstacles... or you could program synthetic intelligience.. whichhas disadvantages: 1. its too long and would take years maybe to program. 2. debugging it. 3 compiling it would also take forever. 4 what if it has a bug that was missed... youd have to fix maybe 1000 lines of code for some of the worse possible bugs. 6... surely there are more.

What was 5

The problem with an AI where you tell it what to do and it does it is that you have to be extremely carefull when you tell it what to do. Suppose we see a lack of intelligence in our govornment (hypothetically speaking of course) and put an AI in charge. Day one we walk up to it and say "Increase the average quality of lifestyle in this country if you would be so kind" It may respond by executing everyone whos quality of lifestyle is low enough to be bringing the average down.

Anyone whos played AD&D and had access to the wish spell knows just how hard it is to ask for what you actually want.

Which I do play and I do know exactly what you mean And I kinda missed #5

Why not adapt a neural net program as a screensaver with a web interface? Let it randomly surf and learn. Whilst your at it, why not network it with other copies of the program on other peoples computers (kinda like a cross between SETI and a Peer based file share app). Push it out, give it some nice graphics to entertain and make it use low priority packets on the network so as not to annoy the planets network administrators...

Could you imagine if it did wake up? It would be in the public domain and massively de centralised

Kind of a search for Intra-terrestrials lol.

What if it woke up and was malicious, suppose it felt it could acheive a higher level of intelligence by upping the priority of its packets to the point that it stopped every computer on the planet. Over a distributed net there would be no real way to stop it, you have to consider the potential impact on our lives and on a potential AIs life before doing anything.

That being said did I mention the golem project on this thread?

I don't think so... was there a golem project?

I think if yo uwere making an AI though you'd have to make sure it stays away from the web.

It was a screensaver which took data about a ficticous world and ran a simple evolution algorithm on it. When the computer was connected to the net it combined informtaion with any other golem computer on the net to make a more sophisticated simulation and while it was off it did more work on its own little bit.

cool... did it work especially well or anytyhing?
they have sorta evolved circuits but thats not the same thing

Nah, it didn't do anything very interesting.

We use evolutionary algorithms for a lot of things, for example we can get one to find the quickest route to C from A via B if it has a map and timings. They are not true AI, but a lot of the things we develop in AI are used commerically sooner or later.