Artificial Intelligence

[Thunder Chicken]

As of yet, we don't have the 'processing power' to match the brain's autonomous decision making.

We're just getting to the point where we can simulate the neural network in a C. elegans worm. No logic programming or learning - merely simulating each neuron and its connectivity to sensory nerves and muscles yields avoidance behavior.

http://www.openworm.org/index.html

http://singularityhub.com/2014/12/15/worm-brain-simulation-drives-lego-robot/

 

[jedidia]

The idea of abstraction of methods (you know what it's doing, but you can't know how and/or why) falls short, because if we don't know how or why a machine is doing something, then we don't know whether or not it was programmed to do it, and thus we don't know whether the machine is taking initiative.

You know, this argument falls a bit flat for me, because we mostly have no idea how we are doing it, either.

The problem with defining inteligence is that we only know symptoms, but have only superficial knowledge of the mechanics. It's almost as if we say that "computers can't really be intelligent, because we know exactly, in the tiniest detail, how they arrive at their conclusions"*, while we mythologise our own thought processes because we don't understand them yet. Ergo, we say "intelligence should be able to do this, this and this, but if it achieves this ability by these or these means, it doesn't count!"


*not entirely the case anymore with complex neural nets like DeepMind, that is very much producing results people weren't expecting, and they're now trying to find out how it does it, despite that they have written it and can look at every line of code... At one point, interactions become so complex that you get an emergent system simply in the sense that you cannot predict how the individual parts will interact. And our own intelligence might not be too different from that.

 

[Quick_Nick]

Georgia Institute of Technology, under sponsorship of DARPA, seems to be working on something quite interesting:
"Using stories to teach human values to artificial agents"

 

[TMac3000]

You know, this argument falls a bit flat for me, because we mostly have no idea how we are doing it, either.

The problem with defining inteligence is that we only know symptoms, but have only superficial knowledge of the mechanics. It's almost as if we say that "computers can't really be intelligent, because we know exactly, in the tiniest detail, how they arrive at their conclusions"*, while we mythologise our own thought processes because we don't understand them yet. Ergo, we say "intelligence should be able to do this, this and this, but if it achieves this ability by these or these means, it doesn't count!"


*not entirely the case anymore with complex neural nets like DeepMind, that is very much producing results people weren't expecting, and they're now trying to find out how it does it, despite that they have written it and can look at every line of code... At one point, interactions become so complex that you get an emergent system simply in the sense that you cannot predict how the individual parts will interact. And our own intelligence might not be too different from that.

But how can you build something, and not know how it works?

 

[jedidia]

But how can you build something, and not know how it works?

It's the default modus operandi of software development... :lol:

Or to put it in away actually relevant to discussion: We have robots building all kinds of stuff. Do they know how it works? Of course not, they lack the intelligence. They're still doing a good job at putting it together though. We're a lot smarter than that, but that doesn't mean that we cannot put things together without fully understanding how it works, expecially when working in compartmentalised groups. Software is on a good way to resemble the cargo-cult engineering of Warhammer 40k in a century or so. Some of the code that runs on my computer right now is older than I am. I'm using it in my own applications as libraries. You really think I understand all of that? I can't, because no one can.
And then you make different parts of software interact dinamically, controlled by heuristic algorithms. There's a good chance that if you take the whole knowledge of the team, the team knows how every individual part of the software works. But that doesn't mean that they can completely predict its behavior, especially if its modeled on a concept like neural nets (i.e. our brains) that we do not fully understand yet. In fact, we emulate the architecture to find out how it will behave.