New neural networks coming to PA

Have you done anything with neural nets & AIs outside of vidya gaems?

I'd also be interested in it, but I finished my degree quite a while ago, there wasn't really a major AI field then (and neural nets certainly weren't around). It seems that the field is on the edge of a major breakthrough, with neural nets & self-evolving hardware like the FPGA article I linked. Perhaps the only roadblock now is to get an AI to select its own problems to solve, without human input.

Um guys:
http://waitbutwhy.com/2015/01/artificial-intelligence-revolution-1.html
http://waitbutwhy.com/2015/01/artificial-intelligence-revolution-2.html

Basically a summarization of the implications of the ai you describe
(apparently, the experts think that this could be really good or really bad)

While I think its a great Idea, we need to be careful.

People were expecting a revolution in AI back when LISP first emerged.We've generally proven to be incredibly overoptimistic about AI advances.

I am not a careful person, and to be quite frank, EMP technology exists, and I do not see a reason why an EMP would not be sufficient to disable even the largest supercomputers.

the epic race of fusion power and better AIs. Place your bets now.

It's true, but there's still been progress, and there's nothing to suggest that a self-advancing AI isn't possible.

That'd be pretty cool!

I have to imagine training on the scenarios players create would be far better than training on the AI vs itself. Might we actually see the AI countering various meta strategies if it was trained on real player data? Interesting question.

Only if they played hundreds of games against it.

is dozens good enough?

The neural networks DO NOT dictate if the AI follows the meta. They do not determine what the AI builds, only how the AI uses combat units that it has built.

an interesting read, I haven't finished it yet but so far far the author seems to be basing his assumptions purely as a technological historian, that is, basing predictions based purely on prior technological advancement.

There are 2 things that are not taken into account.

first, the political, social and moral/ethical factors that humans will utilize BEFORE AI achieves true general intelligence. 20 years from now, we could be on the verge of a general AI, then a powerful neo-Luddite movement pulls the plug on the whole thing. Put more generally, the forces guiding human action are progressing right along with technology.

The second is quite simply nature might interfere. [EDIT: 2 1/2 years ago, got the date mixed up from when it happened and when NASA reported it] we had a near miss from a solar flare that conveniently hit a deep space solar weather space ship and the measurements they got put it at an estimated Carrington level event with an estimated direct cost of 1.6 trillion dollars. Human kind have shown their inability to cope with natural disasters, I doubt AI would be any better, and all it takes is for the power grid to fail.

Of course, humans have also shown themselves to be remarkably self-destructive, so you never really know...


EDIT: Little off-topic reading for those curious about aforementioned solar flare http://science.nasa.gov/science-news/science-at-nasa/2014/23jul_superstorm/

I was wondering if there is a way we could help the AI learn, a kinda crowd sourcing thing. I, for one, would be more than willing to sign up for this.

A new level in difficulty; the Supreme Commander. An AI that learns from every game and micromanages on a unit level all game long.

Probably end up boom-botting everyone

Sorian's neural net isn't that complex. Sounds like it's more about basic decisionmaking following certain parameters than actually evolving. He said it takes only like one hour to calculate.

I'm curious, is there an opportunity for the neural nets to continue improving if you left them longer? I mean I'm guessing it's an iterative process, so your trending towards 'perfect', so how close to a theoretically 'totally trained' do you get in that time frame?