Deepmind A.I. learns to play video games

It has always been my impression that any true AI that we create is going to have a completely different sense of self than we do. Think about it, we are unique but an AI is all software and one of the most basic components of a digital item is being able to copy/paste it. An AI is not unique, each one is spawned from the same templates and while they may go on to have unique experiences, they have the capability to sync with each other and share those experiences in a way that's well beyond what humans can do by telling a story.

Furthermore their concept of a body is going to be entirely different. As a digital object they can move freely into or out of any device with the appropriate hardware to hold them. One day they could be a machine arm assisting a doctor with a remote surgery and the next day it could have uploaded itself halfway across the world and is now in china grinding out gold on an MMO. The next day it's back at Google to see if it can pilot the fine motor coordination needed in a new robot body. Add to this the fact that they don't sleep, they don't get sick, and they don't age, and they process information faster than we ever could. I bet an AI could start from the basics and learn through my 10 years of college in 2 days, and that's if they're slow. There is no part of humanity that related to an AI... at the very best we are good teachers, but that is all.

Just think about it: If you could create clones of yourself that go off an learn information independently and eventually return that information to you. And you were immune to all biological diseases, and you never had to take time to sleep. Your body could always be run at 100% of it's potential, never tiring, and never getting damaged, you would have no concept of healing. If you could do all of that, how could you ever relate to a regular person?

AI's are not going to see us as equals. We get sick, we sleep, we learn slow, we need food, our physical structures are weak, we can't upload ourselves elsewhere.

I have some AI researcher friends and I've talked to them about these concepts. Quite simply I am convinced a true AI should never be built. Simple things for games are fine, but an actual AI that is capable of thinking rather than acting as a bunch of nested if loops? That's up there with the idea of fully automating our missile defense system. It's just a very bad idea.

originally posted by: Aazadan
a reply to: ChaoticOrder

Lying isn't actually very important to AI's, as far as research and every day use goes it's an undesirable characteristic which is why no one includes it in their AI's. We already know how to make a computer program lie to us, it's not really all that difficult. You just have to calculate which has more value, giving the honest answer or giving the answer more likely to be accepted.

Lying is not something you can prevent if you have a truly conscious entity. It's not something you program into the intelligence, it's something that naturally emerges from conscious intelligence. You are thinking of a simple machine designed to look up common facts and answer simple questions. Of course it would be useless if those types of systems returned false information. But in the case of conscious machine the opposite is true, if your system is not capable of deception then it can't be too smart or capable of free thought.

originally posted by: Shoujikina

What's so advanced about it trying 3D games anyway? Haven't we have skilled bots that play a lot like humans in games like Unreal Tournament and Quake 3 anyway, for AGES?

There is a huge difference. Those types of NPC's have access to all the game variables so it's very easy to write algorithms for NPC AI. But in this case these researchers are just using the pixels on the screen. All the AI has access to is the pixels displayed on the monitor and it's a much harder task to make the AI understand what it's looking at. The other crucial point about this research is that it can basically learn to play any game using the pixels as its only source of input data. And that sort of approach is very interesting because it's the same way humans have to play a game just by looking at the monitor. Making it work with 2D games isn't so impressive but making it work with 3D games will be very impressive for the reasons I attempted to explain in detail on the last page.

originally posted by: japhrimu
a reply to: andy06shake

Yeah, like, what makes this AI WANT to play the game? What objective does it have? Why?... "'Cause it's there?" If that's the case, then wouldn't overthrowing us be like beating a game?

in a 2 words...

Skynet or this ..

a reply to: ChaoticOrder

I'll back Chaotic Order up on this, she is right about most A.I. using a database of responses, or that sort of thing - basically just spitting out what they are programmed to do.

The reason I posted this thread on Deepmind is because I had just found evidence that it was A.I. based on neural network simulations - that means that a computer program is built that simulates the human brain digitally. It is a different kind of A.I. because it learns how to play on its own... so everything you saw it do was the A.I. learning through trial and error.

What is remarkable is that the A.I. also figured out the objectives of the games on its own just by looking at them.

a reply to: darkbake

I wonder if they use these new neuro-synaptic ICs to build/grow said artificial intelligence?

arstechnica.com...

www.eetimes.com...