Google deepmind playing starcraft II

DavidBVal

Developer
Staff member
Administrator
Joined
28/02/2015
Messages
7,427
Highly unsettling. I had no idea AI was at such an advanced stage. Brace yourselves, the world is going to change and very quickly...
 

p4ran0id

Staff member
Moderator
Joined
27/01/2017
Messages
2,063
FULL ACK, Alphastar was after 1 week of training similar to a human who played 200 Years.

But still the AI (better call it machine learning) can only solve known problems at the moment. It's not possible to find something new, that's still something that only a human being can do. But still it will change our live in the next years. That's a really big impact for everyone.
 

Alan_SP

Staff member
Administrator
Joined
11/01/2016
Messages
1,957
AI is in much better position.

Its capacity can grow, while man's capacity usually diminishes after some time in his life (i.e. we grow old, slow, forgetful, etc.)

And I don't think that AI can't do new things. It is just question of programmed limits, if someone programs it without limits, there would be no limits what AI can do.

That's the story about three robotic's laws. Without them (and we still don't apply them to everything we do), we are "fair game".
 

p4ran0id

Staff member
Moderator
Joined
27/01/2017
Messages
2,063
Yes and no, AI can really good recognize patterns, but if something new accurs, it will always fail. So it's not possible to use AI for example in astrophysics, it can recognize the known phenomenon but not more. It can't show us something unknown. That's the actual limit of every machine learning algorithm. And I am not sure if we ever would be able to pass that.
 

Alan_SP

Staff member
Administrator
Joined
11/01/2016
Messages
1,957
I think that problems like this (recognizing something new) is much easier to program, then it is to learn humans to react in correct way. You know, old dogs and new tricks...

I'm not saying that I know how to solve it, but with advances, there would be more and more ways how to program AI and how AI will react to own surroundings.
 

p4ran0id

Staff member
Moderator
Joined
27/01/2017
Messages
2,063
Like I said, the limit is not the code. It's just a technical limit

Next step would be to teach an AI a few multiple things (different tasks). That would be the limit of neuronal networks with actual hardware.

An AI would be able to recognize new things, if it can learn everything else at once. That's not possible with actual neurological networks, it would need a quantum computer it's a n^n problem and it's not possible to solve that with actual architecture (faster or more CPU/GPU won't be enough, like with prime numbers for encryption)
 

Top