Total Posts:9|Showing Posts:1-9
Jump to topic:

Deep-learning software defeats human

PetersSmith
Posts: 5,819
Add as Friend
Challenge to a Debate
Send a Message
1/29/2016 2:31:59 AM
Posted: 10 months ago
Apparently, a Google program called "AlphaGo" managed to defeat a champion "Go" player in five back-to-back matches. "It has since gone on to win 99.8% of its matches against other programs, and has a match against South Korean professional Lee Sedol scheduled for March." It should be noted that AlphaGo "learned" to be "this good": "DeepMind's A.I. program studied 30 million positions from expert level games, then played against itself across 50 computers. It got better with each game it played, putting in on par with other Go programs, but DeepMind then gave it the ability to examine and interpret Go boards in order to pick its next move."

Apparently, the program is to go up against Sedol soon, "who many say is the best player in the world, DeepMind co-founder Demis Hassabis said he feels 'pretty confident' about the outcome". To quote Stephen Hawking: "The development of full artificial intelligence could spell the end of the human race." Luckily, this program cannot seamlessly go from playing the game to some other tasks, so it's non-adaptive in that area. You can read the full article here: http://www.nature.com...

So what do you guys think? Will AI soon start becoming "smarter" than humans? Will AI begin to take over our day-to-day trivial tasks? Is the Singularity upon us? Did I waste my time typing this? Discuss.
Empress of DDO (also Poll and Forum "Maintenance" Moderator)

"The two most important days in your life is the day you were born, and the day you find out why."
~Mark Twain

"Wow"
-Doge

"Don't believe everything you read on the internet just because there's a picture with a quote next to it."
~Abraham Lincoln

Guide to the Polls Section: http://www.debate.org...
Discipulus_Didicit
Posts: 3,088
Add as Friend
Challenge to a Debate
Send a Message
1/29/2016 3:22:20 AM
Posted: 10 months ago
At 1/29/2016 2:31:59 AM, PetersSmith wrote:
So what do you guys think? Will AI soon start becoming "smarter" than humans?

In any discussion like this, I think it is important to keep in mind something that you hint at near the end of your post. Specifically, the fact that for any software program we can concieve, it is literally impossible for it to do anything it is not programmed to do.

So, keeping this in mind let us say that the program gets smart enough to defeat anyone at any time at the game of go, no matter who. No matter what. Does this really make the program smarter than, or even equal to humans? No, not by a long shot.

Did I waste my time typing this?

In my opinion, no. Others may disagree.
Cobalt - You could be scum too.
Matt - I suppose. But I also might not be.

Kiri - Yeah, I don't know what DD is doing.
Vaarka - He's doin'a thingy do

DD - The best advice most often goes unheeded.
Wise Man - KYS, DD.
DD - Case in point ^
PetersSmith
Posts: 5,819
Add as Friend
Challenge to a Debate
Send a Message
1/29/2016 3:32:00 AM
Posted: 10 months ago
At 1/29/2016 3:22:20 AM, Discipulus_Didicit wrote:
At 1/29/2016 2:31:59 AM, PetersSmith wrote:
So what do you guys think? Will AI soon start becoming "smarter" than humans?

In any discussion like this, I think it is important to keep in mind something that you hint at near the end of your post. Specifically, the fact that for any software program we can concieve, it is literally impossible for it to do anything it is not programmed to do.

So, keeping this in mind let us say that the program gets smart enough to defeat anyone at any time at the game of go, no matter who. No matter what. Does this really make the program smarter than, or even equal to humans? No, not by a long shot.

Did I waste my time typing this?

In my opinion, no. Others may disagree.

I was referring to the technological capabilities that we're at now that we can even develop a program like this (like a "what's next" thing), but okay.
Empress of DDO (also Poll and Forum "Maintenance" Moderator)

"The two most important days in your life is the day you were born, and the day you find out why."
~Mark Twain

"Wow"
-Doge

"Don't believe everything you read on the internet just because there's a picture with a quote next to it."
~Abraham Lincoln

Guide to the Polls Section: http://www.debate.org...
Torton
Posts: 988
Add as Friend
Challenge to a Debate
Send a Message
1/29/2016 3:34:36 AM
Posted: 10 months ago
At 1/29/2016 3:22:20 AM, Discipulus_Didicit wrote:
At 1/29/2016 2:31:59 AM, PetersSmith wrote:
So what do you guys think? Will AI soon start becoming "smarter" than humans?

In any discussion like this, I think it is important to keep in mind something that you hint at near the end of your post. Specifically, the fact that for any software program we can concieve, it is literally impossible for it to do anything it is not programmed to do.
Even if it's only theoretical at this point, if we want true (strong) A.I., there needs to be some sort of way for an intelligence to be able to program itself (recursive self-improvement).
So, keeping this in mind let us say that the program gets smart enough to defeat anyone at any time at the game of go, no matter who. No matter what. Does this really make the program smarter than, or even equal to humans? No, not by a long shot.

Did I waste my time typing this?

In my opinion, no. Others may disagree.
Agree.
Discipulus_Didicit
Posts: 3,088
Add as Friend
Challenge to a Debate
Send a Message
1/29/2016 11:03:44 AM
Posted: 10 months ago
At 1/29/2016 3:34:36 AM, Torton wrote:
Even if it's only theoretical at this point, if we want true (strong) A.I., there needs to be some sort of way for an intelligence to be able to program itself (recursive self-improvement).

More specifically, able to do self-improvement on tasks that it has no base 'knowledge' of (for lack of a better word). The program above, for example, learned new methods of playing go from experience, but if we tried to make it play chess without programing into it the bare basics (bishop moves diagonal, rook moves in a line, definition of checkmate, etc.) It would have no idea what to do. We programed the bare basics of go into it, then it used self-teaching to play well.
Cobalt - You could be scum too.
Matt - I suppose. But I also might not be.

Kiri - Yeah, I don't know what DD is doing.
Vaarka - He's doin'a thingy do

DD - The best advice most often goes unheeded.
Wise Man - KYS, DD.
DD - Case in point ^
Floid
Posts: 751
Add as Friend
Challenge to a Debate
Send a Message
1/29/2016 12:31:41 PM
Posted: 10 months ago
At 1/29/2016 11:03:44 AM, Discipulus_Didicit wrote:
At 1/29/2016 3:34:36 AM, Torton wrote:
Even if it's only theoretical at this point, if we want true (strong) A.I., there needs to be some sort of way for an intelligence to be able to program itself (recursive self-improvement).

More specifically, able to do self-improvement on tasks that it has no base 'knowledge' of (for lack of a better word). The program above, for example, learned new methods of playing go from experience, but if we tried to make it play chess without programing into it the bare basics (bishop moves diagonal, rook moves in a line, definition of checkmate, etc.) It would have no idea what to do. We programed the bare basics of go into it, then it used self-teaching to play well.

Many computer learning techniques like neural networks are perfectly capable of figuring out the rules of a game as long as you give them feedback when they perform an action if it were legal or not. So for example if every time it tried to move a bishop horizontal or vertical you feedback a no it would eventually figure out the valid moves for a bishop. Figuring out the goal of the game (checkmate) can work in the same way except it is very, very slow because you are making dozens of moves that don't directly result in the goal (checkmate).
keithprosser
Posts: 1,933
Add as Friend
Challenge to a Debate
Send a Message
1/30/2016 4:02:47 PM
Posted: 10 months ago
Many computer learning techniques like neural networks are perfectly capable of figuring out the rules of a game as long as you give them feedback when they perform an action if it were legal or not.

While that is true, it is only true if the system was explicitly set up to learn which rules were legal and which weren't. Doubltess AlphaGo is good at winning at go, but it wouldn't be able to learn the rules of chess without being modified. No doubt a competent (human) programmer could make the modifications easily enought, but AlphaGo wouldn't/couldn't modify itself.

The problem of making an AI that is 'deep but narrow' (such as chess or go playera) have been solved, at least to a great extent, and from that point of view AlphaGo is nothing new. But making an AI that is flexible to enough to be even mediocre at a lot of unrelated things is still in the far future.
Ramshutu
Posts: 4,063
Add as Friend
Challenge to a Debate
Send a Message
1/30/2016 4:56:26 PM
Posted: 10 months ago
At 1/29/2016 2:31:59 AM, PetersSmith wrote:
So what do you guys think? Will AI soon start becoming "smarter" than humans? Will AI begin to take over our day-to-day trivial tasks? Is the Singularity upon us? Did I waste my time typing this? Discuss.

AI is already taking over our day-to-day trivial tasks, and already has been for a long while. It's just so ubiquitous we don't call it AI any more.

While it's not smart enough to take over the world by any means, learning computer systems are all around us, and they are key to almost every thing we do.
famousdebater
Posts: 3,938
Add as Friend
Challenge to a Debate
Send a Message
2/1/2016 9:32:45 PM
Posted: 10 months ago
At 1/29/2016 2:31:59 AM, PetersSmith wrote:

It doesn't seem to be a huge surprise to me. Computers are beating champions at games all of the time these days. It wasn't that long ago that the worlds greatest chess player was beaten by a computer. This stuff is pretty normal whereas people probably would have been a lot more shocked 10 - 20 years ago. This is probably why the thread hasn't got that many posts on it.
"Life calls the tune, we dance."
John Galsworthy