Why does Facebook want the machines to win at Go?


#1

[Read the post]


#2

Let’s not teach the computers all of our tricks.


#3

What do you mean? The smarter the machines are, the more money Zuckerberg makes with no effort. Everybody wins! (Except almost everybody.)


#4

Speaking of algorithms that need improvement, your image search gets a big F here: If that’s a Go game then I’m a 9 Dan.

This is like illustrating a post about chess with a board that has white pawns on the first rank.


#5

Not all smart people are good Go players, but the best Go players are also the smartest people.


#6

Sounds like it would be a good problem for quantum computers.


#7

Machines don’t “win” at anything, because they have no goals. They are tools which function as an extension of the user, an agent. Just like you aren’t racing your own car when you drive.

With mechanical tools, I think this was intuitively obvious for most people. But with electronics - and particularly programmable systems - people are more likely to overlook this. You don’t talk to a telephone, rather, a telephone mediates between you and another party. Likewise, you are not “on” the internet, but rather are using servers and scripts to communicate with other people. Even software based upon indeterminate input is still written by people. Software written by other software is simply a once-remove from typical human goals written into it.

This can, and arguably does, have the effect of people hiding behind technology to obscure their manipulations of others. It is even more opaque than finance and economics, which is really saying something! But I think it is crucial to avoid the temptation to assume that computers have agency, because this discourages people from scrutinizing those whose tools impinge upon them.

FWIW I am fine with real autonomous computers, because unlike in dystopian sci-fi, their requirements and goals are likely to be so alien to mine that there is little overlap.


#8

Speak for yourself :smiling_imp:


#9

Do you mean racing against your own car, or racing your car against other people in theirs?

I should have been more clear, but I was referring to the former.


#10

Oh,I know. It’s tricky and usually involves a big sunroof.


#11

I can imagine…


#12

More relevant…


#13

It seems like it’s 50:50 whether a random image of a Go board even makes sense :smile:

“Ah hell… Why are you asking me?! I don’t know how to play! Just put 'em where it looks good and take the photo.”


#14

Why is Mao so far down on that list? From the description on the Wiki page, it just has a fixed set of rules (which may be different per game, but no matter) that can be deduced through inductions. It sounds like exactly the kind of thing that a machine learning algorithm can do many times better than a human.


#15

Maybe the choice of image is an ironic statement about the state of machine Go players…


#16

I would put Fizzbin high on the list of “games at which computers will never beat humans”


#17

But will computers ever be able to beat a human at Go Johnny Go Go Go Go?

Go Johnny Go Go Go Go

The game is a cross between Hoover and 8 men down.

The Rules:

Jacks are worth 10 Kings are worth 3 apart from one eyed Jacks which are Wild cards. Round one you get a hand of 9, round two you get a hand of 7.

Two’s are wild cards, apart from diamonds, which retain their face value except for the king of diamonds, which is worth the same as all kings, 3.

You play in sequence, unless you can match a card in an ascending or descending order.
If you can then that’s a Go Johnny Go Go Go Go, then you stand up and shout Go Johnny Go Go Go Go and pick up all the cards on the table.

The winner is the one with the most tricks after 15 hands.

There simple.

And remember, when reciting your pairs you can’t look at them.


#18

One of the linked articles is really interesting. They’re using monte carlo simulations, which is basically a brute-force computational approach to the problem, but they’re doing stats on a whole bunch of monte carlo simulations, thus rendering the problem largely parallelizable. There’s a certain amount of time it’ll take a core to finish a simulation, but you can throw as many cores at it as you have money for, and get more accurate results.


#19

Actually it looks like a neat semeai tesuji exercise, but without knowing the full board state, it could just be a pretty pattern.


#20

That sounds like something someone who has lost a lot of matches against bots would say.