I was talking to a friend today about the AlphaGo games, and I had a thought that the Go community isn't really getting much out of this. I'm not saying it isn't amazing, or entertaining. And I'm not upset or anything.

It's my understanding that Google is using Go as a jumping off point for an AI that could go on to do great and helpful things, like maybe curing cancer. That is certainly noble and I hope that actually happens. I know that if AlphaGo wins some of the prize money will go to UNICEF, and some to Go charities. Andrew Jackson said that to run AlphaGo you might need to spend 7,000 dollars on a computer that could handle it. (Not an exact figure of course, just a conversation piece to show that it isn't some smartphone app.)

Deep Blue beat Kasparov and years later we had programs bundled into Windows that could beat grand masters. But if Google's intention is to just come in, beat Lee Sedol, and then move on to saving the world (again, a very admirable cause) What do we as Go players come away with? Other than a little money to the AGA, a week of good press, and the knowledge that there is a higher level of play, but no real way to study it? Is it possible we could actually have AlphaGo on our smartphones?

_________________
Lose 1,000 games? Challenge Accepted.

As a chess player, it's definitely nice that once and for all, Go players will stop pointing out that chess is an easy game because computers can play it well, while Go is the more difficult and more worthy of respect because computers can't compete with humans.

This is indeed the sort of thing go players say. And, as on some other issues that require a bit of peripheral vision, they are wrong.

https://tools.wmflabs.org/pageviews/#st ... s=Go_(game)

I think we are, and will, get a lot. Even if Google DeepMind don't release a public version (and they might, or sub-licence it or whatever) of AlphaGo for everyone to buy and play (with appropriate beefy hardware), I don't think it will take more than a year or two before other teams essentially re-implement their approach which has been proven to be so successful (though that will of course require a large amount of computer science expertise). We can already see this underway with team Deep Zen: the two authors of Zen have teamed up with some neural net folks at a Japanese university.

And I wouldn't dismiss all this extra press as nothing. I did an interview on local radio yesterday and got to plug our tournament and teaching session this weekend, so hopefully that will attract some new players. Today I did a talk on Go at work.

7 000 USD sounds like a miniscule fraction of the cost of AlphaGo's hardware. Perhaps that's the cost of the compute time for a single game. Didn't they say it was running on over 1 100 CPUs and over 300 GPUs or something? (On a side note, does anyone have that quote or the exact numbers?)

Personally, I'd like to see a battle between "Makers". AMD and Facebook and other teams are all working on their next-generation Go AIs, Google DeepMind are just the first to début against professionals. AlphaGo has been practising against itself. I want to see regular tournaments between the different blood lines and I want to see what happens when these AIs start practising against each other, like Insei.

It would be helpful, I think, for the professionals to have access to these engines. I never want to see one on my laptop. I think Lee Se-dol learned something about Go and about his own play style in the last two games but, for me, AlphaGo would simply be an opponent that I could never defeat. You can't really turn the difficulty down on these bots, either - they just become more and more random when you make them weaker and that does no good to a learning SDK player.

My only use for such a bot would be cheating, online, and game analyses. Cheating with AIs must be prevented at all costs and I don't think that bots are any good for analyses. Firstly, they can't explain their reasoning, and, secondly, in kyu or low dan games, there are probably many moves that win the game and AlphaGo inherently doesn't choose the "best" move - it chooses the one with the highest probability of winning. We should not study to emulate that.

Finally, it is entirely unnecessary to play against a bot when you have an "automatch" button on most servers.

Computer go tournaments are not exactly a new idea. In fact, come to think of it, they are my idea, from 1984. But not one I am specially proud of, in that programs tuned to play against other programs have not, so far, been specially useful.

AlphaGo plant the seed into computer go community, now Zen, Crazy Stone, Many Faces of Go, Aya, and the upcoming Baidu's go AI is now adapting deep learning. It'd be sad if AlphaGo quit but the idea is already injected.

Yes, from a commercial point of view, if pro-level go software is a development project rather than blue-sky research, then it's a question of what the market is. And in East Asia it would be big.

Indeed, Myungwan Kim 9p on the latest video said he would pay hundreds of dollars for it, and then corrected himself and said he would even pay thousands.

Computers didn't need neural nets to beat you guys...

_________________
Patience, grasshopper.

But so far this is true.

I can probably run a chess engine on my phone that can beat strongest chess players or at least play on their level.

When I run SF on my pc no GM could beat it.

The hardware deepmind uses is something like 1200 CPUs and 200 GPUs (something similar) which costs few million dollars.

Also it is not known if you can just feed random position to AlphaGo and let it continue playing (like you can do with chess engines and proper GUI).

Until we get go engines on our smartphones that can be run on the phone hardware, I don't agree with you.

When smartphones have the computational ability to run AlphaGo at the level we've seen this week, they'll also be able to do a lot of other things even more interesting before then, so I'll take such a phone even if there are no go programs for it.

No human will ever play perfect chess. It's too deep.

..Is it really fair to go around saying that everybody should play Go instead, because hey, it's even deeper?

People always bring computers into this as proof.

Well, hopefully Go will never again be introduced as the magical game that not even computers can play well. Take you guys down a notch, better late than never. That's all I was saying :D

I for one would be glad if it isn't introduced that way. I didn't start playing for that reason. Although, it was a nice "perk", if you can call it that, Go has other qualities that to me are far more attractive. The handicap system, for one, is unparalleled in board games. I have had so many great handicap games against both stronger and weaker opponents.

From the numerous "traing sessions" of AlphaGo vs. AlphaGo it should be possible to derive a komi, where the winrate of white vs. black becomes 50:50. This would be probably a more accurate value for proper komi than ever before. And that's at least something.

If go follows the chess model, there will be more cheating at first when strong go programs become available, but shortly thereafter, those same programs will be used to prevent cheating, which could be done at a rudimentary level simply by comparing the moves of the potential cheater with a computer's moves.

As far as analysis, I think anyone who truly wants to improve would be a fool to not use a computer, once the strong programs become available easily. A computer can give a list of moves with win percentage, including the list of possible follow-up moves. It should become standard pretty soon for someone like me to just run a quick program to analyze a go game I just played, and see instantly where my big mistakes were.

I am more worried about the global economic impact, as well as lack of a conscientious human touch, that the possibility of general purpose neural net robots will have on the world.