There is a graph in the nature paper they released in january that shows how the strength scales with hardware.
Basically it scales a lot from 1 to 2 GPUs, but then looks like it growth logarithmically at best. : ]

Looking forward to future announcements. They kind of hinted that they might work on a single-machine version, so we could actually get "our own" AlphaGo at some point. :D

Glacierz

United States1244 Posts

April 01 2016 13:46 GMT

#422

One of the event organizers, Hajin Lee 3p, recently posted a video review of the games:

She had close access to Lee Sedol and offered some really interesting reflections of the moves from him after each game.

micronesia

United States24628 Posts

January 02 2017 22:58 GMT

#423

Have people been playing go recently? I started getting back into it (having nothing to do with the ai defeating the top pro). What struck me is how, after several years off, I started off right back at the strength I was at when I stopped playing (~13-14 kyu). After playing some matches I seem to be appropriately ranked at 12 kyu on KGS. I'm still weak enough that there is a ton of variation in the results of my matches... I can win by 100 moku one game and lose by nearly that much in another, even though the handicap system should minimize that (assuming the difference in ranks isn't too big).

I think I'm starting to get good enough now where it is actually worthwhile for me to study some of the more common joseki, but I'm not 100% sure I will do it... we'll see. I have been doing some life or death problems which I know, in theory, is very helpful. For whatever reason I seem to be much better at solving the problems when they provided to me in isolation than I am at applying what I learned into real games. I suppose playing more games will help with that.

I'm also forcing myself to be a bit more flexible with time controls. I used to only play long games because I didn't like to feel rushed. This time around I'm doing somewhat shorter games, and while it's frustrating at times, I'm also seeing it help me against slower opponents, so I can't really complain. For perspective, I used to only play 30 minutes + 5 x 30 seconds, but now I'm also doing 20 minutes + 5 x 20 seconds (for many players I'm sure that's considered slow as well).

The funniest thing is how I'm slowly marching my way towards actually being a good player, and even when I reach low single-digit-kyu I still will not know how to play go on a real-life board or be able to score without the computer helping me

nimbim

Germany983 Posts

January 03 2017 00:01 GMT

#424

You don't really get weaker if you stop playing, just the reading gets worse over time and at kyu level you will rarely see people read more than 1 variation or 4 moves ahead.

For joseki there are 2 proverbs that go together: memorize joseki to get weaker and study joseki to get stronger. If you just memorize the moves, you will be confused when people stray from joseki. If you know why move x is played in the sequence, you can properly benefit from your opponent's mistake or decide to go out of joseki yourself because it is actually benefical for the whole board situation.

Time settings should just be what you're comfortable with, imo. I know people who play exclusively 10s blitz games, but their fundamentals are just awful, because they rely on their opponent not finding the proper refutation for their overplays in time. That's not true for everyone of course, but you see it a lot in blitz games. It's also a matter of practice, when my reading was best I could read several variations out in 10s, so a blitz player is not automatically a sloppy player.

I think learning how to count in a game is more important than knowing how to move stones on a real board to make scoring easier, but that probably depends on whether or not you play a lot of real board games. Counting definitely isn't very important until you reach dan level.

I don't play very often on KGS or Tygem anymore, but I always have about 5 simultanous games on DGS and watch a lot of professional games. I also spend a significant amount of my go time on creating and fixing problems for goproblems.com

micronesia

United States24628 Posts

January 03 2017 00:05 GMT

#425

Thanks for the insights. Actually, goproblems.com is what I use.

How do you study joseki? Obviously I can look up joseki and read the little tidbits associated, but I think what you are suggesting is more involved than that. Keep in mind I'm probably not at a point where I'm really going to go that far out of my way right now, but it would be good to know.

nimbim

Germany983 Posts

January 03 2017 01:00 GMT

#426

You just go from the simple to the complex. For example the common 3-4 joseki with high approach
[image loading]

What is the purpose of each move?
1 is a standard move to make territory in the corner, because if Black gets to play again, he can make an enclosure at 3, 4, 2 or 6.
2 is a standard approaching move with emphasis on influence (4th line).
3 attaches to a single stone, that means both sides will get stronger locally, it also starts the sequence that gives Black solid corner territory.
4 is touching back and as a general rule you should always respond to an attachment directly. 4 stops Black from further extending along the top and reduces the liberties of 3, leading to
5 draws back to give Black a solid shape, which also leaves a cutting point at 6, so White defends.
After 6 the Black shape has a weakpoint, so he defends that.
8 makes a base for the White stones, the position is now about equal. Black gained territory and White got a stable group to limit Black's expansion from the corner.

If you are aware of the purpose of every move, you know why it is ok/not ok to omit it. For example if Black tenuki at 7, White can enclose Black and gain a lot of influence, because Black didn't defend the weakpoint in his shape.
[image loading]

This may actually be good for Black, if playing on top is super important and White won't be able to use the influence he gains here.

micronesia

United States24628 Posts

January 03 2017 01:04 GMT

#427

Is there a place where I can go to read more analysis like this? That is really helpful but I don't expect you to write a page or so on every common joseki lol.

nimbim

Germany983 Posts

January 03 2017 01:20 GMT

#428

Indeed, lol. I bit off a lot to chew there^^ However, I really don't mind helping out or reviewing games for TL users.

A good starting point are joseki dictionaries, they contain these explanations and some continuations. josekipedia.com has a lot of variations, you can also watch lectures by stronger players who explain these things.
I learned a lot from gocommentary.com, a mistake in this particular joseki is explained in this video http://www.gocommentary.com/free-videos/how-to-punish-overplay-03.html
My teacher reviewed a lot of Shusaku games (another go proverb is "study the shusaku fuseki") and explained the more basic concepts for kyu players https://www.youtube.com/playlist?list=PLS1Sn1wo0vB3da6Mlnpfl1fyDJUeypt09

Paljas

Germany6926 Posts

January 05 2017 14:46 GMT

#429

I'm sure most people here know this already, but the "Master" account which terrorized Tygem and FoxGO was revealed to be a new version of AlphaGo.
Considering the rise of other Go programms (Zen, various chinese programs and the mysterious "GodMoves"), the future of Go seems to take an interesting turn.

neobowman

Canada3324 Posts

January 11 2017 18:25 GMT

#430

I really wish more pros would release some analysis on the Master games. I'd super want to get some insight on this stuff but it's so hard to come by english content. I know a couple of lower dan pros have done commentaries and that's some great insight but I was hoping for some top pro input.

Paljas

Germany6926 Posts

March 28 2017 17:57 GMT

#431

^ again, im sure many people know this already but the aga youtube channel has released a lot of high level master reviews in the last couple of weeks

Ilovekatrina

1 Post

April 01 2017 21:47 GMT

#432

Hey guys.

If any of you guys want a teaching game on KGS, I'll happily give it.

I'm 1kyu so I can teach from "don't know the rules" to 4 kyu.

Drop me a message there my nickname is maggrig.

Akara12345

164 Posts

April 02 2017 03:00 GMT

#433

I still can't believe Go has been conquered by AI.

Paljas

Germany6926 Posts

May 10 2017 19:54 GMT

#434

For those that are interested, a new AlphaGo/Master live match even will happen around 23-27th of May
https://deepmind.com/blog/exploring-mysteries-alphago/

Glacierz

United States1244 Posts

May 22 2017 17:24 GMT

#435

It's happening in 9 hours, live stream can be found here:
https://events.google.com/alphago2017/

Excalibur_Z

United States12235 Posts

May 23 2017 01:41 GMT

#436

20 minutes! Let's go!

BrTarolg

United Kingdom3574 Posts

May 23 2017 13:06 GMT

#437

Damn alphago is nuts

The game is way too complex for me to understand haha

hypercube

Hungary2735 Posts

May 23 2017 13:51 GMT

#438

It seemed like Ke Jie was doing ok at the start (compared to the Master games) then he was 10 points behind.

Glacierz

United States1244 Posts

May 23 2017 14:38 GMT

#439

The program used less than 1 minute per move in the beginning, and end game plays were just a few seconds. It feels so much quicker than the Lee Sedol match, pretty scary how much it's improved.

Ke Jie stuck with a game plan of taking early territory, but ended up losing quite a bit on the top corner from a white reduction move. He then tried to reduce the center, but it wasn't enough for him to win the game.

As expected, AlphaGo made passive plays that simplified the board state and played loose endgame moves that didn't really maximize the points it wins by. The game isn't as close as the final score suggests.

BrTarolg

United Kingdom3574 Posts

May 24 2017 13:44 GMT

#440

It's pretty much at this point alphago is just playing teaching games with ke