Life In 19x19

How strong is Zenith Go, the commercial version of Zen?

I ask this question because of an interesting event in China. Weiqi TV there produced a series of programs to discuss AlphaGo. In today's program, Luo Xihe 9-dan, a former Samsung Cup winner, gave an unique view on AI. He suggested that any AI would have weakness which, once discovered, could be exploited by human players. To proof his point, he played a live game against Zenith Go , giving the AI a 9-stone handicap and beat it easily. He said he played many games against Zenith Go so presumably he knew that AI's weakness quite well.

Then there's this shocking opinion: Luo said he could probably give Alpha Go 4 stones. It doesn't look like a joke.

I think it's 4D with normal time setting and fast computer (not that server's fast)
But the latest experiment version is 6D now (it's 7D in fast game in KGS)
But the exploitation of AI weakness is nothing new, it happened in computer chess too. But there'll be the time human just can't win against AI unless in a handicap game. Computer chess was already passed that point and I see no reason why computer go can't reach that point too.

pookpooi wrote:I think it's 4D with normal time setting and fast computer (not that server's fast)
But the latest experiment version is 6D now (it's 7D in fast game in KGS)
But the exploitation of AI weakness is nothing new, it happened in computer chess too. But there'll be the time human just can't win against AI unless in a handicap game. Computer chess was already passed that point and I see no reason why computer go can't reach that point too.

So what? Making a strong program is a good achievement for AI scientists but, perhaps controversially, it has nothing to do with humans playing go. Would we feel it had any meaning for human abilities if a robot that moved on two "legs" could "run" 100 meters in 5 seconds?

gowan wrote:

pookpooi wrote:I think it's 4D with normal time setting and fast computer (not that server's fast)
But the latest experiment version is 6D now (it's 7D in fast game in KGS)
But the exploitation of AI weakness is nothing new, it happened in computer chess too. But there'll be the time human just can't win against AI unless in a handicap game. Computer chess was already passed that point and I see no reason why computer go can't reach that point too.

So what? Making a strong program is a good achievement for AI scientists but, perhaps controversially, it has nothing to do with humans playing go. Would we feel it had any meaning for human abilities if a robot that moved on two "legs" could "run" 100 meters in 5 seconds?

We'd feel very happy, for example, if driverless car is prove to drive 100x better than the best human driver with zero accident, I'd feel very happy to buy one because it means I've own a good car.

Interesting. Is there a sgf of his game against Zenith Go?

On a related note, I have downloaded Hiratuka on my computer and sometimes play against it. It is strong enough for the games to be interesting but it has trouble realizing that some one-eyed group are dead. I never really tried to find which situations exactly triggers this but it is always surprizing to see the bot throwing away points because he thinks he is so much ahead, while in fact the game might be close or already lost for him.

I don't know if AlphaGo has flaws of this kind but it would be nice if we could see more games by him. I remember from Go Word that in many commentaries of title matches between two pros that had never played each other in a final, the commentary sometimes makes clear that after an even start (let's say one game each), one of the player seems to "take the measure of the other". Sometimes so much so that the remaining games are more one-sided. It seems Lee Sedol is denied this by not being able to play against AlphaGo before March.

That video is a joke because the settings for Zen was completely wrong (set level rather than time). Can see how that mistake is made (Zen has a pretty counter intuitive setting where "5d" is weaker than 5 seconds per move and you can only select either strength or time), but can't believe no one at weiqitv realized and pointed it out.

macelee wrote:How strong is Zenith Go, the commercial version of Zen?

I ask this question because of an interesting event in China. Weiqi TV there produced a series of programs to discuss AlphaGo. In today's program, Luo Xihe 9-dan, a former Samsung Cup winner, gave an unique view on AI. He suggested that any AI would have weakness which, once discovered, could be exploited by human players. To proof his point, he played a live game against Zenith Go , giving the AI a 9-stone handicap and beat it easily. He said he played many games against Zenith Go so presumably he knew that AI's weakness quite well.

Then there's this shocking opinion: Luo said he could probably give Alpha Go 4 stones. It doesn't look like a joke.

Luo is full of sh*t.

I would like him to try and give 4 handicap to AlphaGo.

I think Luo would struggle* to give 4 handicap to Zen19X which has been doing well at 7d on KGS, but there is a chance for him to win if he can practice a lot against it to find its weaknesses (for example rsun (user who plays bots a lot) says Zen has some flaw where if you approach 4-4, Zen plays one space low pincer and then you double approach in a position in which that is an overplay which should be punished by 3-3 (such as san ren sei) then Zen doesn't play 3-3 but a usual local joseki of attaching). Would be interesting to see this TV game vs Zen to compare its play. Win on 9 stones I would eat my hat. Then again I might have said I would eat my hat if last week you said a computer could already beat a pro even, but maybe not as that improvement from around EGF 5d (see code centric challenge vs FJ) to EGF 8d (Fan Hui level) is less than the difference of 4 vs 9 stones to Luo; and also being unaware of advances in AI level is different to looking at known bot levels and judging how far from pro they are.

* Edit: Maybe struggle is too strong, but not be assured of victory at least. If we say Zen19X is KGS 7d (maybe it is 8 though?) and Andy Liu is 9d and Lee Sedol is 11d+ (beat Andy on 2 stones easily). Maybe Luo is particularly good against bots so if we say he and Lee are KGS 12d then that would be 5 stones or so. Maybe Zen is not particularly good at using the handicap stones though?

pookpooi wrote:

gowan wrote:

pookpooi wrote:I think it's 4D with normal time setting and fast computer (not that server's fast)
But the latest experiment version is 6D now (it's 7D in fast game in KGS)
But the exploitation of AI weakness is nothing new, it happened in computer chess too. But there'll be the time human just can't win against AI unless in a handicap game. Computer chess was already passed that point and I see no reason why computer go can't reach that point too.

So what? Making a strong program is a good achievement for AI scientists but, perhaps controversially, it has nothing to do with humans playing go. Would we feel it had any meaning for human abilities if a robot that moved on two "legs" could "run" 100 meters in 5 seconds?

We'd feel very happy, for example, if driverless car is prove to drive 100x better than the best human driver with zero accident, I'd feel very happy to buy one because it means I've own a good car.

Yes. I would buy such a car if I could afford it. My point related to how many go players seem to think that the existence of a program that can play go well enough to defeat the strongest pros somehow diminishes humanity. Of course that is silly. There are many human activities that computers have long ago surpassed human abilities. For example adding a list of 1000 10 digit integers.

gowan wrote: There are many human activities that computers have long ago surpassed human abilities. For example adding a list of 1000 10 digit integers.

Having said that...

https://en.wikipedia.org/wiki/Kim_Peek wrote:Aged 18, he got a job working out the payroll for 160 people: this was a task that took him only a few hours, without the need for a calculator. However, he became unemployed a decade later when his employers decided to computerize payroll accounting, and he was replaced by two full-time accountants and a computer.

gowan wrote:Yes. I would buy such a car if I could afford it. My point related to how many go players seem to think that the existence of a program that can play go well enough to defeat the strongest pros somehow diminishes humanity. Of course that is silly. There are many human activities that computers have long ago surpassed human abilities. For example adding a list of 1000 10 digit integers.

You mean this?
https://www.reddit.com/r/baduk/comments ... akthrough/
also there's top comment in mark zuckerberg post about his progress in Go AI like 'why don't you leave an ancient board game alone' but that comment is gone.

illluck wrote:That video is a joke because the settings for Zen was completely wrong (set level rather than time). Can see how that mistake is made (Zen has a pretty counter intuitive setting where "5d" is weaker than 5 seconds per move and you can only select either strength or time), but can't believe no one at weiqitv realized and pointed it out.

Is there some documentation somewhere on how the different settings relate to the program's strength?

I am using Zen go 6 and I noticed that it only uses 25% of CPU even on 7 dan.

Is there a way to change it so it uses all 4 cores on full capacity?

pookpooi wrote:
We'd feel very happy, for example, if driverless car is prove to drive 100x better than the best human driver with zero accident, I'd feel very happy to buy one because it means I've own a good car.

Would be off the topic except we are discussing whether AI's might have bugs affecting their performance in practice.

The self driving cars have just had their first fatal accident (broadside into the side of a semi across an intersection where the color of the semi's trailer matched the background)

As a Brit, I always chuckle when an American says "a semi". For the unaware, he means a truck/lorry.