I think this is a high possibility! It kind of fits in line with the first AlphaGo Q&A video, the Zero opening study, and now the AlphaGo teaching tool. Seeing as how they are just getting the news of AlphaZero in this video, I think that the next video will be up to date with the release of the AlphaGo teaching tool.
In the position at 10:46 is it unreasonable to play o13 with the idea that it is the border of both moyos, with the intention of trying to turn b and outgrow him by making a moyo on the right?
It's bad in this game because the lower left black is so strong that white moyo is difficult to become territory and W O13 B M14 you can see that black moyo is much larger than white. White needs to invade black's moyo in this game.
44:30 Chris, I think you meant korean Sunjang Baduk. The first 16 stones are all preset on the 4th line and the game starts directly with the middle game fighting.
Neither Zero nor Master found move L9 to work. There's got to be a variation nobody is seeing. I can't believe both versions (the two strongest Go players in the world) would miss it.
It is very complicated. there could be one very weird one that is hard and unnatural to see for redmond. Remember black doesnt have to get very much to win the game
@Thomas Hancock while you will never be certain if a belief will lead you to success but opt to not form any beliefs because any beliefs are potentially false is no different than not looking for directions on foreign soil
FOR THOSE WORRYING : everybody is ok - the series lounch at feb 2nd with a sneak peak on jan 31st. Go check out the oficial AGA website for more information. Cheers! : )
What I'd like to see is a huge data dump from one of AlphaGo's games - the new paper mentioned that it evaluates 15,000 moves per second, which is definitely low enough to actually fit into an SGF file and have it be just a few tens of MB. It would be quite tough on current software, which wasn't written with this many variations in mind, but browsing it would definitely be achievable with some filtering schemes, perhaps only looking at branches that go a "large" number of moves away from the actual game, or opening just variations from a single point in the game, filtering out everything else. Custom software would help here.
In your variation, wouldn't W19 cut at H9? I didn't read it out, but it seems to be troublesome for black at least. Not sure B can afford to hane with 14.
I had Leela crunch the position for some time and it came up with similar things - a tough capturing race. It's no pro level AI when it comes to capturing races though, so I can't tell how that goes. Such a complicated position where both sides have options to extend their liberties.
I would say thank you for the hard work and the public communication at every video, but let's just consider it an iteration for the past and next ones ;) i was wondering if you could eventually get an announcement about your future plans ? For example : are you going for the reviews of the whole Zero games ? Do you plan on going back to Master's self played ones ? And although I love Redmond sensei commentaries as well as the duo with M. Garlock (i think you found out your rythme for this show) and can't wait for my weekly dose, do you plan to again make some human games commentaries like top tournaments finals with Myungwan Kim and the whole pro crew (i kinda miss them ^^) ? Thanks again and good luck for the future !
It's something about language, it's not intentional. I can't for the life of me say one guy's name I've worked with for a year. I have tried, and tried, and my mouth won't make the sound I hear other people say. It's pretty common with other languages without the same sounds.
31:45 So I'm not a strong go player but now 5 years later we have access to strong go AIs to help with analysis. I thought this game was interesting so I re-analysed it with KataGo on a powerful PC with a lot of playouts. So what happens if white cuts at L9? KataGo considers cutting here to be a very bad move for white which would almost certainly win the game for black. In KataGo's opinion the only valid option here is the game move: extending at H10. If white cuts at L9, KataGo doesn't play the response for black that Michael suggests (considering it a bad move). Instead it pushes at N8, white is forced to connect at M7 and black attacks the two white stones at H10. In KataGo's resulting playout white kills the black stones at M10 but black uses the aji of those stones to take a lot of the right side which is just too much territory for white to lose.
10:20 Was the invasion on the left side premature? KataGo says no it's fine, although it also slightly preferred to play on the top at L16, but the difference is very small.
12:49 Michael's suggested play on the left at D15. KataGo doesn't like his suggestion. It agrees the white group is alive but it views this outcome as an advantage for black (58.7%). Interestingly KataGo attacks this group with black and then lets it live. I'm not sure if the idea here is just to gain some extra territory for black in the exchange, or whether KataGo's plan is to use it for aji/ko threats.
16:50 Does playing the ko on the left side work for white? KataGo says no white can play (and win) the ko and strongly prefers it to the game move. It does choose very different ko threats to Michael though: S3, R2 and then ignores black's ko threat on the right side. KataGo agrees that this sequence as played resulted in an advantage for black and that L17 was a small mistake.
25:48 Is the peep by black at P5 a mistake? KataGo doesn't love it either and considers it a minor mistake by black. KataGo suggest playing at L9 instead.
Redmond at 32:12 identifies the so-called losing move. I disagree with his analysis 100%. White has 2 weaknesses after he immediately cuts by his L9--the move Redmond suggests is correct. The counter to white L9 is N10 which sets up M13 threat next for black, a severe attack. If white chooses to follow N10 with N13, then black can go the opposite direction with N8. White follows with M7, black with N7, and there is a big fight in that area which looks like alot of complexity to me.
In a word, yes. Redmond and Garlock took a break from the videos over the holidays to recharge and to focus on their AlphaGo book project, “which is coming along very nicely,” says Garlock. More on that soon, he promises. Redmond also took the opportunity to work on sgf commentaries for both Zero and Master games, as well as some of his own recent tournament games, and the two - along with producer Michael Wanek - have returned to the studio to begin production of the new series, which is expected to launch on Friday, February 2, with a sneak peek update from Redmond and Garlock on January 31. “The response to the videos has been quite amazing,” Redmond said. “I’m really pleased that so many go players are enjoying them and look forward to continuing to explore AlphaGo’s exciting games in the months ahead.” this is from the aga homepage. seems we will get a "sneak peak” in a short while!
Your pronunciation of 'Jie' is incorrect. IMO, the easiest way to think of it is as "yeah" with a 'j" instead of a "y'. It's kind of like the word chyeah that people used to say a lot around 10 years ago lol.
They also trained AlphaZero on Go, and it beat the 3-day trained AlphaGo Zero after just 8 hours of training. This should probably be taken with a grain of salt, though, since I understand they were throwing more computing power at the training process this time around (meaning it could play more games faster). One interesting thing I gleaned from the paper is that unlike AlphaGo, AlphaZero was trained to score as many points more than its oppoonent as it could rather than just win. This means it would be interesting to see play since it's unlikely that it would start playing slackly when it was ahead, and not play stupid moves when behind, since it wants to at least have a close score if it's going to loose. Sadly, they haven't released any of the set of games it played against AG0 20 block. AlphaZero Chess was interesting, especially since it became the world's best chess player after only 4 hours of training. If only they would introduce a ko rule to chess instead of this resulting in a draw things would be even more interesting! This would probably be impractical in human play, though, since the way chess pieces move around the board rather than being added means you can easily have a position repeat 20 moves later (making it really hard to remember if this position had happened before), unlike Go, where it would almost always (multiple ko situations are possible, but are rare to outright theoretical) repeat the very next turn if you let it. This is probably why they make the position repeat 3 times before declaring a draw, just to avoid accidental draws.
Just about everything in keldor314's comment is incorrect. 1. AlphaZero achieved parity with AG Lee after 175k steps/8 hours. AlphaZero, after 700k steps/3 days, was the one that defeated AlphaGo Zero 3-day (20 blocks), also 700k steps. Thus, the length of time is not only irrelevant, but also misleading. It's not clear whether they were using better hardware or not, but there doesn't seem to be a reason for them to use *different* hardware than before. 2. About the score, that is completely made up and does not exist in the paper. 3. Also untrue. AlphaZero achieve parity with 2016 Stockfish after 300k steps/4 hours. Immediately after, it improved slightly and then hit a hard plateau.The final games were actually played after 700k steps/8 hours of training. Eliminating threefold repetition doesn't mean that there won't still be a lot of draws.
ConsciousBreaks "AlphaGo Zero estimates and optimises the probability of winning, assuming binary win/loss outcomes. AlphaZero instead estimates and optimises the expected outcome, taking account of draws or potentially other outcomes." My interpretation of "estimates and optimises the expected outcome" is that it's talking about score. The wording could be a bit more direct, though, since it is, admittedly, kinda unclear. "Training proceeded for 700,000 steps (mini-batches of size 4,096) starting from randomly initialised parameters, using 5,000 first-generation TPUs (15) to generate self-play games and 64 second-generation TPUs to train the neural networks.1 " "1The original AlphaGo Zero paper used GPUs to train the neural networks." My copy-paste is refusing to work with the AlphaGo Zero paper, but briefly they mentioned 64 GPU workers and 19 CPU parameter servers. It's hard to compare these systems since GPUs are different from TPUs, and it doesn't mention which GPUs were used, but I'd hazard to guess that AlphaZero's training hardware was significantly more powerful. AlphaZero did indeed beat AlphaGo Lee after 8 hours and AlphaGo Zero 20 block after 34 hours. My earlier statement was a result of me misreading the paper.
Going from "optimising the expected outcome" to "score as many points" is a huge leap. It could be that optimizing the expected outcome, i.e. a win, means making the winning probability higher. I agree that the wording is ambiguous though. Also, thanks for pointing out that bit about the 64 GPUs. However, you should note that the 64 2nd gen TPUs and 64 GPUs were used to train the neural network, and *not* to generate the self-play games, which require the lion's share of computing resources. Both AlphaGo Zero and AlphaZero would have used the 5,000 1st gen TPUs to do so, it would seem. Thus, although the training hardware is better, the hardware for generating self-play games is comparable.
Michael and Chris, thank you so much for this series! I can't even begin to read the games as you do, but I always learn something nonetheless.
A pleasure, as always. I can't wait for Leela Zero to get close to this strength so that we can "Ask AlphaGo" what it's thinking.
How did I miss this video last week? lol
Watching it now!
Would it be possible to get a video of Michael going over variations with the AlphaGo teaching tool?
I think this is a high possibility! It kind of fits in line with the first AlphaGo Q&A video, the Zero opening study, and now the AlphaGo teaching tool.
Seeing as how they are just getting the news of AlphaZero in this video, I think that the next video will be up to date with the release of the AlphaGo teaching tool.
In the position at 10:46 is it unreasonable to play o13 with the idea that it is the border of both moyos, with the intention of trying to turn b and outgrow him by making a moyo on the right?
It's bad in this game because the lower left black is so strong that white moyo is difficult to become territory and W O13 B M14 you can see that black moyo is much larger than white. White needs to invade black's moyo in this game.
44:30 Chris, I think you meant korean Sunjang Baduk. The first 16 stones are all preset on the 4th line and the game starts directly with the middle game fighting.
+
Thank you
Neither Zero nor Master found move L9 to work. There's got to be a variation nobody is seeing. I can't believe both versions (the two strongest Go players in the world) would miss it.
It is very complicated. there could be one very weird one that is hard and unnatural to see for redmond. Remember black doesnt have to get very much to win the game
Yep, think so too.
+1
@Thomas Hancock
while you will never be certain if a belief will lead you to success
but opt to not form any beliefs because any beliefs are potentially false
is no different than not looking for directions on foreign soil
Katago says L9 is about 6pt mistake, and pushing at N9 is where the punishing from that mistake begins
Alphago partied so much for the New Year Eve that it got a hangover for a month
FOR THOSE WORRYING : everybody is ok - the series lounch at feb 2nd with a sneak peak on jan 31st. Go check out the oficial AGA website for more information. Cheers! : )
What I'd like to see is a huge data dump from one of AlphaGo's games - the new paper mentioned that it evaluates 15,000 moves per second, which is definitely low enough to actually fit into an SGF file and have it be just a few tens of MB. It would be quite tough on current software, which wasn't written with this many variations in mind, but browsing it would definitely be achievable with some filtering schemes, perhaps only looking at branches that go a "large" number of moves away from the actual game, or opening just variations from a single point in the game, filtering out everything else. Custom software would help here.
24:30
40 points-really Chris?! xD
I'm sure a DDK can tell that it's much smaller than that, LoL.
That was painful to watch Chris....
I think he said 14 points
Listen again-he definitely said 40. Whether or not he *meant* 14 is another thing.
i think he was trying to add to the total
The total of what? 60+15 = 40?
I think maybe Chris meant 14 and said the wrong thing. That, or he just completely hallucinated, lol.
35:45
i.imgur.com/JqSItKm.png
It will be hard for white to win the semeai.
In your variation, wouldn't W19 cut at H9? I didn't read it out, but it seems to be troublesome for black at least. Not sure B can afford to hane with 14.
i.imgur.com/PwJvFA1.png
i.imgur.com/JWpJmzs.png
It's impossible for white to capture J9 & K9.
I'd love to know what Redmond thinks about this variation!
I had Leela crunch the position for some time and it came up with similar things - a tough capturing race. It's no pro level AI when it comes to capturing races though, so I can't tell how that goes. Such a complicated position where both sides have options to extend their liberties.
Cool! Thank's for video!
Great video
I would say thank you for the hard work and the public communication at every video, but let's just consider it an iteration for the past and next ones ;)
i was wondering if you could eventually get an announcement about your future plans ? For example : are you going for the reviews of the whole Zero games ? Do you plan on going back to Master's self played ones ? And although I love Redmond sensei commentaries as well as the duo with M. Garlock (i think you found out your rythme for this show) and can't wait for my weekly dose, do you plan to again make some human games commentaries like top tournaments finals with Myungwan Kim and the whole pro crew (i kinda miss them ^^) ?
Thanks again and good luck for the future !
I don't understand why he has so much trouble pronouncing Ke Jie. I don't mind, I'm just confused.
It's something about language, it's not intentional. I can't for the life of me say one guy's name I've worked with for a year. I have tried, and tried, and my mouth won't make the sound I hear other people say. It's pretty common with other languages without the same sounds.
31:45 So I'm not a strong go player but now 5 years later we have access to strong go AIs to help with analysis. I thought this game was interesting so I re-analysed it with KataGo on a powerful PC with a lot of playouts.
So what happens if white cuts at L9?
KataGo considers cutting here to be a very bad move for white which would almost certainly win the game for black. In KataGo's opinion the only valid option here is the game move: extending at H10.
If white cuts at L9, KataGo doesn't play the response for black that Michael suggests (considering it a bad move). Instead it pushes at N8, white is forced to connect at M7 and black attacks the two white stones at H10. In KataGo's resulting playout white kills the black stones at M10 but black uses the aji of those stones to take a lot of the right side which is just too much territory for white to lose.
10:20 Was the invasion on the left side premature?
KataGo says no it's fine, although it also slightly preferred to play on the top at L16, but the difference is very small.
12:49 Michael's suggested play on the left at D15.
KataGo doesn't like his suggestion. It agrees the white group is alive but it views this outcome as an advantage for black (58.7%). Interestingly KataGo attacks this group with black and then lets it live. I'm not sure if the idea here is just to gain some extra territory for black in the exchange, or whether KataGo's plan is to use it for aji/ko threats.
16:50 Does playing the ko on the left side work for white?
KataGo says no white can play (and win) the ko and strongly prefers it to the game move. It does choose very different ko threats to Michael though: S3, R2 and then ignores black's ko threat on the right side. KataGo agrees that this sequence as played resulted in an advantage for black and that L17 was a small mistake.
21:23 Should white have played the peep at N16?
KataGo gives the peep at N16 and the game move the same score (with 5M playouts).
25:48 Is the peep by black at P5 a mistake?
KataGo doesn't love it either and considers it a minor mistake by black. KataGo suggest playing at L9 instead.
no sound 6 minutes in
I am unequivocally the worst Go player of all time, but I love watching machines do amazing, and sometimes ridiculous seeming, things. Great vids!
Fight me, I'm the worst. 😅
Well... I'm so bad, I think having someone "in Gote" means you fed your opponent to a goat. Also, hello fellow enthusiast.
Redmond at 32:12 identifies the so-called losing move. I disagree with his analysis 100%. White has 2 weaknesses after he immediately cuts by his L9--the move Redmond suggests is correct. The counter to white L9 is N10 which sets up M13 threat next for black, a severe attack. If white chooses to follow N10 with N13, then black can go the opposite direction with N8. White follows with M7, black with N7, and there is a big fight in that area which looks like alot of complexity to me.
after white cuts at L9, black keep overall initiative with P10. The cut i no good--white has too many small groups near the attacked black group.
Don't worry about your jobs, guys, you're awesome
"Thanks and see ya next week."
three weeks later..... D;
+
now im just getting kinda worried.... is everyone okay???
Same sentiment here. What’s going on? Has anyone seen any announcement?
In a word, yes. Redmond and Garlock took a break from the videos over the holidays to recharge and to focus on their AlphaGo book project, “which is coming along very nicely,” says Garlock. More on that soon, he promises. Redmond also took the opportunity to work on sgf commentaries for both Zero and Master games, as well as some of his own recent tournament games, and the two - along with producer Michael Wanek - have returned to the studio to begin production of the new series, which is expected to launch on Friday, February 2, with a sneak peek update from Redmond and Garlock on January 31. “The response to the videos has been quite amazing,” Redmond said. “I’m really pleased that so many go players are enjoying them and look forward to continuing to explore AlphaGo’s exciting games in the months ahead.”
this is from the aga homepage. seems we will get a "sneak peak” in a short while!
I hope Mr. Redmond is okay...
Ke Jie- Ke as in Cup without the 'p', Jie as in Jazz, with the 'zz'.
Your pronunciation of 'Jie' is incorrect. IMO, the easiest way to think of it is as "yeah" with a 'j" instead of a "y'. It's kind of like the word chyeah that people used to say a lot around 10 years ago lol.
Master on Tilt
I am from the chess community and I came to go because Lee Sedol lost to Alphago Lee.
As a weak player, I look at the right side and see what seems to be a huge white territory. That makes the resignation perplexing for me.
comes back at 7 minutes 55 seconds! Weird
On the other hand: "Humanity, you have to be communist in the long run, tending to 1 tax as we approach the Singularity"
"Capitalism is an impossibility in motion"
They also trained AlphaZero on Go, and it beat the 3-day trained AlphaGo Zero after just 8 hours of training. This should probably be taken with a grain of salt, though, since I understand they were throwing more computing power at the training process this time around (meaning it could play more games faster).
One interesting thing I gleaned from the paper is that unlike AlphaGo, AlphaZero was trained to score as many points more than its oppoonent as it could rather than just win. This means it would be interesting to see play since it's unlikely that it would start playing slackly when it was ahead, and not play stupid moves when behind, since it wants to at least have a close score if it's going to loose. Sadly, they haven't released any of the set of games it played against AG0 20 block.
AlphaZero Chess was interesting, especially since it became the world's best chess player after only 4 hours of training. If only they would introduce a ko rule to chess instead of this resulting in a draw things would be even more interesting! This would probably be impractical in human play, though, since the way chess pieces move around the board rather than being added means you can easily have a position repeat 20 moves later (making it really hard to remember if this position had happened before), unlike Go, where it would almost always (multiple ko situations are possible, but are rare to outright theoretical) repeat the very next turn if you let it. This is probably why they make the position repeat 3 times before declaring a draw, just to avoid accidental draws.
Keldor314 where in the paper did you read AlphaZero is trying to score as many points as possible?
Just about everything in keldor314's comment is incorrect.
1. AlphaZero achieved parity with AG Lee after 175k steps/8 hours. AlphaZero, after 700k steps/3 days, was the one that defeated AlphaGo Zero 3-day (20 blocks), also 700k steps. Thus, the length of time is not only irrelevant, but also misleading. It's not clear whether they were using better hardware or not, but there doesn't seem to be a reason for them to use *different* hardware than before.
2. About the score, that is completely made up and does not exist in the paper.
3. Also untrue. AlphaZero achieve parity with 2016 Stockfish after 300k steps/4 hours. Immediately after, it improved slightly and then hit a hard plateau.The final games were actually played after 700k steps/8 hours of training. Eliminating threefold repetition doesn't mean that there won't still be a lot of draws.
Also, the question "why not have a ko rule in chess?" is incorrect.
ConsciousBreaks "AlphaGo Zero estimates and optimises the probability of winning, assuming binary win/loss
outcomes. AlphaZero instead estimates and optimises the expected outcome, taking account of
draws or potentially other outcomes."
My interpretation of "estimates and optimises the expected outcome" is that it's talking about score. The wording could be a bit more direct, though, since it is, admittedly, kinda unclear.
"Training proceeded
for 700,000 steps (mini-batches of size 4,096) starting from randomly initialised parameters,
using 5,000 first-generation TPUs (15) to generate self-play games and 64 second-generation
TPUs to train the neural networks.1 "
"1The original AlphaGo Zero paper used GPUs to train the neural networks."
My copy-paste is refusing to work with the AlphaGo Zero paper, but briefly they mentioned 64 GPU workers and 19 CPU parameter servers.
It's hard to compare these systems since GPUs are different from TPUs, and it doesn't mention which GPUs were used, but I'd hazard to guess that AlphaZero's training hardware was significantly more powerful.
AlphaZero did indeed beat AlphaGo Lee after 8 hours and AlphaGo Zero 20 block after 34 hours. My earlier statement was a result of me misreading the paper.
Going from "optimising the expected outcome" to "score as many points" is a huge leap. It could be that optimizing the expected outcome, i.e. a win, means making the winning probability higher. I agree that the wording is ambiguous though.
Also, thanks for pointing out that bit about the 64 GPUs. However, you should note that the 64 2nd gen TPUs and 64 GPUs were used to train the neural network, and *not* to generate the self-play games, which require the lion's share of computing resources. Both AlphaGo Zero and AlphaZero would have used the 5,000 1st gen TPUs to do so, it would seem. Thus, although the training hardware is better, the hardware for generating self-play games is comparable.