Post AI Go
August 25, 2020 7:59 PM Subscribe
Impact of Go AI on the professional Go world by Hajin Lee: "At the time, which is not a very long time ago, no one questioned that Go was a path you walk for a lifetime. The life of a Go player was considered similar to a philosopher, a scholar, an artist, or a monk...The term “divine move” is used as a metaphor for an ultimate level of play. With AI, however, we all realized that the best way to reach the highest possible level of Go is not through thinking about it for a lifetime. It’s actually to buy more powerful GPUs and a well-trained deep neural network and have it play Go. So, suddenly, we players felt an enormous sense of loss." A response.
The plot of Real Steel notwithstanding...
posted by Scattercat at 8:41 PM on August 25, 2020 [4 favorites]
posted by Scattercat at 8:41 PM on August 25, 2020 [4 favorites]
I learned Chess as a grade-schooler, but only played very casually. As a young adult I started studying user interface design, which led into all kinds of sociology and educational theory tangents - a few of which used Chess as a teaching tool or metaphor. I was inspired to explore Chess again.
This was in the mid/late-90s, so the internet wasn't as much of a thing. Still, there was enough info on the early usenet and web to make it clear that Chess was not only a solvable problem, (i.e. for each board condition there was an objectively best move) but that many people were hard at work - using lots of computers - attempting to solve it.
This immediately took all of the fun out of it for me. Since most of the early game conditions were very unlikely in high level play (due to being disadvantageous/dumb) it meant that Chess could/would essentially become a game of memorization. Not terribly interesting or fun in my opinion.
Go is both more and less complicated to model mathematically, but it is still solvable - for many of the same reasons.
So I can totally understand, and am very sympathetic the damage this is causing in the Go community.
But I'm also very excited and intrigued when the mechanics of any game - especially one like Chess or Go, which have such high-brow traditions - are exposed to basically be longer versions of Tic-Tac-Toe.
It turns out that you can't ever really win, but only lose. Whatever the board presented to you, there was always a Best move, statistically. It's both empowering and frustrating.
posted by Anoplura at 8:49 PM on August 25, 2020 [1 favorite]
This was in the mid/late-90s, so the internet wasn't as much of a thing. Still, there was enough info on the early usenet and web to make it clear that Chess was not only a solvable problem, (i.e. for each board condition there was an objectively best move) but that many people were hard at work - using lots of computers - attempting to solve it.
This immediately took all of the fun out of it for me. Since most of the early game conditions were very unlikely in high level play (due to being disadvantageous/dumb) it meant that Chess could/would essentially become a game of memorization. Not terribly interesting or fun in my opinion.
Go is both more and less complicated to model mathematically, but it is still solvable - for many of the same reasons.
So I can totally understand, and am very sympathetic the damage this is causing in the Go community.
But I'm also very excited and intrigued when the mechanics of any game - especially one like Chess or Go, which have such high-brow traditions - are exposed to basically be longer versions of Tic-Tac-Toe.
It turns out that you can't ever really win, but only lose. Whatever the board presented to you, there was always a Best move, statistically. It's both empowering and frustrating.
posted by Anoplura at 8:49 PM on August 25, 2020 [1 favorite]
It's really interesting to see how different people see this from a different angle.
I play video games. While I don't speed-run, I do follow the speed-run community. They work to meticulously tear apart the games to find ways to save time. Certain games are non-random, and you can program a bot to run the game for you from pre-written instructions. This is called a tool-assisted speedrun, or TAS.
The TAS is interesting, and it's a good source of inspiration for some folks, but the community is all about the human record, not the computer record.
These games (Chess, Go) are sufficiently complex that even if computers can play them perfectly from all positions, there's no way a human can internalize it all. There's still so much room for execution and creativity.
posted by explosion at 9:01 PM on August 25, 2020 [5 favorites]
I play video games. While I don't speed-run, I do follow the speed-run community. They work to meticulously tear apart the games to find ways to save time. Certain games are non-random, and you can program a bot to run the game for you from pre-written instructions. This is called a tool-assisted speedrun, or TAS.
The TAS is interesting, and it's a good source of inspiration for some folks, but the community is all about the human record, not the computer record.
These games (Chess, Go) are sufficiently complex that even if computers can play them perfectly from all positions, there's no way a human can internalize it all. There's still so much room for execution and creativity.
posted by explosion at 9:01 PM on August 25, 2020 [5 favorites]
Chess is very far from being a solved game, and it is not clear if a solution (in the sense that checkers has been solved) is possible in practical terms. Perfect play has been computed for endgames with 7 pieces (including both Kings); those positions take about 140 terabytes to store. If you want to despair, the fact that computers are unbeatable by human Grandmasters now will have to do.
posted by thelonius at 9:07 PM on August 25, 2020 [6 favorites]
posted by thelonius at 9:07 PM on August 25, 2020 [6 favorites]
The second point in the response is on the money: there's been a huge shift, but it should open up new possibilities in the long run, even as people engage with the ai in pretty simple ways in the short term.
The real question is whether the go community survives at a sufficient scale to realize the gains. If it's less viable to be a pro and fewer people learn the game because the ai is better than any human, it'll be hard to keep up a population of players big and competitive enough to keep exploring the to level of play.
I kinda expect there will be some ebb and flow, but that the game will survive and maybe even thrive. The game got a huge amount of attention with Alpha Go, all over the world, which will lead to more players. We are perhaps seeing a changing of the guard, as some of the older to players question what they're doing. But the next generation of players won't have those questions, because a world with Alpha Go is the only world they'll know.
It's worth looking at the history of chess, post deep blue. My impression is that there are fewer pros, with more problems with cheating (as mentioned in the response). But I don't have a good sense of the overall changes...
posted by kaibutsu at 9:14 PM on August 25, 2020 [5 favorites]
The real question is whether the go community survives at a sufficient scale to realize the gains. If it's less viable to be a pro and fewer people learn the game because the ai is better than any human, it'll be hard to keep up a population of players big and competitive enough to keep exploring the to level of play.
I kinda expect there will be some ebb and flow, but that the game will survive and maybe even thrive. The game got a huge amount of attention with Alpha Go, all over the world, which will lead to more players. We are perhaps seeing a changing of the guard, as some of the older to players question what they're doing. But the next generation of players won't have those questions, because a world with Alpha Go is the only world they'll know.
It's worth looking at the history of chess, post deep blue. My impression is that there are fewer pros, with more problems with cheating (as mentioned in the response). But I don't have a good sense of the overall changes...
posted by kaibutsu at 9:14 PM on August 25, 2020 [5 favorites]
They should just increase the board size every time AI catches up. Never stop.
posted by fleacircus at 10:37 PM on August 25, 2020 [3 favorites]
posted by fleacircus at 10:37 PM on August 25, 2020 [3 favorites]
The boxing machine analogy is somewhat lacking. You don't have to worry about boxers looking up good punches during a break, but a go player could have a computer help them find a good move in a critical situation. You need rules for competitive go that say when the players are allowed access to their phones or other computers and how this is controlled during long matches. Traditionally, at least, top level games lasted multiple days.
In other words, even if you say computers can't take part in a human competitive go, the easy availability of computer analysis will still affect how people play competitive go and how people can potentially cheat. I'm sure these issues can be solved well enough that most players will be satisfied, but it's not quite as simple as just saying no computers allowed.
posted by primal at 11:42 PM on August 25, 2020 [6 favorites]
In other words, even if you say computers can't take part in a human competitive go, the easy availability of computer analysis will still affect how people play competitive go and how people can potentially cheat. I'm sure these issues can be solved well enough that most players will be satisfied, but it's not quite as simple as just saying no computers allowed.
posted by primal at 11:42 PM on August 25, 2020 [6 favorites]
Certain games are non-random
Amusingly enough (to me) I was watching the recent GDQ run of Doom Eternal (recent version), and a big chunk of the voiceover was spent discussing how the "Hell on Earth" level sucked because it wasn't fully predictable so it slowed you down.
[meanwhile, I swear the dance rhythm game guy has legs that can break the speed of sound. Three fans pointed at him and he was still overheating.]
posted by aramaic at 7:21 AM on August 26, 2020
Amusingly enough (to me) I was watching the recent GDQ run of Doom Eternal (recent version), and a big chunk of the voiceover was spent discussing how the "Hell on Earth" level sucked because it wasn't fully predictable so it slowed you down.
[meanwhile, I swear the dance rhythm game guy has legs that can break the speed of sound. Three fans pointed at him and he was still overheating.]
posted by aramaic at 7:21 AM on August 26, 2020
So the "game", the competition in speed running is not the original game but analyzing the software. In an automated speed run a neural net is trained by running every variation of the game on a giant array of thousands of virtual simulations.
But train an incredible system on Chess, Go, Doom Eternal, Poker and Flappy Birds to the point that it's unbeatable on each one. Then let's sit down at a game of Monopoly. Or Shoots and Ladders.
Taking bets.
posted by sammyo at 9:04 AM on August 26, 2020
But train an incredible system on Chess, Go, Doom Eternal, Poker and Flappy Birds to the point that it's unbeatable on each one. Then let's sit down at a game of Monopoly. Or Shoots and Ladders.
Taking bets.
posted by sammyo at 9:04 AM on August 26, 2020
Of course, the first comparison that comes to mind for many of us is chess. Deep Blue vs. Garry Kasparov happened 24 years ago, and we all know the leaps and bounds technology in general—and AI in particular—have taken during this time. (Human) Chess is not only alive, but finding new outlets (previously, on MeFi) in a world where the best chess AI is not only available for download, but fully integrated into lots of chess apps, bots and sites.
Sure, they can point to what is wrong with my move, why it's wrong, what would have been the best move... but they are pointing this out to a human, not to another computer. Teaching is more than merely pointing out mistakes and the general direction of optimal play. Teaching a human being starts with the assumption that they will learn and apply that knowledge as humans. Computers teaching each other is easy(-er) because they can share network parameters and values, but a brain works with its own individual heuristics. Teaching/learning "the AI chess style" means translating the former into the latter. This, in my opinion, still hasn't happened in chess.
A skilled player can, of course, learn with an AI tutor but I'm not convinced this is true of the majority of human players. We've taught the AI to play chess, not to teach chess. With such AI players, the burden of learning falls mostly on the student.
Without data, I can accept the author's comment that the demand for Go teaching has plummeted, but I believe (and hope) that this is mostly because of the buzz after AlphaGo only 4 years ago (CNBC: "Why the buzz around DeepMind is dissipating as it transitions from games to science"). Maybe it's still too early to see if the above observations on chess will also apply to Go. Maybe this is another example of "Oh but AI can't do *this* as well as humans. Time will tell, but I can see why many people can't wait.
posted by andycyca at 10:12 AM on August 26, 2020 [1 favorite]
[The author's pro friends] told me that in the pro circuit, it’s a unanimous belief that you need to play like an AI to win, and every serious player has an AI set up at home.This, if true, is saddening. But I hope is also temporary. After all, the post-AI Go world is younger than the post-AI Chess world, and lots of chess players are still learning with humans. AI Chess players are indeed wonderful, but AI Chess teachers/tutors? Very underwhelming in my experience.
Sure, they can point to what is wrong with my move, why it's wrong, what would have been the best move... but they are pointing this out to a human, not to another computer. Teaching is more than merely pointing out mistakes and the general direction of optimal play. Teaching a human being starts with the assumption that they will learn and apply that knowledge as humans. Computers teaching each other is easy(-er) because they can share network parameters and values, but a brain works with its own individual heuristics. Teaching/learning "the AI chess style" means translating the former into the latter. This, in my opinion, still hasn't happened in chess.
A skilled player can, of course, learn with an AI tutor but I'm not convinced this is true of the majority of human players. We've taught the AI to play chess, not to teach chess. With such AI players, the burden of learning falls mostly on the student.
Without data, I can accept the author's comment that the demand for Go teaching has plummeted, but I believe (and hope) that this is mostly because of the buzz after AlphaGo only 4 years ago (CNBC: "Why the buzz around DeepMind is dissipating as it transitions from games to science"). Maybe it's still too early to see if the above observations on chess will also apply to Go. Maybe this is another example of "Oh but AI can't do *this* as well as humans. Time will tell, but I can see why many people can't wait.
Before AI, most strong players had distinctive flavors of play (...) Today, everyone is trying to imitate the AI style, and the pros judge each other only by who is better at playing like the AI.My mind immediately goes to that Innuendo Studios video about Smash and wishes that the same happens here. More people learn the AI style, more people learn to counter the AI style, the AI style falls from grace, ad infinitum.
posted by andycyca at 10:12 AM on August 26, 2020 [1 favorite]
It's fascinating to recall that just like Go, chess used to be considered the hallmark of intelligence and refinement. The feeling of dislocation top Go players are experiencing is understandable. It's likely that most other games that rely on a limited number of moves or probability will go the same way, with top players being replaced by computers.
The final, most difficult challenge is obviously going to be duplicating highly social games. At this point I can state that for the last few years I have been part of a team attempting to use machine learning to duplicate the character actions of the highly social game Dungeons & Dragons, with separate AI systems taking the roles of different characters. In addition to the usual group of programmers and machine learning experts, I brought in an old friend whose name I will not reveal, but who has been involved with D&D as a player and writer for over 40 years, back to the early days of RPGs. They were actually very skeptical that machine learning could duplicate the D&D experience , but I felt their insights would be helpful.
The programming and debugging actually went more smoothly than I thought, up until last week when we tried our first "live" scenario, where everything went bizarrely, randomly wrong with the character AIs. The thief tried to pickpocket the wizard, the fighter didn't respond to anything, the cleric attacked the thief, and the wizard cast a point blank fireball. We're currently reviewing to see where in the learning process things went wrong. However my gamer friend simply looked at the results, looked at me, said "You asshole", and hasn't spoken to me since.
posted by happyroach at 5:59 PM on August 26, 2020 [5 favorites]
The final, most difficult challenge is obviously going to be duplicating highly social games. At this point I can state that for the last few years I have been part of a team attempting to use machine learning to duplicate the character actions of the highly social game Dungeons & Dragons, with separate AI systems taking the roles of different characters. In addition to the usual group of programmers and machine learning experts, I brought in an old friend whose name I will not reveal, but who has been involved with D&D as a player and writer for over 40 years, back to the early days of RPGs. They were actually very skeptical that machine learning could duplicate the D&D experience , but I felt their insights would be helpful.
The programming and debugging actually went more smoothly than I thought, up until last week when we tried our first "live" scenario, where everything went bizarrely, randomly wrong with the character AIs. The thief tried to pickpocket the wizard, the fighter didn't respond to anything, the cleric attacked the thief, and the wizard cast a point blank fireball. We're currently reviewing to see where in the learning process things went wrong. However my gamer friend simply looked at the results, looked at me, said "You asshole", and hasn't spoken to me since.
posted by happyroach at 5:59 PM on August 26, 2020 [5 favorites]
Reading up a bit on history of chess post-deep blue, one is confronted by the ENORMOUS CONFOUNDING FACTOR of the internet. Online play exploded, and has likely made it a bigger game than ever. We can definitely say that chess didn't disappear because of deep blue, though.
I found some old 'president's reports' from the US Chess Federation which touch on this question sometimes. They are the US ratings organization, so give some indication of the size of the 'serious' player base. It sounds like they've had gradual attrition of adult players, and maybe-slightly-growing participation amongst kids, who then fail to convert to adult members. Sounds like they recognize that online play is still growing, but that converting people to actual members is hard... For the bigger picture, this indicates that the number of serious tournament players is slowly shrinking, while the number of amateur online players continues to grow.
Overall, this doesn't seem unhealthy to me. My estimate is that the internet makes people better at their hobbies on average due to the crazy amount of sharing and education available. More casual and semi-serious players means more chances to find people who are seriously talented, who will end up engaging at the highest levels... So even if the overall size of the 'formal' organizations drops off a bit, the game has stronger roots in the much larger, more dynamic informal organizations.
posted by kaibutsu at 8:02 PM on August 26, 2020
I found some old 'president's reports' from the US Chess Federation which touch on this question sometimes. They are the US ratings organization, so give some indication of the size of the 'serious' player base. It sounds like they've had gradual attrition of adult players, and maybe-slightly-growing participation amongst kids, who then fail to convert to adult members. Sounds like they recognize that online play is still growing, but that converting people to actual members is hard... For the bigger picture, this indicates that the number of serious tournament players is slowly shrinking, while the number of amateur online players continues to grow.
Overall, this doesn't seem unhealthy to me. My estimate is that the internet makes people better at their hobbies on average due to the crazy amount of sharing and education available. More casual and semi-serious players means more chances to find people who are seriously talented, who will end up engaging at the highest levels... So even if the overall size of the 'formal' organizations drops off a bit, the game has stronger roots in the much larger, more dynamic informal organizations.
posted by kaibutsu at 8:02 PM on August 26, 2020
The kind AI that defeated go had brought a lot of new insights into chess and chess positions. New kinds of deep tactical ideas are being found that older brute force style systems seem to have missed. We’ve also seen a shift towards more events with shorter time controls for players to make their moves . This tends to make the memorization of opening lines less important and incentivize the players to find novelties and early branches out of common openings to equal positions so they can just try to out play their opponent in the middle game.
posted by interogative mood at 3:20 PM on August 28, 2020 [1 favorite]
posted by interogative mood at 3:20 PM on August 28, 2020 [1 favorite]
« Older An alcoholic can run out of booze or money. You... | 🐶 So long 🐕 and thanks 🐩 for all 🦮 the treats... Newer »
This thread has been archived and is closed to new comments
posted by jim in austin at 8:21 PM on August 25, 2020 [2 favorites]