Interestingly, the AI winning changed the way Go was played altogether. Traditionally, the goal was to capture as much territory and win my as big a point gap as possible without overextending. The AI didn't care about the score gap, only that it won. So the AI would play theoretically 'smaller' but safer moves with the aim of quickly securing just enough to win but rarely if ever actually pushing too far. So professionals these days play with those smaller joseki (commonly accepted best local patterns) in mind.
hi ! 5-3kyu player here ! Did you EVER see the infamously known "AI josekis" ? they take as much as half the board for, and I quote a book referencing one of them here ; "At move 101, the result is considered even for both players". And nah, in pro games they play long sequences, they just hange depending on the "outside forces".
Also, did you ever hear this saying ; "A good go player only tries to obtain 51% of the board" ? This one is OLD. Like, I read it in 1950's book "attack and defense". The thing the AI did however, true, was put more importance into early territory by playing both territorial AND adaptive moves. Lee chang-ho (the stone buddha) is a counter-example to your "tradition", he is a very old go player who was famously known for winning by half a point in a lot of his games.
Humans are better at go again because they discovered a weakness in the neural network and this required an adversarial approach to fix this weakness. The solution (which drives us closer to AGI) has a basis built on perturbative theory. This idea to deviate the neural network beyond its own deemed "optimal" solution is showing us that we are blinded by our own local understanding. This deviation in perturbation theory begins to explore beyond our local intelligence. Now, build a neural network that tabula rasa discovers these perturbation deviations and you have created AGI.
A small correction, there is more than one rule when it comes to go ! Another rule is how the board state can't be consecutively repeated (this is to prevent infinite KO). As noted by reddit user cs_referral
you didn't talk about how territory is counted or the fact that the game ends only if a player gives up or they both pass one after another, either. And didn't talk about the "seki" rule, which, by the way, was used to completely DESTROY "leelazero" last year, and didn't talk about the fact that the leelazero at that time was equivalent to alphago in 2017. And an amateur canadian player used a loophole in the AI's way of playing to win 13 games out of 15. Anyway, the AI plays weirdly but if you take the concepts we as humans developped over the last three millenia and push them to the limit, you obtain what the AI does. The AI just accelerated research in the way of playing go. The sequences the AI invented ask for more things to betaken into account for a normal average human brain to comprehend, but if you can understand how a battlefield in real life or in a RTS game works, you easily can get how it came up with these ideas. I'm no historian, but a lot of the things the AI did that were "shocking", in fact, were done before by proffessional players thinking things like "I don't know why but I find this move beautiful". Anyway, sorry for the rant, but I think you can even use just the infos I gave you here to make another video on the subject. VERY Old example of an AI move being used : "the red-eared game of honinbo shusaku" (yes my name comes from that don't pay attention to it) other examples can be found in other games of the same player or you can also search into "go seigen"'s history, he was one of the first people playing in a mind-boggling way, even though his playstyle is far from AI, he used some of the "tricks" the AI uses as soon as the Eighties.
Well, as we all know, 2+2 can equal 5 if you ask ChatGPT - but that's why they have plugins with Wolfram if you're doing anynhing you canot do on a calculator !
That's because it is a language model, it is not designed for arithmetic. However, it is good at writing code. So now, when you present ChatGPT 4.0 with a math problem, it will just write python code for the actual arithmetic's and run it through a python interpreter. Works very well!
Good video. Go is a very good game. Better than chess. Very strong computer programs cause some problems but in general they make games much more accessible than ever. Once you needed a teacher in chess or go to become strong. Today a computer program can be your teacher.
Anybody who says go rules are simple must have never heard about case when Go Seigen discussed the rules and was proved wrong and he was the strongest player that time
Go rules are in general principles very simple, but Japanese rules of go use territory scoring which produces some very tricky edge cases. Chinese and AGA rules are simpler in use.
Interestingly, the AI winning changed the way Go was played altogether.
Traditionally, the goal was to capture as much territory and win my as big a point gap as possible without overextending. The AI didn't care about the score gap, only that it won. So the AI would play theoretically 'smaller' but safer moves with the aim of quickly securing just enough to win but rarely if ever actually pushing too far. So professionals these days play with those smaller joseki (commonly accepted best local patterns) in mind.
hi ! 5-3kyu player here !
Did you EVER see the infamously known "AI josekis" ?
they take as much as half the board for, and I quote a book referencing one of them here ; "At move 101, the result is considered even for both players".
And nah, in pro games they play long sequences, they just hange depending on the "outside forces".
Also, did you ever hear this saying ;
"A good go player only tries to obtain 51% of the board" ?
This one is OLD. Like, I read it in 1950's book "attack and defense".
The thing the AI did however, true, was put more importance into early territory by playing both territorial AND adaptive moves.
Lee chang-ho (the stone buddha) is a counter-example to your "tradition", he is a very old go player who was famously known for winning by half a point in a lot of his games.
Humans are better at go again because they discovered a weakness in the neural network and this required an adversarial approach to fix this weakness. The solution (which drives us closer to AGI) has a basis built on perturbative theory. This idea to deviate the neural network beyond its own deemed "optimal" solution is showing us that we are blinded by our own local understanding. This deviation in perturbation theory begins to explore beyond our local intelligence. Now, build a neural network that tabula rasa discovers these perturbation deviations and you have created AGI.
A small correction, there is more than one rule when it comes to go ! Another rule is how the board state can't be consecutively repeated (this is to prevent infinite KO).
As noted by reddit user cs_referral
you didn't talk about how territory is counted or the fact that the game ends only if a player gives up or they both pass one after another, either.
And didn't talk about the "seki" rule, which, by the way, was used to completely DESTROY "leelazero" last year, and didn't talk about the fact that the leelazero at that time was equivalent to alphago in 2017.
And an amateur canadian player used a loophole in the AI's way of playing to win 13 games out of 15.
Anyway, the AI plays weirdly but if you take the concepts we as humans developped over the last three millenia and push them to the limit, you obtain what the AI does.
The AI just accelerated research in the way of playing go.
The sequences the AI invented ask for more things to betaken into account for a normal average human brain to comprehend, but if you can understand how a battlefield in real life or in a RTS game works, you easily can get how it came up with these ideas.
I'm no historian, but a lot of the things the AI did that were "shocking", in fact, were done before by proffessional players thinking things like "I don't know why but I find this move beautiful".
Anyway, sorry for the rant, but I think you can even use just the infos I gave you here to make another video on the subject.
VERY Old example of an AI move being used : "the red-eared game of honinbo shusaku" (yes my name comes from that don't pay attention to it)
other examples can be found in other games of the same player or you can also search into "go seigen"'s history, he was one of the first people playing in a mind-boggling way, even though his playstyle is far from AI, he used some of the "tricks" the AI uses as soon as the Eighties.
Kind of late to the party, are we not?
Meanwhile, when I use ChatGPT it can’t even do simple addition.
Well, as we all know, 2+2 can equal 5 if you ask ChatGPT - but that's why they have plugins with Wolfram if you're doing anynhing you canot do on a calculator !
Give it time! It is only two or three years old. How much arithmetic could you do when you were 3?
That's because it is a language model, it is not designed for arithmetic. However, it is good at writing code. So now, when you present ChatGPT 4.0 with a math problem, it will just write python code for the actual arithmetic's and run it through a python interpreter. Works very well!
awesome video thanks
Good video. Go is a very good game. Better than chess.
Very strong computer programs cause some problems but in general they make games much more accessible than ever. Once you needed a teacher in chess or go to become strong. Today a computer program can be your teacher.
Underated
Anybody who says go rules are simple must have never heard about case when Go Seigen discussed the rules and was proved wrong and he was the strongest player that time
The rules of go are not simple. But cases of genuine rules complexity come up only very rarely
Go rules are in general principles very simple, but Japanese rules of go use territory scoring which produces some very tricky edge cases. Chinese and AGA rules are simpler in use.