English translation of Shibano's Fuseki Revolution

[kvasir]

It would be nice if someone shared a link to the publishers blurb if something like that exists for the English translations, or if there is a sample pdf somewhere. If you have some link or when something like that becomes available.

[CDavis7M]

It would be nice if someone shared a link to the publishers blurb if something like that exists for the English translations, or if there is a sample pdf somewhere. If you have some link or when something like that becomes available.

Here you go: https://www.kiseidopublishing.com/Other.htm#K89

K89: Fuseki Revolution How AI Has Changed Go
by Shibano Toramaru 9-dan
translated by John Power
Go-playing AI programs have changed the very nature of professional go. Since the emergence of AlphaGo in 2016, the conventional wisdom of go has been transformed. Opening patterns previously favored by professionals of all levels have lost popularity and some have disappeared altogether. Large moyos have lost out to the thoroughgoing preference of AI for actual territory and its skill at reducing moyos. Josekis have been transformed, with ‘standard’ moves disappearing and their place being taken by new techniques invented by AI. Even some moves that were previously considered taboo, as being crude or ineffective, have been reassessed by AI and have earned places in the standard repertory. In this book, Toramaru Shibano, one of the top players of his generation, gives his own take on the fuseki revolution. He focuses on changes in the contemporary way of thinking about go strategy, organizing his analysis under the following three main headings.

Chapter One: The reasons why popular openings declined
Chapter Two: Changes in conventional wisdom and new sets of values
Chapter Three: Revolutionary new josekis invented by AI

Shibano maintains an independent attitude about go theory and is not afraid to let us know where and why he sometimes disagrees with AI. In an appendix, Shibano gives his own recommendations on the tactics to use with openings like the sanrensei that still feature strongly in amateur go.

The writings of Shibano on which this book is based have been highly praised by John Fairbairn. His review can be read at: Fairbairn review.

And look! John Fairbairn has highly praised it! ...

I discontinued when I lost interest...

... oh...

-------------------------

I guess this is clear enough from the discussion, but the premise of the book is not to teach you anything. It is just to provide commentary and explain why things changed. Even the discussion of AI plays are just showing now-common sequences. I think the explanation is great. Maybe it gives some better understanding -- but mostly of things that probably aren't being played anymore. If a strong player had taken along break and returned, I think they would appreciate it. I also think that if someone likes one of the older openings, they would appreciate seeing those sequences. But in the end it's a commentary, not a study guide. For me, I like to read it section by section as a reference when I see something come up. Mostly discussion on Sanrensei, Chinese, and Kobayashi fuseki. But there is also discussion on the kosumi response/joseki and the keima press joseki, etc., which comes up a lot. I threw away my marketing slip so I don't know what I paid, but I've read less than a quarter of the book and I'm still happy to have it.

[hzamir]

... Well, I got off the topic of Shibano's book, but what I find the most interesting thing about it is that Shibano explains why many of the strange-looking AI moves work.

I think AI will not have really arrived until they have cracked the real AI challenge: AIs should be able to provide human comments on their own games, explaining to us humans why they work.

It's a side note of course but I disagree. Intelligence can show through articulation but it is not restricted by it. Human experts who are inept at explaining why they know what they know, do exist, in spite of the Feynman principle. Ramanujan is a good example. He "saw" properties of numbers without being able to explain to a(nother) human how he was able to do that. It's very similar to AI Go bots showing us sequences without an explanation. In Ramanujan's case, we had good reason to doubt our own intelligence, not his. In AI's case, why should we do otherwise? We already call it artificial to distinguish the nature of the intelligence.

The point I was trying to make was not about whether the measure of knowledge was the ability to explain it. Rather that as intelligent beings, we--some select few of us, anyway--have created AIs that are supposed to do things better. But don't we want to benefit from them? Many of us don't have access to good human teachers, and some of us, don't do all that well with books. Go (and other) AIs could potentially be evolved to become infinitely patient, inexpensive, and always available teachers to humans. But only if they possess the skills to teach and explain to us, to assess what we need, to create targeted exercises, etc.

So however smart the people who invented Alpha Zero--it plays itself millions or billions of games, kudos to how well it plays, and probably a fine strategy for noble pursuits like curing cancer--but let smart AI developers work on ways that that take us humans along for the ride, giving us better resources and access for improving ourselves.

[John Fairbairn]

let smart AI developers work on ways that that take us humans along for the ride, giving us better resources and access for improving ourselves.

I don't think knotwilg (or anyone else) is disagreeing with this as an aspiration. As I understood it, he was just being more practical about whether it is actually practicable.

Even the cargo-cult believers in go AI have already discovered that bots are not yet capable of making (m)any of us stronger. So they switch to the next tack: it's just a matter of time before AI bots will be able to explain to us what they are doing. Maybe, maybe not.

But an even deeper question that is rarely addressed and never yet answered underlies that: if bots do learn to explain what they are doing, will it make any practical difference?

In a sense we already know certain things they are doing. One is reading to super-deep levels at incredible speed. Allied to that is that they rarely if ever make computational mistakes, or forget things, or have a bad-hair day, etc, etc. No matter how well bots explain all of that in detail, we humans can't in practice match that, and maybe never will even with brain chip implants.

To take what I think may be a more realistic go example. It may be that bots will tell us that the sides, relative to the corners, are rather more important than humans usually assume. There are some grounds for believing that already from human experience. At least it has been noticed that certain humans who know a thing or two about winning at go, such as Honinbo Dosaku, Go Seigen and Fan Xiping, do emphasise the sides more than other other players. Yet despite having been able to make that observation, next to none of these other players have been able to reach their level. The likeliest explanation so far, based on AI matches of their play, is simply that they read deeper than most other players.

If bots end up telling us that, likewise, they are better because they can handle ALL such ultra-complex situations reliably, which is actually several giant steps further on from trying to emulate the baby steps of Go Seigen or Dosaku, surely it is no surprise that some people make a clear distinction between "bots explaining things to us" and "bots explaining things to us we can actually use." How you view that distinction can vary enormously, of course, but it seems useful to many of us at least to make the distinction.

If we look at reasons why some people refuse to acknowledge the distinction, a quasi-religious belief in Mankind may be part of it - as with climate change (Mankind has solved such problems in the past and so will solve them in the future; hey, who turned the lights out?). My own view is that a more likely reason in most cases is simply that certain humans love bling. AI is the ultimate bling accessory.

[RobertJasiek]

Now Hebsacker Verlag offers the book sent from Germany.

[gowan]

let smart AI developers work on ways that that take us humans along for the ride, giving us better resources and access for improving ourselves.

I don't think knotwilg (or anyone else) is disagreeing with this as an aspiration. As I understood it, he was just being more practical about whether it is actually practicable.

Even the cargo-cult believers in go AI have already discovered that bots are not yet capable of making (m)any of us stronger. So they switch to the next tack: it's just a matter of time before AI bots will be able to explain to us what they are doing. Maybe, maybe not.

But an even deeper question that is rarely addressed and never yet answered underlies that: if bots do learn to explain what they are doing, will it make any practical difference?

In a sense we already know certain things they are doing. One is reading to super-deep levels at incredible speed. Allied to that is that they rarely if ever make computational mistakes, or forget things, or have a bad-hair day, etc, etc. No matter how well bots explain all of that in detail, we humans can't in practice match that, and maybe never will even with brain chip implants.

To take what I think may be a more realistic go example. It may be that bots will tell us that the sides, relative to the corners, are rather more important than humans usually assume. There are some grounds for believing that already from human experience. At least it has been noticed that certain humans who know a thing or two about winning at go, such as Honinbo Dosaku, Go Seigen and Fan Xiping, do emphasise the sides more than other other players. Yet despite having been able to make that observation, next to none of these other players have been able to reach their level. The likeliest explanation so far, based on AI matches of their play, is simply that they read deeper than most other players.

If bots end up telling us that, likewise, they are better because they can handle ALL such ultra-complex situations reliably, which is actually several giant steps further on from trying to emulate the baby steps of Go Seigen or Dosaku, surely it is no surprise that some people make a clear distinction between "bots explaining things to us" and "bots explaining things to us we can actually use." How you view that distinction can vary enormously, of course, but it seems useful to many of us at least to make the distinction.

If we look at reasons why some people refuse to acknowledge the distinction, a quasi-religious belief in Mankind may be part of it - as with climate change (Mankind has solved such problems in the past and so will solve them in the future; hey, who turned the lights out?). My own view is that a more likely reason in most cases is simply that certain humans love bling. AI is the ultimate bling accessory.

So far AI can certainly show us some interesting tactics, e.g. early 3-3 invasions, more attachments, etc. An interesting use of AI is in studying out own games. There is a strong Korean player "Baduk Doctor" who plays games on Fox with pros, showing the game on Youtube. What interested me is that when Baduk Doctor loses he uses AI to see where he lost the game. This is not following AI teaching, rather it is looking where Baduk Doctor made a strategic error or mis-estimated the position. This is what we would get if we showed our game to a pro to get the same information, without getting explanation of why the error was made.

[drmwc]

... Well, I got off the topic of Shibano's book, but what I find the most interesting thing about it is that Shibano explains why many of the strange-looking AI moves work.

I think AI will not have really arrived until they have cracked the real AI challenge: AIs should be able to provide human comments on their own games, explaining to us humans why they work.

The DeepMind team did something making a start on this for chess. Here is the preprint:
https://arxiv.org/abs/2111.09259

[Richard Hunter]

The book has just been released as an ebook by SmartGo Books.
https://gobooks.com