AIRevolution/Discussion

RobertJasiek: Simultaneously, human go theory research by Berlekamp, Spight, me et al creates another, non-AI revolution concerning especially endgame evaluation and explanation, which is (becoming) much more accurate than previous endgame theory. AI does not take it all and render human invention void. Besides, AI lacks knowledge explanation so far.

Uberdude: I wouldn't call that a revolution, it has hardly had any affect on what people are playing. However, if you look at top pros' games today, or even random amateur kyus or dans you will easily see many moves which a few years ago would have been considered highly unusual or bad have now become a standard part of the repertoire as a direct influence of bots. The post-AI world of Go looks quite different to the pre-AI one. The post-Berlekamp/Spight/Jasiek yose theory improvements world looks much the same as the pre-.

RobertJasiek: Berlekamp: that is because people do not apply it, although it would improve their game. Spight / Jasiek (to be published soon): it also depends on whether people apply it, but everybody ought to apply it because of significantly to greatly improving their endgame. Of course, all endgame looks like "endgame", but different endgames moves mean better scores so "looking the same" misses the point of good endgame, which is about "better scores despite looking the same to the naive beholder". The impact of correct vs. wrong endgame is like the effect of correct vs. wrong status assessment due to tactical reading. It all "looks the same" because all is just variations. Endgame all "looks the same" because all is just values. Better variations! Better values! The "unusual" of better tactics or endgame is how small (or inexistent) the remaining error is. (Close to) perfect reading would be or (close to) perfect evaluation is (going to be) a revolution! Not noticed at a glance like the 3-3 under 4-4 change but only the lazy refuse to improve their game dramatically WRT the endgame. Soon, every amateur can learn better endgame evaluation (the one without complex tactics) than 9p because the theory will be applicable for everybody willing to calculate numbers. BTW, I am very much more impressed by the AI's emulated tactical reading skill than by their openings, josekis, fancy middle game moves. That revolution has also been mostly overlooked thus far because, like endgame, it does require meticulous study to really appreciate and profit from it.

Herman: Rather than discussing it at a sub-page under AI Revolution, we should consider whether it is warranted to have and "Endgame revolution" article. Something is called a revolution if it involves a sudden large shift in people's behaviour and opinions (eg. AI has caused a sudden large shift in the behaviour and opinion of players w.r.t. several joseki and opening patterns). I've not observed such a shift yet in how endgame is played, but I've not really been paying attention to it either. Is there such a shift? If not, at most one might say that endgame theory research contains the seeds for a potential endgame revolution, or perhaps that some of the research is "revolutionary" (ie. involving insights vastly different to previous research on the subject).

RobertJasiek:

• Discussion about possible new pages (or the new parent page) affects a) history of go theory / play revolutions (Dosaku, New Fuseki etc.) for which there can be, if not already there, a new page, b) the difference between go theory revolution and popular play / teaching revolution (the initial parent page has not cared clearly for the difference and should clarify and elaborate), c) a possible new page for revolutions by mathematical go theory (most of whose development has been a slow process, most of which is still only known to mathematicians or a few experts, some of which about endgame theory however is under current heavy research especially by Spight / me and has great impact on play if only it is applied).
• Concerning (b), the current version of the parent page still contains misleading information. First, AI has not created go theory (but has created game records); it is us humans who interpret the AI games as giving rise to (revolutionary) new go theory. Second, the difference between theory and popularity must be clarified. Third, the parent page must not give the false impression that other revolutions or major but slow changes after the New Fuseki did not exist. Concerning popularity only, maybe one can claim that the current AI wave is the next revolution. Concerning theory, it is hard to separate huge slow progress achievements of both tradtional theory and formal theory from theory revolutions.
• IMO, there have been the 1994 Mathematical Go Endgame revolution and the 2016+ Spight / Jasiek endgame theory revolution. Both revolutions have evolved from a previous, slow, long preparatory research phase and then culminated into systematic breakhtroughs from the point of view of Go application. Before 1994, microendgame was mainly seen as CGT for abstract games - since 1994, microendgame has become very Go-specific at least on the theoretical level (later translations have made great parts applicable for players). Before 2016, Spight's research was more theory for its own sake than ready for application and covered only parts - since 2016, application is becoming reality and I have been filling lots of gaps. A thermography researcher might not care for such Go-specific advances and rather insist that thermography would have been a revolution for CGT. As a Go player, my view is contrary: pure thermography is essentially inapplicable during playing a Go game and is too unspecific. IMO, the mentioned two revolutions are some because of their great impact on improving play by providing correct evaluation (within their scopes of application): microendgame really gets the last point, Spight / Jasiek evaluation gains lots of points which prior informal guesswork loses because of making evaluation mistakes.
• Clearly, so far, the (what one might call) endgame theory revolutions are revolutions of theory but not of popularity. The AI revolution is one of AI research and AI playing strength, is not one of AI-generated go theory, is one of popularity and is one in the human interpretation of AI games as go theory. All this can change in future: more aspects of a change might assume the characteristic of a revolution. Neither major change is a revolution in all respects. That every revolution emphasises difference aspects must not hide the other revolutions just because of difference emphasis.

Uberdude: Robert obviously has a rather different understanding of the word 'revolution' to me.

RobertJasiek: See above. Revolution of theory versus revolution of popularity. Revolution of AI itself versus revolution of human interpretation of AI output. Etc.

isd: Senseis Library is not here to sell books.

RobertJasiek: We know, but what is SL? A highly selective extract of knowledge about Go. One of the heavily treated topics on SL is contents trying to explain contents of the book Mathematical Go Endgames. Much of that treatment on SL is hard to understand. I have read much of that but what really helped me better understanding that book was reading it again twice. Books do fulfil important purposes a Wiki cannot fulfil. So 24 years after reading that book first I finally understood the core of the contents. Since I expected some opinion like yours, I have not mentioned a book from which somebody might quickly understand the contents of Mathematical Go Endgames. The consequence of hiding books can be that others share my experience of needing 24 years before their understanding will come. Instead of cultivating the hiding of books and knowledge, it is better that sources are mentioned when relevant!

AlphaGo's research and games have been promoted for free for Go players and the world but that does not mean that their development would have been free. It has been very expensive. Not every go theory can be developed under Google's supervision. Usually, go theory is developed under different circumstances and books are an important part of that, whether by Asian pro players or Eastern or Western amateur researchers.

isd: Fine. SL should not have popular topics hijacked in order for people to sell their books. If you want to talk about revolutions in Game Theory everyone is happy. Go ahead and do it, leave pages about AI related innovations alone.

RobertJasiek: The removal of the word theory now avoids the problems. There is no need for meta-discussion about what you call hijacking or book sales; a meta-discussion which misses the prior point of concern: misrepresented theory. When I will have the necessary time, I might create some page about history of modern go theory (not during the following months).

Leaving pages alone is a bad suggestions. Pages (other than homepages etc.) of a Wiki are not for only their initial creators or the fans of the initial creators' opinion. Pages shall represent different views and approach some sort of facts-orientated, neutral or balanced view.

Dieter: let's say we don't call it a revolution but X = "a massive shift in playing behavior of Go players". It is fair to say that the AI playing early 3-3 has resulted in X, but Robert Jasiek's endgame research has not (yet) resulted in X. If you would agree to X = "revolution", then we're all on the same page. The semantic problem may be that your work is "revolutionary", in the sense that it breaks with a tradition and potentially sparks a "revolution", in the above sense. We can solve the problem by calling your work "ground breaking" and reserve "revolutionary" for ground breaking work that has actually resulted in a revolution.