AI vs Dummies - Dummies win?

[John Fairbairn]

I have long been suspicious that the AI bots can teach us anything. They can certainly beat us, and they can certainly stimulate us into our own research. But can they teach us anything?

I ran LeelaZero for some time on the following position.



LZ stuck to move A for Black as the best move even though there can be no possible justification for this, even with billions of teragurgles of extra computing power. Yet the problem is trivial for humans.

Even if another bot or another version of LZ came up with the right answer, they presumably "think" the same way as my LZ, and I'm guessing that the right answer would then be more a question of luck or randomisation in how the search starts.

It therefore seems that were the current AI bots to learn to speak that they will still not be able to explain even trivial moves like this, and a fortiori they could never explain complex concepts like thickness. All they can ever say is, in effect, "this is the best move in this particular position."

Is this right?

[jlt]

LZ didn't care about ko threats, as there is no ko fight on the board.

[Uberdude]

I have long been suspicious that the AI bots can teach us anything. They can certainly beat us, and they can certainly stimulate us into our own research. But can they teach us anything?

I have learnt that approaching a 3-4 is more important than answering an approach to a 4-4.

As for the problem, indeed bots can't perform the abstract logically reasoning that allows us to quickly solve this problem 'correctly'. I actually put corrrectly in quotes, because now I just realised the traditional correct answer is not unconditionally better than the wrong answer. The standard rationale is it only leaves 1 instead of 2 ko threats. However, the wrong answer has the benefit that it you ignore white's ko threat then black's remaining dead stones have 2 liberties instead of 1. It would be pretty unlikely that this would matter in a game, but it is possible for the benefit to become manifest if a ladder or semeai ended up smashing into the surrounding white walls.

I also wonder what whole board position John used, if you made a realistic game position in which living here was the correct idea maybe LZ would find the better answer.

[splee99]

Regarding the title, it's far from win by answering this local problem correctly.

[Uberdude]

Quick try with LZ 201, slightly prefers correct answer.

Screen Shot 2019-02-04 at 16.53.07.png (1.06 MiB) Viewed 14157 times

[Knotwilg]

they could never explain complex concepts like thickness.

I'm pretty sure that even if they could, they would be told "no, that's not it".

More to the point, bots speak a language that is probably much closer to the language of professionals than the language of Go scholars, even if one person hosts a pro and a scholar. They both "see sequences" as I have argued in another thread. Because amateurs don't see the same sequences, scholars have abbreviated and generalized the professional sequences into heuristics.

People like Uberdude have already conveyed modern (AI/bot) opening theory to this crowd. If you've been using LZ for some time in game analysis, its sequences have sufficiently spoken for you to articulate its "thinking" into words. I don't know how much programming that would take but I'm not sure talking a conceptual teaching language would be that much harder to achieve than playing at professional level.

[Bill Spight]

I also wonder what whole board position John used, if you made a realistic game position in which living here was the correct idea maybe LZ would find the better answer.

Yes. Why wasn't tenuki correct? That's my choice on an otherwise empty board.

[Bill Spight]

they could never explain complex concepts like thickness.

I'm pretty sure that even if they could, they would be told "no, that's not it".

Neural net go bots are, perforce, global thinkers. Something that humans are encouraged to be. Despite the fact that the top bots are better than humans, humans sometimes are better with regard to local positions. Given their weakness at local thinking, the fact that the top bots are better than humans mean that they are much better than we are at global thinking.

Thickness, in the sense of outside strength, is a global concept. One of my first thoughts when viewing AlphaGo's games was, "There goes thickness." While many of AlphaGo's global plays were reminiscent of Go Seigen's, it was plain (to me) that its understanding of center oriented play went far beyond any current human concepts.

Humans are quite smart, and I believe that we will come up with new and better concepts or modify our current concepts. Thickness as we know it is yesterday's news.

[Knotwilg]

Thickness as we know it is yesterday's news.

Even if John may not agree with my definitions of thickness, it was him who made me understand the difference between local thickness (atsumi) and global thickness (atsuza). I'm forever grateful for that post.

What you saw disappear is global thickness, the value we presumed in being globally with less defects than the opponent, and influential stones working well together, whether or not to form a potential huge central territory, Takemiya-style.

I don't think the value of local thickness is gone to the same extent. The fact that White prefers crawling on the second line in the basic variation of the 3-3 invasion, rather than hane-connect, is very likely due to the opponent's tiger mouth giving local thickness to the wall resulting from that invasion. By crawling, the invader lives and the invaded gets outward influence, not thickness.

If you do this multiple times, you get global thickness, but the bots would still invade.

This is my proof

[sorin]

I have long been suspicious that the AI bots can teach us anything. They can certainly beat us, and they can certainly stimulate us into our own research. But can they teach us anything?

If they beat us and we don't learn anything from it, it's our fault, not theirs. Fortunately, most people learn from being beaten.

I ran LeelaZero for some time on the following position.



LZ stuck to move A for Black as the best move even though there can be no possible justification for this, even with billions of teragurgles of extra computing power. Yet the problem is trivial for humans.

I guess I am as bad as an AI, since A was the only answer I could think of!
It was only after reading other posts when I realized that you are talking between A and another way to defend that leaves one less ko-threat...

[Knotwilg]

Maybe we should ADD a topic “what’s trivizl?”

[tundra]

Why wasn't tenuki correct? That's my choice on an otherwise empty board.

My understanding is even worse, as I am still confused by Bill's choice. If Black plays elsewhere, away from the group, doesn't this allow White to kill the group by playing a white stone at A?

[jlt]

Playing locally is 24 points gote. At the opening stage, it is possible that some tenuki moves are worth more than that.

[tundra]

Okay, I think I can see that possibility. White currently has a small wall facing the empty part of the board, which could help him extend his control of that area. Black, on the other hand, has nothing out there. So perhaps Black should get some stones out there soon, as otherwise White could dominate that empty area? Or maybe I'm still missing something.

[Bill Spight]

Playing locally is 24 points gote. At the opening stage, it is possible that some tenuki moves are worth more than that.

By swing counting, the first move on an empty board is worth approximately 28 pts.