Here's why I think so.

White 314 loses 6% ????? Well, Black is considered to be ahead at this point, so that may just be the win rate approaching 100%. But that underscores the problem of using consecutive win rate estimates to indicate mistakes.

White 310 loses more than 3% ????? Here there is no excuse that Black is ahead. even after White 310 Black is estimated to be ahead.

White 302 gains 1/3 pt. (I am using territory scoring.) Is this a mistake at area scoring? Possibly, depending upon the ko fight. At territory scoring canonical play is for White to take the 1/2 pt. ambiguous sente on the top (White 304) and then take the 1/2 pt. gote at 303. Then Black can take the ko at 302, but White has a local ko threat. (Edit: I have looked at the ko fight. Black is komaster, but I don't think he has enough ko threats to make White filling the ko superior.)

Black 301 gains 1/2 pt. It is miai with 303, but worse for ko threats. Canonical play is at 304, ambiguous 1/2 pt. reverse sente.

White 300 gains 1/2 pt. Again, canonical play is at 304.

White 298 gains 2/3 pt. (ambiguous reverse sente). Canonical play is at 299, gaining 3/4 pt.

White 294 (saving 2 stones) loses more than 1%, with White ahead?

Black 267 gains 3%????? With White ahead? Surely that indicates a previous misestimation.

White 256 loses 6%????? Even versus the variation? Really?

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

LZ has improved (in self-play Leelo at least) quite rapidly in the last few weeks, here's a game (LZ black) with 214. Opponent held his own fairly respectably until he died in the wrong way at top right (because he didn't make miai of saving the dead group or doing mean things to bottom right corner). A fun ko at top left, with him making nice use of the technique of playing out lower side to create some large ko threats, with LZ's decision to end the ko because it knew it could sacrifice the lower stones in an effective way to build some centre points was impressive (answer threat was not bad, but just something like 96 instead of 97%).

A game with latest netwoek #223. Not sure if the opponent was a bot, he always used 25 seconds per move for a while and only lost advantage slowly, but then messed up a fight and didn't resign for ages. I thought LZ's judgement of when to play solidly to settle, when to sacrifice, when to play severely was interesting. LZ's winrate slowly grew until he messed up a fight and game over. Highlights:
- move 26 white should have d9 sente first to stop black making some eyespace there.
- 41: LZ wanted r7 extend but lighter r8 was a promising blue circle, I played out a few moves and winrate stayed higher so I picked it.
- 48 was mistake, should n4 or l7 and make a mess. So why didn't LZ want to atari there before j4? And why did it want to atari there now rather than directly l7 (-1%). I wonder if something to do with white j2 in a sequence wanting to be gote. Or is it just if black makes the territory it's double sente yose?
- 65 don't often see LZ make this block, but quick life and makes r11 wrong
- 69 cut was preferred to (natural to me) o16 (-7%) getting out even though it then didn't run with it and just went back to living in corner. I think the option of q12 atari to help reduce a w moyo and make w bad shape was why.
- 76 o16 slow, should f17
- 77 interesting that LZ picked this way to attack, it is mutual base of 2 unsettled groups but struck me as slow compared to top side (e.g. g17 -3%). But it judges strength and weakness of groups here as urgent and doesn't mind if white extends on top, because by attacking this group we can expect to keep top under control
- 81 voila, black d9 group is now essentially alive whilst white group weak and only now does black play top side in good rhythm.
- 83 another lesson, attack from 2nd line not centre!
- 95 W should answer this peep in centre and black is 65% (see sgf var - it's in a way that means if black cuts white group has exit route and black has ugly shape. e18 was an invasion point but he got timing wrong. Now he starts making big mistakes, maybe his bot crashed.

Down to 8d, many bots at 9d on Fox. This was clearly against a human 8d, a beautifully clear example of the LZ strategy of taking territory, give walls, make a live group inside the moyo and 99% by move 80. Of course the 8d opponent helped with his slight mistakes, but still impressive. The paratrooper of move 50, kind of on the sector line aiming at p11 or q11 for sabaki was a highlight.

LZ 229 (black) using a ko threat factory (main branch):

"slight mistakes"?

He lost 13% at move 19 because of joseki fail, lol

then he appeared to lose fewer points, but actually it's because his score was so close to zero there was nothing much to lose. In isolation his other joseki mistake would lose 8%

Where should B play with 19?

The move chosen is very natural - I suspect a very large number of humans would select it.

My LZ want the solid connection at E13

LZ views the outside atari of 15 as -2% mistake and prefers the connection which some might find ugly as it causes a previous stone to become an empty triangle. I seem to recall that outside atari and squeeze being recommended in a joseki book which said connect was bad (probably Japanese). I expect LZ not liking the wall following hanging connection is not only the general tendency of strong bots to value centre walls as less valuable than prior humans, but a lot to do with the easy-to-play g14 peep (see also http://josekipedia.com/#path:pdqfncrdqhqcqerepfqgpgoe for a less obvious peep that changed value of wall and justified prior ugly playing into hane-at-head-of-two bad shape). LZ prefers e13 solid connection because then when white cuts at c13 black can atari at c12 and after white capture can tenuki because d12 is not atari. The g14 peep / f14 cut is still a significant weakness though so the wall is not really thickness and can even become a weak group.

And yes the n2 double hane is normally a bad move and n4 too is slack territorially, but it's the kind of move I can well imagine a Japanese top pro from the last century playing as part of a consistent moyo strategy. The reason it's so bad with the L3 counter is not a new AI move but one lesson I've reinforced from AI study. I did think it was odd he played n3 though and didn't even get s3: maybe he feared tenuki? LZ would have answered.

And worth bearing in mind top pros of today losing 8% (in LZ 229 opinion) a move is not unusual.

Me, too.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

That seems strange to me. If you don't like the empty triangel here, don't make the hanging connection in the first place.



When I was a kyu player I read something (by Sakata, I think) about this double peep connection, to the effect that answering one peep was OK, but then answering the second peep was not so good. That is not so easy to imagine when the peeps are on the 5th and 6th lines. And, as you point out, the ability to atari White C-13 looks big. As for the wall not really being thickness, that doesn't seem to bother the bots.



Also worth noting that the margin of error of the winrate estimates actually increases with the strength of the bots, not because they make more errors, but because the swings increase. With a bot that plays perfectly, the swings are 100%. For Leela 11 I reckon that plays that lose less than 3% are probably not errors. For Elf 1 I think the figure is higher, maybe 5%. Nobody really knows, because the research on margins of error has not been done.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.

Not played much as LZ recently; little time, and not really doing new things: it beats human Fox 9ds, sometimes loses to other bot 9ds with faster computers, and I'm unlikely to ever play top pros to test vs them. But here's something new, playing with LZ running on a smartphone. It was a one-way victory vs a human 9d. Phone is a Huawei Mate 10 Pro (high end phone released October 2017), with specs below. I was usually getting a hundred or two playouts per 20-30 second using a 15 block network Ms5 in Ah Q Go.

Now that you've played lots of games on Fox, as a bot... Has it been worth your time? What practical benefit have you gotten from this experiment so far?

_________________
be immersed

Kirby, yes, I think so, though I did end up playing other bots a bit more than I would have liked. By practical benefit do you mean has it made me stronger? Maybe a little, but I see various other benefits:
1. I have a better idea of the strength of LZ on moderate hardware vs pros: it's generally much stronger than a mid-level pro but sometimes can still mess up ladders.
2. But unfortunately I didn't get to play top pros, so still don't really know how it would fare against them with a few thousand playouts. My feeling is generally would win if it can avoid the sorts of positions it struggles with.
3. I've enjoyed watching skilful play, it's like when you kibitz a high dan game but have some control over the direction the game takes (e.g. there are several choices similar win% and I want to see how LZ handles an invasion in a particular place I can choose to play it and see how it unfolds, trying out variations during the game if time allows).
4. I have become a power-user of Lizzie/LZ, I understand the features well and the strengths and weaknesses of the bot which allows me to better analyze my own and pro games (e.g. https://lifein19x19.com/viewtopic.php?p=247941#p247941, https://lifein19x19.com/viewtopic.php?p=240352#p240352).
5. I have a better understanding of AI opening theory (Opening gospel according to LZ: https://lifein19x19.com/viewtopic.php?p=241104#p241104)
6. I have posted some games here which others have found interesting.
7. As for my strength, there is probably a bit of learning by osmosis and I think I've got a bit better at the opening and my policy network has broadened to play more interesting AI-style moves. I think I can play better when ahead, learning from LZ's solid style, and look out for exchanges to lead to overconcentration more now. But my reading of complicated positions is still probably about the same, and I'm still slow and prone to blunder under pressure. So if I can get a lead in the opening and hold onto it then I can beat people I wouldn't have in the past, but if they make it complicated I can easily mess up. I briefly got up to 6d on KGS in February when I was playing as LZ a lot (wife away) so the LZ style was coursing in my veins. I've not played much real-life Go this year, though I did have a nice game vs a Chinese 4-5d (Fox 8-9d) at the Cambridge tournament in March where I played LZ-style opening and then sabaki with a weak group in complex fighting and had a lead for most of the game, but messed up in byo-yomi as I often do. In May I qualified for another British championship final but was disappointed to lose to my friend/rival Andrew Kay: it was the first game of the tournament so I can half-reasonably blame not being warmed up, but I made a mistake in an early fight and then it was over. I won the other games but not particularly convincingly given the weak opposition.

So what is the explanation for this straight win with only hundred or two playouts? No strong opponent yet?

Fox 9d could be as weak as EGF 6d or even a strong 5d, so LZ with a few hundred playouts beating such a player doesn't surprise me much. Ah Q Go also says it has some modified engine to address ladder weakness. I've played 3 games with the phone since: lost vs a bot, winning (90% winrate and SE + 30 points) but had to abandon in early endgame vs a 300W 200L human, lost after keeping in pretty even for most of opening but then downhill vs a 215W 85L dunno, that game I could see better moves rising up the playout ranking (one crucial time it didn't make some double sente forcing exchanges before taking the big point, and giving it more time to analyse later it did prefer those moves quite a bit).

Played first game on Fox as KataGo, I liked its style, but then it's easy to look good vs a player as weak as a 160 win / 100 loss Fox 9d . Highlights were:
- move 12 classic bot opening theory of ignore approach to attack wall, his 13 extend was better as corner hand and then double hane as he has ladder but minor mistake
- 22: hane connect was bad as hurt r14 and that side aji was important next place, better to just o18 slide (AG teaching tool had some exmaples of this shape too in pincer joseki).
- 28 was important to r10 invade because it makes eyes for the wall before black has chance to peep o5 and reasonably disconnect and fight.
- 30 connect a little soft (and can see it was a good exchange later when black sacrificed corner), should s3 e2 tesuji combo and start ko immediately as probe
- 35 interesting as I thought this shape was often slow and loses m17, but here moyo more important. Black 5 points ahead now.
- 43 was aiming for white to block left and jsut black connects to reduction stones on 5th line, nice judgement that ok to give territory (overconcentrated?) here and just connecting is good enough and bolsters the moyo on right side.
- 50 white resisted and started fight, KG showed how to handle it with lots of nobis.
- 66 one line too close
- 67 cap at j10 looked natural to me and KG considered it at first but then liked c11 more. Very interesting judgement to go for local settling rather than attacking (if w answers c11 it's obviously good exchange and then cap), attacking has problem that black has weakness at m16 so top black stones not strong either.
- 68 off topic. m16 fight.
- 79 just give up ko highlight of the game. I understand it as b12 and b10 settle black AND also unsettled white c9 group on left, so black has created a splitting attack on the centre white group for the future. White can handle 1 weak group, but not 2. Black leads 15 point after b10.
- 82 should defend d5.
- 83 splitting attack lesson commences.
- 89 fix m16 weakness first before attack centre, B+22 now
- 92 although threatens a squeeze not sente enough, should break the line around m13
- white lives in gote, but black gets strong wall
- 117 now go after next group by taking solid profit thickly as still hard to run
- 118 if kosumi b goes for tail, but knight black cuts and kills. Game over +42 points.

Interesting sgf that you used.

_________________
Dave Sigaty
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21

Another KataGo game from Fox, displaying nice sacrifice and then use of aji that justified its decision to tenuki from top left. This sgf also includes my analysis of (var 2) Ilya Shikshin (B) vs Stanislaw Frejlak and (var 3) Ali Jabarin (W) vs Artem Kachanovsky from the Silk Road tournament currently being broadcast on KGS from Vienna.

Another KataGo game against a bot. KG had a really bad blindspot in not seeing g19 descent was better than f19 capture and then thought it was 5-6 points behind, but I kept playing to see if my probably LZ opponent would let it get closer and then mess up form assuming komi was 7.5 but actually Fox is 6.5, and indeed he filled in his own territory to lose by half at the end But KG (who I told 6.5 komi) thought it was winning even if he didn't do that which is weird.