Back to game 3 (sorry, I'm not quite keeping up with the pace here) -- a couple of interesting variations:

Position after move 35:

Click Here To Show Diagram Code

[go]$$Wc White to play
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . O . . . . . . X . . . |
$$ | . . . . . . X X O . . . . . O . . . . |
$$ | . . . . X X O O X X . . . . . . . . . |
$$ | . X X . X O . b O . . . . . . . O . . |
$$ | . . O . O . O X O . X . . . . . . . . |
$$ | . . . O . . . a . . . . . . . X . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ +---------------------------------------+[/go]

Here, Otake played at a. I wondered if this was too slow and if white could play in the lower right instead. "Obviously" if white tenukis, black can cut at b. But I had trouble reading out a clear continuation. Here's one possibility, with a little help from LZ-258.

Click Here To Show Diagram Code

[go]$$Wc White tenukis, and then?
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . O . . . . . . X . . . |
$$ | . . . . . . X X O . . . . . O . . . . |
$$ | . . . . X X O O X X . . . . . . . . . |
$$ | . X X . X O 3 2 O . . . . . . . O . . |
$$ | . . O . O . O X O . X . . . 1 . . . . |
$$ | . . . O . . . 4 5 a . c . . . X . . . |
$$ | . . . . b . 8 0 6 7 . . . . . . . . . |
$$ | . . . . . . . . 9 . . . . . . . . . . |
$$ +---------------------------------------+[/go]

Next white a, and black b or c in either order. It looks as though black loses the capturing race but can squeeze on the outside, so on the whole black is ahead (around 60% winrate).

Later, after move 84:

Click Here To Show Diagram Code

[go]$$Bc White's status?
$$ . . . O . . . . . |
$$ O . . . . . . . . |
$$ O X X O X X X X . |
$$ X O O O O X O O . |
$$ . X . . X O . . . |
$$ . . . X X O O O . |
$$ X . . . O B . . . |
$$ . . . . O B . O . |
$$ . . . . . B . X . |
$$ . . . . . . . . . |
$$-------------------+[/go]

It looks to me like white needs one more move to make a living shape. But black can't afford to try and kill, because the circled stones are fragile.

If black does try a blunt attack now:

Click Here To Show Diagram Code

[go]$$Bc Capturing race
$$ . . . O . . . . . |
$$ O . . . . . . . . |
$$ O X X O X X X X . |
$$ B O O O O X O O . |
$$ . B . . B O a 7 3 |
$$ . . . B B O O O . |
$$ B . . . O X 1 2 0 |
$$ . 6 . . O X . O 9 |
$$ . . . . 4 X 5 X 8 |
$$ . . . . . . . . . |
$$-------------------+[/go]

Black can follow up at a and rip off three stones, leaving the rest of the corner as seki -- but black's marked stones probably die.

My impulse is to play Otake's move at a, and Elf agrees. Maybe more later.



Without doing any reading myself, I suspect that the weakness of the squared stones is also a factor. In game Honda extended from them to make a base, and then Otake played at a.

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

Visualize whirled peas.

Everything with love. Stay safe.

Up to game 4 now :-) I also did a double-take when I saw the iron pillar at move 19. But what's just as strange (to my eyes) is that LZ-258 wants white to reply not at O7 as in the game, but with the attachment at P7. This leads to a nice tesuji which I was actually able to spot because I've just been looking at the relevant chapter in the Segoe dictionary!

Click Here To Show Diagram Code

[go]$$Bc After this, white has a tesuji
$$ . . . . . . . . |
$$ . . . . O . . . |
$$ . . . 3 4 . . . |
$$ . . 6 2 X 1 . . |
$$ . 9 7 5 8 . . . |
$$ . . . . O . . . |
$$ . . O . O O O . |
$$ . . . X X O X . |
$$ . X . . . X X . |
$$ . . . . . . . . |
$$-----------------+[/go]

Speaking of game 4, a brief note about White 126.

KataGo prefers the block at F-09, yielding a 67% winrate for White with 56k rollouts. Elf, in its commentary, also likes F-09, devoting almost 143k rollouts to it and giving it an 79.8% winrate. However, even though Kudo's play, the atari at R-10, is almost entirely off Elf's radar, getting only 4 rollouts, Elf gives it a 77.7% winrate, only 2.1% less than the block. That is well within Elf's margin of error. (With only 4 rollouts, Elf does not calculate a winrate directly for R-10, but inherits its winrate from Black 127. Elf does that for any play with fewer than 500 rollouts. )

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

Visualize whirled peas.

Everything with love. Stay safe.

This is an interesting one: the various bots disagree here. LZ-258 doesn't find the game move on its own (on 10k playouts), but awards it +1% when shown. Of course LZ-258 thinks komi is 7.5 not 5.5. Trying with KataGo 1.2 and 1.3 (out since you first posted this!): for me, KataGo 1.2 has the game move at -4% (different random number seed?) whereas 1.3 says -7%, setting komi at 5.5. The numbers move around a little if you change komi to 5.0 or to 7.5, but S15 often looks like a possible small loss not a clear mistake. (KataGo disagrees with LZ-258 even on 7.5 komi.) Interestingly, if you set komi to zero, KataGo searches a bit more broadly.

This is also a nice test case for KataGo 1.3's new score estimation (median instead of mean!) Changing the komi between 5.5 and 7.5 makes the score estimate move by about 1.7 or 1.8 points, whereas the same change for KataGo 1.2 results in an almost 4 point change!

After the interlude of "Otake's perfect game", I wanted to publish the 7th game in the series, against Kajiwara. It's by far the most complex analysis I've conducted and I'm not at all at the end of it.

Problem: after analyzing with LZ, the file has become too large.

The interesting thing about this game is that - for John's sake! - the pros seem to consider all the complex variations Kata sees. They play the important moves, just in a different order.

The exception may be move 94, a very natural extension from atari, maintaining pressure, which Kata suddenly evaluates as a 23% loss.
Instead Kata offers a variation which can lead to a furikawari, but it ramifies into a complexity I can't contain. From the game, I have no hint Otake or Kajiwara considered this line.

Otherwise though, the game makes a case for John's point that it's the human errors/blunders that make the difference, not the ingenuity of the bot, or its better developed intuition. The "blunder" in question is Kajiwara's 223, a first line hane which seems like it was meant to be a ko threat, keeping sente, but it wasn't.

Well, if by blunder we mean a large reading error, there are at least three sources of error, 1) incorrect calculation of variations; 2) poor choice of candidate moves; 3) incorrect evaluation of results. As for 1) bots are perfect and human pros, nearly so, (Edit: OC, the bots can make a much more extensive search in the same amount of time than humans can do consciously; as for human unconscious search, quien sabe?) 2) is related to the bot's policy network, and 3) is related to the bot's evaluation network. It is the advances in both the choice of candidate plays and the evaluation of the results that enabled the bots to leap to superhuman play.

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

Visualize whirled peas.

Everything with love. Stay safe.

Although, there is some room for better ingenuity too? One or two interesting points came up in a game review Michael Redmond streamed here recently about the right side battle in one of his own games. It doesn't seem quite right to call human pros missing the moves merely blunders, it feels more like, the bots have developed some understanding of some kinds of probes and flexible shapes that humans haven't quite internalized yet.
https://www.twitch.tv/videos/540764912?t=2h40m20s

If you've analysed and saved the file using Lizzie, it will have added a lot of tags of the form:

(It says "LZ" even if you used KataGo.)

Those tags allow Lizzie to remember the evaluations and hover-over variations next time you open the file. But they make it a lot bigger. If you have a text editor that can do search and replace using regular expressions, then you can probably get rid of all the LZ[...] tags by replacing "LZ\[.*?\n.*\];" with ";". (I say probably because regex isn't perfectly standardised and can be tricky.) If this doesn't work for you, I'm happy to discuss other options either here or by PM. I'm keen to see your analysis :-)

Analyze with Katago and annotate in Sabaki a clean second copy is another option.

(Sabaki should probably add an option to get rid of the tags in the clean sgf dialog.)

I got this far in understanding the problem but haven't yet found an editor with regex capability. I was down to using python but then sloth got the better of me. If I send you the file, would you be so kind ... ?

You can open the file in Lizzie, select "clear evaluations" or some option like that in the menu, then save it. It will still have all the analysis tags there, but they'll be wiped to just a dummy "0 visits, no evaluation" tag, which takes much less space.

In the "Analyze" menu, there's an item that says "clear analysis", but for me it only removes the analysis info from the current move, not from the entire SGF file. Am I missing something here?

No problem. Email is probably easiest.

Eeew. Okay then. Regex it is.

Thanks to xela, here's the sgf:

Many thanks to Knotwilg (and xela) for showing us this interesting game.

Quick comment on . Looking at the Elf commentary, I expect that the suspicious drop in White's winrate estimate is the result of too few rollouts, which makes the winrate estimate unreliable. For instance, Elf recommends the jump attachment at F-09, giving it 107k rollouts, while Otake's play gets only 2.1k rollouts. Elf estimates the White winrate after as 68%, but after , which Elf thinks is correct, it estimates White's winrate as 71½%, with 31.6k rollouts. Do we really think that was a mistake that boosted White's chances by 3½%? Or do we think that it provides a better winrate estimate with 15 times as many rollouts?

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

Visualize whirled peas.

Everything with love. Stay safe.

Similar remarks apply to the winrate estimate after . KataGo estimates White's winrate at 25%. But on how many rollouts?

Elf estimates White's winrate after at 23½%, but with only 678 rollouts. It then agrees with Kajiwara's play, , and then estimates White's winrate at 34½% with 30,641 rollouts. Do we think that despite being both Kajiwara's and ELf's choice, was a mistake that increased White's chances by 11%? Or is 34½% a better estimate of White's chances, based upon more than 45 times as many rollouts?

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

Visualize whirled peas.

Everything with love. Stay safe.

Like KataGo, Elf missed , giving it only 1 rollout. It didn't entirely miss it, it just wanted to play the sequence W D-11 - B C-11, W B-10, B B-09, first, thereby sacrificing an extra stone on the left side unnecessarily. After Otake's it estimates White's winrate at 49%, while it estimates White's winrate after its top choice for White of D-11 at 35½% (with 59.8k rollouts). IOW, it rates its top choice as losing 13% to Otake's play. (BTW, its estimate of 49% for that play is not based upon only 1 rollout. It inherits its estimate from , which has 14.2k rollouts.)

OC, Elf's blindspot makes its winrate estimates for and earlier plays incorrect. But that does not justify the winrate estimate for on only 678 rollouts. In fact, the earlier estimates should be revised upwards for White.

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

Visualize whirled peas.

Everything with love. Stay safe.

There is a remarkable discrepancy between the winrate estimates of KataGo and the Elf commentary. As we know, Elf tends to favor Black by comparison with other bots, at least at first, and its evaluations tend to be more extreme than other bots'. However, at move 181, where a ko fight breaks out, KataGo gives Black a 61% winrate estimate, at that same point, Elf gives Black only a 24½% winrate estimate (with 47.5k rollouts), a discrepancy of 36½%, not counting the original difference at move 1.

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

Visualize whirled peas.

Everything with love. Stay safe.