I decided to join in the experiment. My process was:
1. Download the uncommented game record from KGS, make diagrams, lay out the stones on a board. While doing so, I wrote down some impressions:
- Moves that looked like clear mistakes
- Moves where I would have played differently but I don't think it makes a big difference
- Moves where I struggled to choose between two alternatives and thought it would be worth asking an expert such as KataGo for an opinion
2. Review with Lizzie+KataGo, check the things I noted from step 1, and also pay attention to moments where the recommended move was in a completely different place from where I was looking
3. Look at Knotwilg's commentary to compare impressions.
There was the usual crop of surprises:
- Things I rated as big mistakes that KataGo said were actually OK
- Ideas of my own that turned out to be awful
- Shapes that were just not on my radar
- Suggestions pointing to positional judgements that seem to go against received wisdom
Some specific examples:
Black 9: My first impression was that it's an awful move! But KataGo rates it equally with the joseki move. Looking at some variations, and thinking about it, I can rationalise this like so: black can look after the groups on each side independently, so being split costs nothing. On the left, of course there's the 3-3 invasion. On the right, I'd forgotten about black's tesuji with the P4 attachment to help settle the group. I've seen this move before, but it didn't come to mind today. So it's good to be reminded about that shape.
Black 19: Again this looked very bad to me, but it's actually KataGo's first choice move! The problem is the followups: after move 25, black has thrown away any hope of invading at C5, and the evaluation drops. Psychologically though, the faulty thinking probably happened on move 19 -- the AI evaluation on move 25 is flagging an error of judgement from six moves earlier.
White 21: Knotwilg asked about N5 versus N4. Again N4 just wasn't on my radar, and I was surprised to see how much KataGo likes it.
White 28: I wondered if white could get away with extending one point further. KataGo said yes, it very slightly prefers the game move but either is OK. But the surprise: KataGo also had a slight preference for tenuki (approach at top right) rather than playing either extension. This seems to go against the proverb "urgent before big". I explored some variations, and indeed white could end up defending a weak and baseless group at the bottom left, but it seems to be worth it considering the points gained at top right.
This might be an example of "an AI move" that humans, or at least amateurs, shouldn't imitate? Getting a good result is conditional on being able to look after the weak group, and it's often easier to find attacking moves, harder to defend well, so the fight might favour black in a game between two people who aren't as strong as KataGo? Hard to say. In a casual game, I might try out the tenuki as a fun learning experience. In a tournament, I'd be more inclined to play the extension, and stick to patterns that I [think I] understand.
White 38: Knotwilg's attachment surprised me, but KataGo says it's not much worse than the usual extension. The other surprise here: white can extend four spaces instead of the usual three (i.e. J17 instead of K17): either is OK according to KataGo. Wait, if you extend too far, don't you just get invaded? Yes, but you make a base at K17+G17, hurt black's top left corner, take a bit of damage on the right, and gain as much as you lost.
That's another "AI move" that I'd be hesitant to try in my own games. But looking at the variations, it's helping me to think more flexibly, and to practice judging the position in terms of trades, not just static territory.
White 46: This is one that I hadn't noted for investigation, but "the blue move" popped up in a place where I wasn't looking for it. The game move is fine -- white has the initiative either way -- but KataGo suggested attaching on top of black's one space jump, playing F14 instead of F16.
Is this violating another proverb, "don't attach to weak stones"? The variations get interesting: if black looks after the two stones below, then the top group can end up weak, since white has the attachment at G17 to play with. This is an idea I actually would feel comfortable trying out in a serious game.
Black 51: I correctly listed this as a mistake, but my alternative was nearly as bad. I didn't see KataGo's recommendation, the shoulder hit at D11. So here's another not-entirely-new-but-forgotten shape for me to relearn.
Black 95: On my first pass, I thought black was clearly lost by this point. But here, KataGo says it's still close, and shows me some very nice defensive techniques (am I allowed to use the word "shinogi"?) starting with an attachment at O5. Well, really, it's as simple as crosscutting to make complications in an area where you're weak. But it just didn't occur to me in this particular context.
All up, I've listed eight examples here. Moves 21, 51 and 95 are the ones where a big drop in the score graph pointed to something interesting. The other five examples are not that significant in terms of the raw numbers, but I'm learning things that I think are helpful. The big swings are a useful starting point, helping me find my bearings so that I don't get lost exploring thousands of variations, but they're not the full story. Note also that some of the lessons are around judgement and direction of play: we can learn more than just tactics from the AI.