Kira linked this KataGo match game in LZ discord, which I'm relinking here because of how it highlights quite starkly how different that high-level AI style is when playing with group tax, i.e. "ancient" scoring rules (minus 2 points per living group, the two eyes a group needs to live are not counted as territory). Far less territorialness than people associate nowadays with AI play, and much more emphasis on the center, because of the extra incentive to have fewer groups than the opponent.

And lots of interesting different local moves or joseki, such as playing the *long knight* back-off in response to 4-4 knight approach, presumably taking advantage of the fact that 3-3 is less attractive to the opponent and therefore an extension that leaves it more open is fine.



Here's another one showing off a lot of the same patterns:


And a less exotic one, but still: no early 3-3 invasions, yes early two-space high pincer, some other differences.

The first thing that caught my eye was the choice of large "shimaris" as a dominant feature of the opening. But they were not seen as shimaris by the ancient Chinese masters, and the main concepts to do with the corners came later in the game with the ideas of sharing corners (分) and encroachments (侵). I infer, therefore that katago has "seen" something much deeper in the game, to do with the corners, and much deeper than 3-3 invasions.

I further notice that, like the old Chinese masters, katago is happy to go up into the centre early. But the Chinese mostly limited themselves to what we (bt not they) call a one-space jump. Katago seem to like the large glide (knight's move), which is not known in China but there is an example of this sort of play from Edo Japan, as below. This was a handicap game with Yasui Senkaku (he of cosmic go fame) was playing the young Ota Yuzo.



The shape I referenced above is the one in the lower left.

The final point I noticed was that the 9-3 points did not feature in katago play in the first two examples. They therefore look more like early ancient Chinese games. The last example does begin, in that respect, to look more like the go of the top masters of later times. It took a long time for 9-3 point theory to evolve in China. Is it possible that the last katago example was a deeper training run?

The two space high pincer at the opening stage is certainly a novelty in any old go.

Very cool, thanks for the input and the background! Always fascinating to see these things placed in historical context. As for your question:



These are all games from basically about the same point in the run. In any case, I would not expect move preferences to necessarily evolve in any particular way corresponding to how they did for humans - if it did, in large part it would be by chance as far as we're able to predict, because while bots have validated and resurrected a number of older ideas, and also validated a number of modern ideas too, they've also rejected just as many other human ideas, both old and new, coming up with their own judgments and internally-consistent style.