Predicting a pro game has instant feedback every move of the game. Reviewing your game with AI too, although the lessons can only be drawn in bulk after the game.
Move prediction through transcribing certainly does give feedback on every move, and this type of learning has a deeper value, too - one usually overlooked in other fields. Namely, for each move you are getting data about the exact frequency in which each type of move occurs, and by extension data on the exact frequency of contexts in which each move occurs (and all of this is being done in a way that suits the way your brain's neurons collect data).
As an example of this, one thing I noticed years ago (and wrote an article about it somewhere) was that the type of move most likely to trigger a resignation was a monkey jump. I confirmed this by writing a program to show on which line resignation-inducing moves occurred and finding examples with Kombilo. But because I was intrigued and sensitised, I started noticing (from transcribing) that monkey jumps were often ignored. So I became aware of context. Although I had some fairly reliable data to show that at least considering ignoring a monkey jump was fine, in in terms of what we might call "go knowledge" I didn't really understand this, and so would never ignore the damn simian. But I did find the explanation eventually, in O Meien's celebrated endgame book. So my intuition ("can ignore") was right, and I was wrong to doubt it.
But there was an interesting additional take-away from that episode, for me. Being a words guy, I started wondering whether having a name for each kind of move was a way of optimising how to learn by this process. In other words, if a type of move didn't have a name, would I notice/learn it no matter how frequently it occurred? My gut reaction is to say yes, but with the reservation that I might not learn it as thoroughly as when it did have a name. But I still have niggling doubts. It may simply be that when a move type has a name, you can see it talked about in books or you talk about it with other people, and all this talk has the effect of strengthening the associations already existing within the brain. But it has nothing to do directly with the actual frequencies noticed (subliminally) when transcribing. This certainly feels true of the monkey jump.
This seems to be potentially important in the case of studying with AI. If you want to play like an AI, the correct procedure seems not to review your own games with it, or to study pro games with it, but to transcribe moves from a single diagram of games in which AI plays AI. But AI plays moves we find strange, and because they are strange, we don't have a name for them. It seems our instinct is then to give names to them, and the prime example has to be Direct 3-3. But other moves have defied naming so far. Does that matter? Again, my gut reaction is to say yes, but with reservations.
One reservation became clear to me when looking at the book chess book "Game Changer" that dfan mentioned somewhere. He suggested, I think, that this proved that AI can already offer specific ways to improve in chess, and by extension go. My impression, after just an early reading of the book, is that the authors (Sadler and Regan, the latter being a British go player BTW) don't claim that in the manner that we want to take it. In other words, it won't work that way for oldies like most people on this site. It will work, however, but only eventually, for youngsters who come into the game with clean slates of a brain and who learn by putting in the hours replaying AI games instead of the human pro games we grew up with.
I further get the impression that Sadler and Regan believe that this process for the newcomers can be enhanced by identifying now (i.e. naming) the most distinctive and important chess concepts that they think they can detect in AI play (and do note they take as their basis AI versus AI games, not AI versus humans). As with the French Revolution, it is too early to say they are right. But I find their approach interesting, with interesting implications for go.
What S & R do, as dfan mentioned, is to identify core concepts (rather than move types) such as opposite-side castling, or outposts connected with piece mobility, or advancing rook pawns.
This made me aware first, that chess and go differ markedly in how they use names. Move types are not named very often in chess, whereas just about every move type can be named in go (this may contribute to the over-obsession with shape that many go players have). Even when a move type does seem to have a parallel in both games, the associations are quite different. A knight's move in chess is mainly just to do with explaining the rules, but in go it has to do with attacking, waist-cuts, or even (for some of us) to do with disagreements between Shusai and Kitani. Moves types in general have very rich associations in go. In chess the associations seem to derive more from wider concepts.
But I am wondering now, in the AI age, if we should be following chess. For example, we still talk of shoulder hits even though the AI use of them tends to baffle us (and pros). Maybe we should be inventing concept phrases such as (for the old shoulder hit) "creating an early outpost in the centre". In a way, this is what Direct 3-3 does: it stresses the time element rather than the shape element. I would also argue, incidentally, that the ancient Chinese got there first with their "call & response" concept (zhao ying), and I would further posit that amashi is a Japanese equivalent (i.e. an example of a concept rather than a shape, though in this case with the caveat that Go Seigen thought the Chinese genius Huang Longshi was the first master of amashi).
This urge to name does seem universal. I mentioned elsewhere that shogi players learnt to favour early pushes of the side pawns (the rook pawns in chess), something that horrified generations of chess masters but has now become common. The S & R book told me that a British chess master (Simon somebody) has given the pushed king's rook pawn the name Harry (from being on the h file). I think this is a great example of our naming urge, an urge we probably should extend to go, but with an emphasis, like chess, on concepts rather than shapes.
Going back to my intro, what I meant by repetition having a deeper value that is often overlooked in other fields can be illustrated from my own experience in languages. I remember from early schooldays having to learn the names of flowers in French. One that stuck in my mind was glycine = wisteria. I had no idea what a wisteria was (no internet in those days, no bookshop in my home town), and I never found out until I saw Japanese hanafuda playing cards. Since I later discovered that it was named in honour of Caspar Wistar, and so should really be wistaria, it has a firm, but useless, space in my memory.
But when I was reaching Japanese at university, my students were all naval architects (who only wanted to read Japanese, not speak it). The approach I took, therefore, was to handle the grammar with special look-up tables and for the vocabulary they simply translated technical articles in their own specialised fields. This meant that they only ever encountered vocabulary that they wanted in, each term in the exact frequency it occurred in real life. This approach (using translation as an equivalent of transcribing in go) was so efficient that the course took only 49 hours. At the end they could read technical texts quickly and accurately, though with the back-up of a dictionary (Katago?) and the grammar tables.
Would that we could get good results so quickly in go - but maybe we do, already? Is that - playing over pro games - why the likes of Sumire and Fujita Reo got so good so young? After all, when they were five so they could hardly read books or understand pro talk. I remember being with five-year-old Liao Xingwen. He couldn't read much text but he was a whizz with go diagrams, and kept books full of them under his pillow.
