The September edition of the European Go Journal featured a very nice interview with Ryan Li 3p. In the interview he talks about a broad range of topics, but what piqued my interest was his remark on using AI.

The context focuses on studying with AI, or reviewing your games with AI, and Ryan Li’s recommendation is to generally use AI less, because knowing what the correct move is in a situation does not help furthering your understanding of the game. Unfortunately, besides his additional remark that AI has made go players lazier, not much more is said on the matter.

It can certainly be tempting to have an all-knowing oracle that points out the mistakes in your games whenever you ask it to. I know I’ve been guilty (as recently as yesterday) of having finished playing games, and just immediately loading them into the AI to see where I lost more than e.g. 5 points. A recommendation I’ve read multiple times is to first review the game by yourself and then check it afterwards with AI. But having the AI there, I find it difficult to motivate myself to do that, and perhaps this is where the ‘laziness’ comes into play. If I did not have access to an AI at all, I probably would be much more diligent in reviewing my own games by myself (maybe an experiment worth trying for a month).

It's not just when reviewing games that I notice the 'convenience' of AI. When watching live-streamed tournament games, I personally absolutely abhor the presence of the AI winrate bar. Not only does it completely spoil the outcome of the match, it also gives you less of an incentive to analyse the board for yourself and try to figure out who is ahead by using your own understanding of the game.

I was wondering how others have experienced this. Do you feel that AI has made you lazier as well?

AI for review is somewhat analogous to using a calculator to review your mistakes on a math exam. The calculator can give you the right answer, but it's not going to make you much better at doing math problems.

Sometimes playing around with a calculator can give some insights, but nothing beats practicing your ability to think.

_________________
be immersed

A go game is not a math exam. If you think hard enough you can determine for sure if your answer to a math problem is correct or not. On the other hand, in a go game even an AI is not 100% sure that its move is correct. In addition, an AI can show me moves I wouldn't have thought of, while a calculator rarely gives new insights.

That said, I agree that AI is a double-edge sword. On the one hand it helped me a lot to answer questions that I continually had like "is A or B better?" or "I really don't know where to play, what can I do now?" and to point out misjudgments. On the other hand I admit it made me lazy, I often don't spend enough time checking the AI's answer and exploring variations.

In my case, AI has made me more studious.

Not only does AI tell me the right move, but it also
- gives me other candidates
- shows the main sequences it "sees"
- tells me "how bad" my move was in comparison (or actually pretty ok)
- allows me to explore paths that I had in mind but it didn't consider
- calibrates my positional judgment

It basically tells me where I won or lost the game and what I could have done about it. I find that very motivating.

In the past I would do game reviews, sometimes with stronger players. How could I trust their judgment? My own? It was perhaps a more creative activity but not necessarily a more studious one.

For creativity, reading skill, ... I can still play games and solve tsumego. For reviews, I think AI is perfect. Or it suits me, at least.

I wouldn't take issue with any of that, except perhaps the last item - that may depend on how each of us defines positional judgement.

But a guy with a whip could also make you more studious. What you haven't told us, and I'm intrigued to know why, is did AI make you stronger.

For me the benchmark is the well known case of T Mark Hall. He said (and I know it's true) that transcribing the games of Go Seigen (at that time I think it was about 750 games) got him from 2-dan to 4-dan (and British Champion). I can't remember how long it took him in the course of real life, but he could do about 3 games an hour. (Turns to calculator...) That's 250 hours. How much demonstrably stronger has AI made any dan player here in such a short space of time?

(Note that transcribing from one or two diagrams is very different from playing over a game in Go World. When transcribing, if you want to achieve any decent speed you have to learn to "guess" where the next move will be. Mark excelled at that. And conversely, when he transcribed amateur games he had to spend sometimes more than twice as long because he had no idea where the next crap move would be.)

Yeah, most analogies aren't perfect. The main similarity I see here is that both calculators and AI can give you an answer without requiring you to think.


I'd say that depends on the math problem.


I think using online calculators like Wolfram Online can be useful for playing around with data and observing mathematical properties. *shrug*

Anyway, the main point I want to convey is that both calculators and AI may allow you to get answers with or without thinking. And thinking is the important thing to exercise here.

_________________
be immersed

AI does not make me lazy because still I could not buy a graphics card for an acceptable price yet (prices have been 1.5 - 4 times MSRP for over a year). If I had AI on my computer, I would only use it for specific study while I would continue to trust my own skill of positional judgement, endgame, tactical reading, strategic planning. During one's games, one cannot / must not ask AI to make such thinking but one must be able to do it on one's own. Therefore, my thinking would not become lazy. Besides, I enjoy thinking! During AI study, I would also reflect why AI suggests certain moves.

I don't need AI to make me lazy.

I was comparing the math student with the go student. The math student is generally confronted with problems that have a short solution, and can prove in a significant % of cases that his solution is correct. When confronted with a middlegame go position, the go student can almost never prove that a move is optimal or winning. When reviewing a game by yourself, you may think in circles and debate whether A is better than B, while AI shows you that C was even better.

In which year did he become British Champion? I can't find him on the list

https://www.britgo.org/tournaments/bchamp

Unfortunately, AI does not show you that C is better. I know you know this, but AI can only show you is that C is better for the AI, not for "you." If there were a way to supervise the AI learning based on your own game records then the AI could attempt to find a move that was better for you. I don't think that idea is very realistic because the training set would be too small.
--------------------
I find AI to not be so helpful because many of my bad moves are not labeled as being bad (or not so bad) because the AI can find a way to make them work. Whereas moves that are labeled as being good are actually not good (for me) because I would not be able to make them work. Similarly, some moves that are "good" (for my level, against my opponent) are labeled as bad because they would be bad if my opponent were the AI.

I do find AI useful for identifying points that were honte. What I have noticed though is that the players themselves, even at my level, will often identify the honte (or something suitable) sooner or later. So then, how often does the AI find a move that I would not have found by my own review that I would actually be able to use? ... Geez, I don't know if I've ever even seen such a move suggested by AI. And I'm not sure I would actually learn without making the discovery myself. Though I'm sure that pros love to find these moves and are able to learn from it.

At least for me, using AI would not only be lazy, but mostly unproductive.

Most of the time, AI shows me moves that I understand like:

• Make a kikashi first, then make the move you planned
• A point which is the common base of two groups is very big
• You are too/not enough territorial
• Force on the outside before living locally
• Cutting stones, even if they die, can be useful to get forcing moves.

That I understand these moves doesn't mean they always come to my mind, and doesn't mean I would have found them with self-review witout AI.

Some moves are, or were, really off my radar like those I recorded in this thread: https://lifein19x19.com/viewtopic.php?f=48&t=16894

He won the British Open several times (2000, 2003, 2004, 2008):

https://www.britgo.org/tournaments/britishopen

Though I am not sure of the difference between the Open and the Championship. Is the British Championship restricted to British citizens, while the Open has no such restrictions?

_________________
And the go-fever which is more real than many doctors’ diseases, waked and raged...
- Rudyard Kipling, "The Light That Failed" (1891)

Ah, I found a more detailed post about T Mark Hall: https://lifein19x19.com/viewtopic.php?p=210499#p210499

In particular



Let's say 1000 games in 180 days, that's about 5.5 games/day. If 1 game takes at least 30 minutes to transcribe, you need to spend at least 2 hours and 45 minutes a day (more likely over 3 hours/day), 7 days a week, for 6 months. Hard to reproduce for most people.

There's no particular reason why I didn't tell you other than I tried to answer the original question: has AI made us lazy? In terms of studying Go, it didn't. In other aspects, maybe, but I don't think so. I see it as an enrichment and a lasting one at that.

Now to your question: it's a tough one to answer. There are multiple things at play here.

First, there is the question if I, at 50 years of age and half as old at Go, can still improve after being relatively stuck at a rank that averages at about 2 dan, depending on the calibration of this or that server.

And this is the second, biggest problem: how do you measure improvement? I used to be a stable 2d in Belgium, Europe, measured by real life tournament performance: On OGS I'm 1d. Have I become worse? On KGS I'm 3d. Have I improved? Could it be that I have improved at Go by studying with AI but so did others? Have the bots deflated the ranks? I can't tell if I have improved from my online rank evolution.

If online ranks don't tell the story, can I judge for myself? I feel that my understanding of the game and my move selection has improved. My reading skill, or discipline, hasn't. That could be an effect of studying with AI. It could also be a narrative.

So I revert to what I can more objectively measure: I have played and analyzed more than 120 games this year. That's more than any of the previous ten years.

Now another story dates back 20 years. I was hosting another Go player in my house and we played a kind of jubango, pushing each other to higher or lower handicaps. A friendly rivalry it was. I would analyze those games very intensively and take lessons learnt into the next game. After 6 months of that (and pushing the handicap to 3 stones) I jumped from 1k to 2d in the national ranks pretty quickly. Was it the method? Or the intensity? Or the fact that I was 30 years old and only 5 years into playing Go? I'm not sure.

I am not sure AI makes us lazy. On the other hand it surely does let us indulge in overanalyzes.

When we review games it is often that we go through the game move-by-move questioning everything or we stop at a few places and spend an hour trying to come to some conclusions. It is especially common when a group of players sits down after their match in tournaments, but it is clear that many people do this on their own (I have found myself doing this). I never found it useful to ponder over a game and I have started calling it to overanalyze.

The best review is always a quick review that gives you some idea why the game was won/lost, if you are making the same mistakes as usual and if your important decisions were good (and what you could have done if they were not). Many times the only thing to learn from a game is "never make that mistake again".

First when I started reviewing my games using AI it was just this kind of overanalyzes. I hooked up a server with multiple brand new GPUs (only to learn that I could only use one at a time but it was fast enough anyways) then I downloaded the best leela weights and would just step through the game having the AI ponder on every move. Fairly useless, the top weights easily find fault in normal play and more playouts usually means their logic is irrefutable (except that they can't always find a logical resolution in very complicated local situations). Once you see their moves you just can't un-see them and your chances of learning anything are diminished.

This seems to be similar to how many players use the computer. Just have it do the work for you and then try to understand its reasoning. Be very confused and not try to discern something that you could actually improve.

Later I discovered that if I used weaker weights that they would actual agree with most of my moves and if they were not good they would usually not refute them with too crazy lines. So I started using Katrain on my laptop (that has bad GPU that is fast enough per-se but overheats and slows down after awhile). Now I find it very useful to run analysis to 250 playouts for every move and looking for obvious errors, something around -1.5 points seems to be a big mistake at my Pandanet level most of the time, if I have time (and the ambient temperature where I am is not too high for the laptop to cool off) I'll do a second run to 2500 playouts for every move to make sure the mistakes are actually mistakes (they are not always). If there is something interesting, then I will try to playout some variations (as opposite to thinking about them) to see if I can find a reasonable way to play that I might not easily think of or be unsure what to think of.

Sometimes I just throw the game into Katrain and start another game. Even though that often means that reviewing it will not have very high priority later. That seems to be a good thing because playing more is better than reviewing more.

Someone might say I am obviously lazy now that I have figured out my AI analyzes It is just that I really think that you are better off finding just a single mistake in every single game than you are finding improvements for all your normal moves.

I do review every game nowadays but thanks to KataGo's sizing of the gap, I tend to focus more on the big swings in the estimated score, whereas in the past I may have dwelled on choices that didn't really matter. I find my game reviews have become much more productive.

Good to see the correct use of 'temperature' for once in relation to a Go game.

I don't doubt this example at all, but I wonder if something is awry with the description of timing. T Mark was a fairly longstanding 3 dan (looking back at the online British Go Journal this seems to be since at least the early 80s, if not before). His promotion to 4 dan was in the early 90s I think (1994?).

You may be right and, if so, you can safely assume it's because it involves numbers - they scramble my brain. One difference between Mark and me that we often commented on was that, although not a mathematician, he was a numbers guy. We both travelled a lot as part of our work - somewhere between 60 and 90 countries. He could tell me the exact dates he went to each place. I can't even remember if I've even been to certain places. For example, Houston and Atlanta (which I pick just because of the current World Series). I'm certain I've been to one but not sure if I've been to both. Mark would have been able to give me not just places and dates but his hotel room numbers. He kept all his boarding passes and hotel stationery samples.

I also spent a full ten minutes yesterday looking for my house key seconds after opening the door with it. It was in my pocket. I actually think I have a very good memory. I just choose to use it on things that matter to me.

And using that memory, I was able to dredge up a description that Mark wrote about himself. It may be of interest to younger people here who have heard only the name: