Like for example a 3kyu vs a 6-7dan bot?

I don't think it's very usefull (just my opinion). What is really usefull, is to use a bot to review your games.

By that, I mean to look in your game for places where the bot didn't like your play, and explore why it don't like it and how to punish your mistake (if your opponent didn't punish it). Look where you didn't punish your opponent mistakes, and how to take advantage of them. Use the bot to ask questions "What if I play here?" "What do you think the correct continuation should be?" "But if my opponent play here, how should I answer?".

First perhaps better describe what you mean.

That's too large a difference in strength to be useful playing even games.

a) Can the bot be set to play not quite as strong on your hardware?
b) Would you be willing to try taking say 3 handicap stones and immediately contest the fourth corner (where the bot will almost surely play) with the bot set to say 2-3 dan?

I slightly disagree (just slightly) that the bot that useful showing you where you should have played against a significantly stronger opponent. That opponent probably made that clear by its next move (made clear what you had failed to see)

Sure it helps: take handicap against it, change the handicap after each game, challenge yourself to reduce the handicap over time.

_________________
Sorin - 361points.com

Personally I have never been able to improve against a bot for the following reasons:

• When the bot plays fast, it's hard for me not to follow my opponent and play fast as well.
• When the bot plays slowly, it's a bit boring to wait forever that it makes obvious moves.
• Bots always play with the same style, it can become boring.

If the above points are not a problem for you, than a strong bot could be helpful if, either you play with low handicap and accept to lose most games, or play with a high handicap if the bot still makes "normal" moves and is not too passive. For instance, Hirabot 6d on KGS is quite bad when it gives 6 stones.

However all three of these can apply to when playing against a human opponent (the third if always/usually playing against the same human). The first of these, you really need to train yourself to resist.

I think it's useful. I got to SDK mostly by playing gnu go, which was about 5k at that time. At first I played at 9 stones until I could beat it, then worked my way up. Today's bots are much better teachers.

_________________
it's be happy, not achieve happiness

There are more than 200 weights files for Leela Zero now, in a range from 30k to 9d. You can change the one which is 2-3 stones stronger than you and play with handicap.
Or you can play even games, and after 2 wins increase weights file version, after 3 losses decrease it.
I am 2k KGS and now my LZ opponents are in the range of 50-60, for fast even games.

On the one hand a (super-)pro-level AI is like having unlimited access to a professional who will never get tired of playing with you or taking back moves. But he will only play moves, never explain why he plays them. So you will need to explore many variations until you begin to understand.

But human opponents, and even human pros, can play over-aggressive moves or make weird overplays or play other bad moves that you can use as an opportunity to learn. A human amateur opponent will make those moves by mistake; a pro will make them to test you. An AI will never do that.

So I'm not sure what to think of playing AIs. Of course they are very useful to review your games. But lately I've also read about European professionals and strong amateurs who say that they have learned from spending many many hours with Leela Zero.

Some notes of caution:

- Playing online is often synonymous with not taking time to seriously count during the game. This is especially true if there is a clock that limits thinking time. One thing that bots are pretty good at is quickly assessing the score. So you may find yourself losing close games because you didn't have time to properly evaluate the position.
- It has been stated on this forum that bots often will chose a suboptimal move that guarantees victory instead of the "best" move. This could lead to misleading conclusions if you are trying to learn from the bot's play.
- Bots are also prone to errors. Uberdude has found cases where super strong bots went awry after misplaying a ladder. Therefore, make sure if you repeatedly play a bot, you are trying to win by improving yourself and not by finding the bot's fatal flaws.

The OP implies that playing humans is good for learning to start with.

I don’t think it matters whom you play. You learn most in the analysis stage. Having a tireless pro at your disposal is invaluable. It speaks in sequences though, some of which are hard to understand at the lower levels.

Volunteers to explain them in human language are available here up to the amateur dan level. Beware: we tend to overexplain and be too confident in what we think we know.

These days I no longer review games without Lizzie. I can borrow her sequences and try making sense of them for myself and others. I recommend to review your own games and use Lizzie in creative ways. If you don’t get her, try us.

This is so true!

In fact, I modified my go software so that I am not allowed to play before some time is elapsed (I set it at 10s). This way, I am forced to play slowly against a fast playing bot

_________________
I am the author of GoReviewPartner, a small software aimed at assisting reviewing a game of Go. Give it a try!

I invite the OP to have a look at my latest analysis in my own practice blog.

BUT -- this is NOT just a bot problem. You can experience exactly the same thing with a human opponetn whose playing tempo is very different from your own. So need to learn the same skills to overcome the problem.

I agree, this is a skill to learn or work on.
For some reason, I don't have this issue with "over the board" games.

_________________
I am the author of GoReviewPartner, a small software aimed at assisting reviewing a game of Go. Give it a try!