How big to avoid AIs?

My question is more aiming at computer experts.

At which size of the goban would it be impossible to get a good AI like what we see on 19x19?

(Not debating on the human side)

Are you asking in practical terms or theoretically?

1 Like

The modern approach to building strong AI engines is not fundamentally limited to only 19x19, but can be generalized to larger boards as well.

KataGo is an engine that supports custom board sizes. Here is a post from the author of that engine talking about its performance on larger boards:


Of course, but surely not to whatever size. So when it will be too much ?

1 Like


On a tangential note, lately I’ve been thinking about what strategy would be like in a Go variant in which the board is unlimited and the bounding factor is the number of stones.

(Not my idea originally, I must admit, it’s been on SL for years)

Phrases like “too much” and “a good AI” are quite subjective and your question probably depends entirely on how much time and money one wishes to spend on the development and training. There’s not a hard boundary between possible and impossible, and it is rather a question of funding and motivation. Quantifying any of this with would take substantial experimentation and research.

1 Like

No it’s clear question. Good AI means can beat Lee Se Dol. Too much means that
Today computer can’t.
In 10 years… In 100 years… In never.

I’d like to know what is the feeling or even better the proof if any. Especially by people who works or worked in that. Or mathematicians, physicians…

At least it could be specified how each additional line makes the elaboration of the software harder.

Perhaps good AI is possible for any size. After all, why wouldn’t it, ufufu.

It might be unlimited. With a large enough board, groups far away from each other will have no influence on each other.

I wouldn’t be surprised if current day Go AI already surpasses professional players on 25x25, and it won’t take long before a more general AI is developed that can play not just Go on any size, but any kind of abstract strategy game on any size at professional level.

It should also be mentioned that board size or complexity of the game tree are not the largest limiting factors anymore. AI nowadays relies a lot less on deep search algorithms, like MCTS. Most AI is nowadays based on training them intuitively: let the AI play and correct itself based on random guesses.


So what is limiting then? I mean Ai is only like 2 stones stronger.

Now there is a global enthusiasm about the unlimited performance but but.
Any serious reference on that? Somewhere?

Bots can beat most pro on StarCraft 2 and Dota 2 now - not grid based games
these games have much more “possible moves” than 50x50 Go.


Training time is limiting (in fact, it’s all about the money / energy it costs to have the GPU’s running on the training). Bots will get stronger when trained longer, but the speed with which this happens has a limit. With new techniques that are developed in AI research, this is improving a lot, to the point that professional strength Go AI can be trained in a day by a company such as Facebook (that’s what happened with ELF after all).


Really long ladders?

1 Like

Ladders are already deadly enough in 19x19 to be avoided. AI can learn how to avoid ladders in the same way humans do, without reading the whole thing out and instead skipping most of the ladder only to inspect the ending.

With a large enough board, a ladder will never be worth it.


Yes, i was objecting slightly (and ineffectively) to the idea that groups far away have no influence on each other.

In a way this holds up: with a larger board there are a lot more ladder breakers to be played as well.

1 Like

Well Ai can pick up and learn really fast the chances of humans surpassing Ai is technically impossible

I don’t know about AI, but there surely must be a maximum board size for humans to play good on. If only because a human has a limited life span and a limited reaction speed to play a move.

For example, when playing 1 move per second for 120 years (roughly the maximum human life span), a human would be limited to a game of roughly 3.8 billion moves. So that mostly rules out boards bigger than about 65,000x65,000.


I don’t think there’s any reasonable point where we would expect AIs to become worse. There’s likely going to be a point at which the AI architecture would need to change in order to restrict board conception to be a bit more local in scope and train a good engine for assessment of an arbitrary “block” rather than training a network sized around the whole board, but hat’s already been done in a fundamental sort of way with Starcraft.

I think that human ability to effectively play go will degrade far faster than AI performance as board size increases. My instinct is that a reasonable AI could be constructed whose memory footprint & processing requirements increase logarithmically with the size of the board, focusing in detail only on a few sections, but keeping a rough tally of how important the other sections were when last investigated. Humans would likely struggle to determine when to tenuki and what the biggest parts of the board actually are.