Awesome 
Did it return any sources?
Yes, those were a bit less relevant perhaps:
1 Like
2 Likes
7 Likes
6 Likes
Seems sometimes hesitant to answer:
3 Likes
Question: Show some examples of groups with surprisingly large eyes that can be killed
Answer: Some examples of groups with surprisingly large eyes that can be killed include a group with only one single point eye, a group with a straight four eyespace, and fights between groups with same sized big eyes where one has more liberties.
Sources
Doug's Go Blog/May 2003 at Sensei's Library
DougSGoBlog/November2003 at Sensei's Library
Alive/discussion at Sensei's Library
Eye versus Eye Capturing Race at Sensei's Library
2 Likes
And there, it found the correct source but failed to write an answer:
Question: What is the largest eye that can be killed?
Answer: I don’t know the answer to the question about the largest eye that can be killed.Sources
DougSGoBlog/November2003 at Sensei's Library
Largest capture that still can't live at Sensei's Library
Make An Eye And Die at Sensei's Library
Doug's Go Blog/May 2003 at Sensei's Library
3 Likes
Question
What is the size of the largest killable eyeshape?
Answer
A group that forms a single eye outside the corner with no cutting points that has eight or more points is alive unless the opponent can form an eye inside or a ko. The largest killable eyeshape is Rabbitty Six.
Sources
2 Likes
Question: An eyeshape with 8 or more points is alive. An eyeshape with 6 points is dead if it is a rabbitty-six. What about an eyeshape with 7 points?
Answer: All eye shapes with seven points are alive.
Sources
Really?
Question: What is the status of an eyeshape with 7 points in the corner?
Answer: An eyeshape with 7 points in the corner can be defended to a minimum of life in seki or can form a ko.
Sources
3 Likes
I’m assuming the tool didn’t generate that graphic, right?
2 Likes
5 Likes
Soon we’ll have a new go ruleset
4 Likes
Answer
Introducing ‘Pantheon Go’:
- The board is divided into four quadrants, each assigned to a different ancient pantheon: Greek, Norse, Egyptian, and Hindu.
- At the beginning of the game, each player chooses a pantheon and places one ‘god’ stone, a special stone with a symbol, in their respective quadrant. The god stone cannot be captured.
- Each quadrant has unique rules based on the chosen pantheon’s mythology, which apply to any stones placed in that quadrant. For example, in the Norse quadrant, a ‘thor’ effect could allow a player’s stones to have additional liberties.
- Players can place stones in any quadrant, but the effects of the chosen pantheon only apply to the stones in that specific quadrant.
- Once a quadrant is completed (i.e., all intersections are occupied), the rules of that quadrant no longer apply, and standard go rules apply to those stones.
- The game continues until all quadrants are completed or both players pass. The winner is determined by the same scoring system used in traditional go.
This new ruleset proposes an alternative form of go that integrates elements of mythology and strategy while maintaining the spirit of the original game.
Please note that while the specific qualities and effects of each pantheon would need to be refined and balanced, Pantheon Go as a concept has not been mentioned in the provided sources.
4 Likes
It’s going to need more logic skills to reason about rulesets:
Question
What would the effects be of a change to the rules as follows: captured stones of your color are removed after your move instead of after your opponent’s move. Specifically, how would it affect the need for a separate ko rule?
Answer
The proposed change to the rules, where captured stones of your color are removed after your move instead of after your opponent’s move, would eliminate the need for a separate ko rule. In this scenario, a player would not be able to recapture the ko position immediately, as their own captured stone would already be removed from the board. This alternative approach would prevent the possibility of repeated back-and-forth captures, thus avoiding the need for a ko rule.
2 Likes
I am the one who run the GPT-4 games on reddit
I’ve seen papers using transformers to predict Go game moves, but the results are not very successful, but those are some time ago, from the research about Othello-GPT it seems to imply it build internal representations on deep residual layers, probably not that different from what Go AI have always known, the global pooling layers are critical in policy heads and value heads output.
I wonder if trained on smaller board and even a partially and playing very bad LLMs with Go can also using linear probing. The GPT-4 experiments seem to indicate to me that it has basic understanding of its own turn at least, and some form of tracking of atari.
1 Like
I heard ChatGPT is much better at Chess than Go. And that GPT-4 is harder to persuade to act like grandmaster than ChatGPT.
1 Like
GPT-4 is at least stable enough not to play illegal moves too early that it would eventually hallucinate (the GPT-3.5 might be within 20 moves, GPT-4 at least hold on longer with enough prompt “correction” maybe last to nearly 40 moves). And they both did not act like grandmaster (maybe in their phony reviews which in details all fall apart), in fact act like beginners or even less (sometimes just suicide, like playing into a tiger’s mouth). They seem to only pick up a very local coordinates (like if there were a C3 in the previous move, the next will be C4, C4, B3, D3, D4, etc. for the most part just randomly pick one increment). GPT-4 has the added “bonus” that it holds on to its own opinions the longer the chatlog accumulated (token start to take hold, and earlier ones fade), and it sometimes just stick to what is obvious wrong in its own inference, where GPT-3.5 or earlier, they just spew out what you want them to say when reaching a certain point (and by that time, you might as well playing against yourself, since your suggestion will just be reflected back).
Chess has the bonus that there are existing look up database so huge and detailed in tree like dictionary fashion that it can mimic, and total moves often are low enough in recorded games where it can simply copy-paste. And the opening legal moves are almost all within these dictionaries.