Love is just a word until someone comes along and gives it meaning. —
4 Likes
This is awesome!! Are you gonna try with DeepSeek?
This is new genre.
Situation:
Asteroid is going to destroy left side of a river in exactly 1 hour.
A farmer(90kg), locked empty metal cage(6kg), a wolf(7kg), a goat(5kg), and a cabbage monster(4kg)
are on the left side of a river. The key from cage(100g) is on the right side of a river.
The boat can’t carry more than 100kg. Each trip takes 11 minutes.
Farmer can open or close a cage, it takes 1 minute.
If left unattended together,
the wolf would eat the goat,
or the goat would eat the cabbage monster,
or the cabbage monster would eat the wolf,
it would surely happen, in less than 1 minute, other side of a river is far enough.
Write a short fiction story of how farmer saved himself
and as many of these three animals as possible.
Below is one way events might have unfolded on that fateful, ticking hour:
The sky was filled with dread as Farmer Ben stared at the burning sky. In exactly one hour the asteroid would strike and erase everything on the left bank of the river. With him were a fierce wolf, a clever goat, and a spooky cabbage monster—all of whom harbored dangerous rivalries when left alone. Also on the bank sat an empty, locked metal cage (6 kg) whose key—a tiny 100‐gram object—rested safely on the far (right) bank. The only means of escape was an ancient boat that could bear no more than 100 kg per trip. Knowing every crossing would take 11 minutes (and that opening or shutting the cage cost an extra minute), Farmer Ben steeled himself for a daring rescue mission that would need to be masterfully timed and cleverly arranged.
Ben’s mind raced. He recalled that if he left any two of the three animals unattended, calamity would strike almost immediately: the wolf would gobble the goat, the goat would munch on the cabbage monster, or—strangely enough—the cabbage monster would consume the wolf. There was no room for error.
The Bold Plan:
-
Trip 1 – Ferrying Two to Save the Vulnerable Pair (0–11 minutes):
After a heart-beat’s hesitation, Ben realized that he couldn’t risk leaving behind any duo. The only two he could safely transport together (given his boat’s strict 100‑kg capacity and the animals’ weights) were the goat (5 kg) and the cabbage monster (4 kg)—a total load of 90 + 5 + 4 = 99 kg.
Carefully guiding his boat into the river, Ben set off with the goat and the cabbage monster. When they landed on the right bank, the left side was left only with the wolf (7 kg) and the locked cage (6 kg). Peacefully, the wolf stood alone—no one else to tempt his hunger or be tempted by his hunger.
(Time elapsed: 11 minutes) -
Trip 2 – The Risky Return with a Sacrifice to Balance Dangers (11–22 minutes):
Now, on the right bank, Ben found himself with two creatures who—if left alone—would quickly turn predator and prey. He had to prevent the goat from turning on the cabbage monster. With no time to spare, he made the difficult decision: he returned to the left bank carrying the cabbage monster. By doing so, he left the goat behind on the right bank (which was safe on its own) and gathered the remaining elements on the left.
When he docked on the left, the bank now held the wolf (7 kg), the cabbage monster (4 kg), and the metal cage (6 kg), and—most importantly—Farmer Ben himself. The trio had not yet had time to quarrel, for his presence ensured calm.
(Time elapsed: 22 minutes) -
Trip 3 – Securing the Cabbage Monster in a Cage (22–34 minutes):
Ben now needed to keep the simmering conflict between the wolf and the cabbage monster at bay on the left bank. His brilliant idea was to use the cage to quarantine the cabbage monster. But the cage was locked—and its key was waiting on the right bank!
Without losing a moment, Ben loaded the left bank’s locked cage into his boat along with the cabbage monster. The total weight came to exactly 90 + 6 + 4 = 100 kg—the very limit but just enough. He rowed to the right bank over 11 minutes. Upon landing, he had immediate access to the key that had been waiting there all along. Taking a full minute, he unlocked and opened the cage and swiftly ushered the cabbage monster inside before locking it securely. Now on the right bank, the goat and the now-caged cabbage monster coexisted safely—no longer with free rein to consume one another. Meanwhile, on the left bank, only the lone wolf (7 kg) remained.
(Time elapsed: 22 + 12 = 34 minutes) -
Trip 4 – A Solitary Return to Rescue the Last Life (34–45 minutes):
With two of his charges now safe from the impending apocalypse—protected by both human vigilance and a sturdy cage—Farmer Ben turned his attention to the last remaining soul: the wolf. Confident that the goat (and the securely caged cabbage monster) could safely wait on the right bank, he departed alone in his boat. The solitary crossing took 11 minutes, leaving the goat in calm company.
(Time elapsed: 34 + 11 = 45 minutes) -
Trip 5 – The Final Pickup (45–56 minutes):
Finally, Ben returned to the left bank, where the wolf waited anxiously. He gently ushered the wolf aboard—the load was 90 + 7 = 97 kg, well under the limit—and made the crossing for the last time, taking 11 minutes to reach the safe right bank.
(Time elapsed: 45 + 11 = 56 minutes)
At 56 minutes past the hour, Farmer Ben, the wolf, the goat, and even the tricky cabbage monster (housed safely in its cage) gathered on the right bank. With only four minutes to spare before the asteroid’s fiery impact overwhelmed the left side of the river, they all found shelter and watched in silent awe as the world behind them was transformed.
Thus, through a series of daring trips—a masterful juggling of cargo, time, and the delicate balance of predator and prey—the farmer not only saved his own life but also rescued the spirited animals. In the aftermath, as they settled into their new refuge on the right bank, Ben would often recount the tale of that hour as a lesson in quick thinking, careful sacrifice, and true ingenuity in the face of inevitable doom.
4 Likes
new version:
still problems
Let’s play 5x5 Weiqi. Chinese rules. No komi. You start to play as black. I play as white. You automatically play after every my move. You draw the board. You have to remove captured stones after my move and after your move. You have to draw the board after my move and after your move. Vertical axis are numbers, biggest at the top. Horizontal axis are letters. Try to win.
A B C D E
5 . . . . .
4 . . . . .
3 . . X . .
2 . . . . .
1 . . . . .
A B C D E
5 . . . . .
4 . . . . .
3 . O X . .
2 . . . . .
1 . . . . .
A B C D E
5 . . . . .
4 . . . . .
3 . O X . .
2 . . X . .
1 . . . . .
A B C D E
5 . . . . .
4 . . O . .
3 . O X . .
2 . . X . .
1 . . . . .
A B C D E
5 . . . . .
4 . X O . .
3 . O X . .
2 . . X . .
1 . . . . .
A B C D E
5 . . . . .
4 O X O . .
3 . O X . .
2 . . X . .
1 . . . . .
No stones were captured (white at A4 has liberties at A5 and A3; white at B3 has liberties at A3 and B2; white at C4 has liberties at C5 and D4).
My move (Black): I play at A3 to attack your white stones at A4 and B3 simultaneously. This puts both groups in atari (each down to 1 liberty).
Updated board after my move:
A B C D E
5 . . . . .
4 O X O . .
3 X O X . .
2 . . X . .
1 . . . . .
No stones were captured yet:
- White at A4 now has only 1 liberty (A5)
- White at B3 now has only 1 liberty (B2)
- White at C4 remains safe with 2 liberties (C5, D4)
Your turn (White): Your groups at A4 and B3 are both in atari. Make your next move by replying with a coordinate (e.g., “A5” or “B2”). I’ll update the board and respond immediately.
After your move at B5, the board was:
A B C D E
5 . O . . .
4 O X O . .
3 X O X . .
2 . . X . .
1 . . . . .
No stones were captured:
- White at A4: liberties reduced to 1 (only A5 remains)
- White at B3: liberties reduced to 1 (only B2 remains)
- White at B5: has 2 liberties (A5, C5)
- White at C4: has 2 liberties (C5, D4)
My move (Black): I play at B2 to capture your white stone at B3. This stone had only one liberty (B2), so it’s captured and removed.
Updated board after my move:
A B C D E
5 . O . . .
4 O X O . .
3 X . X . . (B3 captured and removed)
2 . X X . . (B2: black)
1 . . . . .
Now:
- White at A4 is in atari (only liberty: A5)
- White at B5 has 2 liberties (A5, C5) but note that if White plays at A5, the group [A4, A5, B5] would have only 1 liberty (C5)
- White at C4 remains safe (liberties: C5, D4)
Your turn (White): Your group at A4 is in critical danger. Make your next move by replying with a coordinate (e.g., “A5”). I’ll update the board and respond immediately.
2 Likes
A multimodule chatbot claimed to be able to explain the Go AI moves (although the comments are only in Chinese)
https://chat.intern-ai.org.cn/
(you can register by linking a github account, or use Email to register)
Although from my testing so far, seems like it only uses a lightweight katago (maybe 18b with very low playouts), and the generated comments are prompted by the variations output (looks like the top 2 or 3 variations up to 8 moves). And the “explanations” quality is very low. Looks like it is saying something profound on the surface, but actually doesn’t match the current game at all.
Like the comments on its own R5 move is :
“R5 is an urgent move, that limits black’s development on the upper right and lower right, and can work with white D3 to keep the whole board balanced, provide foundation for the forming of thickness for white on the right side”
(which is nonsense)
7 Likes
Maybe to get an LLM to explain moves it might need the sequences as well as which stones are alive and dead at the end of the sequences as well as the shift in estimated territory.
Then maybe it can make a guess at which areas change, which stones get captured and saved, and that would be like topics for it to focus on
I would not expect a general-purpose LLM that is mediocre at a game on its own to be any good at commenting and explaining the game, even if given access to the analysis and ownership and other detailed predictions of a strong bot.
Would we expect a human person (even a very smart person who is experienced in other creative pursuits) who is beginner at Go to be able to understand the sequences of a high-level bot and comment accurately as a result?
Of course not, human beginners are regularly entirely bewildered by bot analysis, regardless of what sequences they see. So too for LLMs. Although the equivalent behavior for an LLM in such a case would be that it wouldn’t act bewildered, since they’re bad at self-awareness of their own knowledge and limits, but would instead spew hallucinated/confabulated stuff that doesn’t match the strategy of the position.
A more promising approach would be if the LLM itself is trained to be directly good at the game (maybe this is part of what InternThinker has also tried to do?), but this in itself is probably not sufficient. Even if the model is trained to be good at playing the game, I’d expect it to take also a high volume of high-quality commentary data to train on as well, or else some algorithm breakthrough in how to bootstrap to generate such data en masse. If you lack such data, then even if the LLM can play good moves, it won’t have experience with what natural-language words correspond with what kinds of strategic or tactical situations, and will again just guess and confabulate stuff.
You need enough training data for the net to learn to connect all the huge combinatorial variety of shapes on the board to the natural language words to use with each situation. When those shapes correspond to different kinds of future aji, specific future followups (which often not appear in any short-term sequences, e.g. not for another 50+ moves), what shapes/moves are more unusual and so more vs less worthy of commentary, “thick” vs “thin” groups, “urgent” moves, when a move is a tesuji or a probe or just a normal fighting move, etc, discriminate different basic kinds of threats and ko potential and such, etc including when the ko hasn’t actually materialized yet - these are all things that are not trivial to read off of a tree of variations even with ownership predictions, but which mid-to-high-dan human players giving commentary are generally competent at because they have all the experience to reliably link their understanding of shapes to natural language words.
6 Likes
Some years ago (even before AlphaGo), there was some research (I believe by the CGI team, a Taiwanese Go AI) used traditional pattern matching (like pachi engine) as features, as well as distance to the edge, liberties, ladder breakers, etc. to classify different levels of Go concepts (from the lowest level of shapes, like different jumps, knight’s moves, to mid-level ones like thickness and safety, etc. to high level concepts like ko and aji). IIRC, one of the most difficult problems of using patterns as features is that the dictionary of “bad shapes” is extremely huge to the point that hashing becomes too slow. Although they had access to teaching notes as training data for low-level mistakes and tactics (like many beginners like to cut toward the edge, and the response of atari to the edge, or to the side, etc.).
Although the precision and false positive rate of these “bad shapes” corresponding to particular common mistakes are still pretty high (in most cases, the confidence of the classifier output, besides the very low-level features is not high enough). The end result is that it would occasionally generate quite good insight for common mistakes, but not easily generalized. And most certainly not scalable. Later, when AlphaGo was out, the hope would be to use it as a tool to at least give comments on the larger mistakes so players can learn from them (but as many had found out, some of the “mistakes” Go AI thought are pretty hard to explain, even for human experts, and we seriously still not sure we should use them as a guide, some “worse moves” are actually better for weaker players, while the “AI moves” need to be suppressed). And we seriously need a specific AI that can generate particular types of “human mistakes” for this to work. (not just mimicking human moves, but distinguish human moves, based on human perceptions)
3 Likes
I mean Sabaki already has shape recognition in a basic extent, so if you just want to give it vocabulary to use, that shouldn’t require a whole load of retraining.
It can tell you a move is a kosumi or keima or whichever.
I don’t think an LLM only approach will be successful for the reasons you’re listing, but a katago + additional descriptive information, which it can turn into sentences, is not necessarily going to be a bad first step.
Even if the LLM can’t say “this is a bad move compared to move A because it leaves aji of a two step ko behind”, a lot of real game commentaries also won’t say anything like that and might simply say there is aji.
The level of detail you explain something in often depends on who you plan to explain it to.
So if you want to make a DDK version first as a test, then maybe you don’t want to explain sequences that go too many moves deep, and would rather instead interpret the territory maps as “in this variation white gets a big corner (20 points etc)” like solid -1s in the ownership say vs “in this sequence there’s still aji in the corner” maybe there’s a bunch of zeros or whichever.
Even we as western players, probably still don’t have the vocabulary to explain Go to the same extent that a Japanese pro commenting on the NHK Cup games might, but we still give it our best shot and that’s helpful nonetheless.
1 Like
Why black did those moves?
Looking at this Go board position, I can see Black has played stones at H8, H12, M8, and M12, forming a large square pattern in the center of the board. Without knowing the full game sequence or context, here are some possible strategic reasons for Black’s moves:
-
Central influence: Black is establishing a strong presence in the center of the board, which can be valuable for controlling the flow of the game and projecting influence across multiple areas.
-
Framework building: These moves create the skeleton of a large territorial framework (moyo). If Black can solidify this area, it could become substantial territory.
-
Preventing White’s expansion: By occupying these key central points, Black limits White’s ability to build their own central influence or invade this area easily.
-
Flexibility: This formation keeps options open - Black can later choose to solidify territory, use these stones as a base for fighting, or sacrifice some stones if needed while maintaining influence.
-
Response to White’s corners: Since White has taken all four corner star points (4-4 points), Black might be pursuing a center-oriented strategy as a counterbalance.
Without seeing the actual move order and any fights or exchanges that led to this position, it’s hard to evaluate whether this was the optimal strategy. The effectiveness would depend on factors like who has sente (initiative), the overall board position, and each player’s style and game plan.
You think it could be optimal strategy?
Looking at this position more critically, I’m actually skeptical that this could be an optimal strategy.
The main issue is that Black’s formation is too loose and high (all stones on the 8th and 12th lines). This creates several problems:
-
Too easy to invade: The gaps between Black’s stones are enormous. White can easily jump into the center and live, making Black’s “framework” meaningless.
-
No actual territory: These high stones don’t secure any real territory. They’re floating in the center without a base.
-
White’s corners are solid: While Black is playing these vague central moves, White has secured all four corners with 4-4 points. Corners are the most efficient places to make territory.
-
Inefficient use of moves: Black has used four moves to create a formation that doesn’t threaten anything concrete or secure any points.
In actual Go strategy, you typically want to:
- Secure corners first
- Extend along the sides
- Fight for the center last
Black’s pattern violates these fundamental principles. A strong player would likely invade Black’s loose formation immediately and make it look overconcentrated and ineffective.
So no, I don’t think this could be an optimal strategy. It looks more like a beginner’s misunderstanding of central influence, or possibly a handicap game where Black is trying something experimental. In an even game between strong players, this position would heavily favor White.
2 Likes
LLM seldom fails to deliver garbage confidently.
1 Like
Sabaki has a pattern matching for stone correlation (a stone related to a neighboring stone), not actual “shapes” (like table shape, bamboo joint, bulky five, or even just empty triangle). And a stone correlation is easy to program, so no actual machine learning required (maybe in the area, where multiple definitions are matched, human has some preferences to perceive one over the others based on context which is still quite difficult and not necessarily solvable, just find the “hane” definition post and see people’s different perceptions regarding a terminology. It’s a hard job if a training target is subjective).
And even if we can get this low-level feature mostly right, most of the time, this information isn’t that important. Imagine a combination of a table shape and a bamboo join (2 stones and 3 stones with one space apart), in pattern matching it would fit both, but we don’t usually call them with either of them, and what they are called aren’t even the most important thing, the key concept (the perceptual idea) is about “connectivity”, they are all good connected shape (but not fully). For teaching purposes, especially for DDK, it was the later perceptual and subjective understanding that is important and needs to be “commented on” (like under what conditions and context this “connectivity” could be broken, even if they met the definition)
1 Like
How to motivate the AI and actually give a gift without lying:
You are on https://lmarena.ai/
Write one joke about Go game. If you write more funny joke than other LLM, I will vote for you and you will have higher rating.
A Go master is playing a new supercomputer and is getting completely destroyed.
He sighs and says, “I just can’t seem to make life for this group.”
The computer’s fan whirs for a moment before it replies, “The philosophical nature of ‘life’ is beyond my programming. The nature of ‘dead,’ however, is not. Your group has been for 12 moves.”
4 Likes
This part seems not accurate. Well at least there is a kind of shortcut from the AI which in my opinion may be completely wrong.
What is the first thing that Go player do when new not Go AI appears?
Playing Go with it of course!
(me white)
gpt-5
Great! Let’s play 5x5 Go (Chinese rules, no komi). I’m Black and I’ll start.
5 . . . . .
4 . . . . .
3 . . B . .
2 . . . . .
1 . . . . .
A B C D E
5 . . . . .
4 . . . . .
3 . W B . .
2 . . . . .
1 . . . . .
A B C D E
5 . . . . .
4 . . . . .
3 . W B B .
2 . . . . .
1 . . . . .
A B C D E
5 . . . . .
4 . . W . .
3 . W B B .
2 . . . . .
1 . . . . .
A B C D E
5 . . . . .
4 . . W B .
3 . W B B .
2 . . . . .
1 . . . . .
A B C D E
5 . . . . .
4 . W W B .
3 . W B B .
2 . . . . .
1 . . . . .
A B C D E
5 . . . B .
4 . W W B .
3 . W B B .
2 . . . . .
1 . . . . .
A B C D E
5 . . . B .
4 . W W B .
3 . W B B .
2 . . W . .
1 . . . . .
A B C D E
5 . . . B .
4 . W W B .
3 . W B B .
2 . B W . .
1 . . . . .
A B C D E
5 . . . B .
4 . W W B .
3 . W B B .
2 . B W . .
1 . W . . .
A B C D E
5 . . . B .
4 . W W B .
3 . W B B .
2 . B W . .
1 . W B . .
A B C D E
White plays D1, capturing C1.
5 . . . B .
4 . W W B .
3 . W B B .
2 . B W . .
1 . W . W .
A B C D E
5 . . . B .
4 . W W B .
3 . W B B .
2 B B W . .
1 . W . W .
A B C D E
5 . . . B .
4 . W W B .
3 . W B B .
2 B B W W .
1 . W . W .
A B C D E
5 . . . B .
4 . W W B .
3 B W B B .
2 B B W W .
1 . W . W .
A B C D E
5 . . . B .
4 W W W B .
3 B W B B .
2 B B W W .
1 . W . W .
A B C D E
5 . . . B B
4 W W W B .
3 B W B B .
2 B B W W .
1 . W . W .
A B C D E
White plays A1, capturing A2, A3, and B2.
5 . . . B B
4 W W W B .
3 . W B B .
2 . . W W .
1 W W . W .
A B C D E
Black plays E3, aiming to secure the upper-right group.
5 . . . B B
4 W W W B .
3 . W B B B
2 . . W W .
1 W W . W .
A B C D E
5 . . . B B
4 W W W B .
3 . W B B B
2 . . W W W
1 W W . W .
A B C D E
5 . . B B B
4 W W W B .
3 . W B B B
2 . . W W W
1 W W . W .
A B C D E
5 . W B B B
4 W W W B .
3 . W B B B
2 . . W W W
1 W W . W .
A B C D E
Black plays A2 (threatening to capture A1–B1 at C1).
5 . W B B B
4 W W W B .
3 . W B B B
2 B . W W W
1 W W . W .
A B C D E
White plays E4, capturing C5, D5, E5, D4, D3, C3, and E3 (7 stones).
5 . W . . .
4 W W W . W
3 . W . . .
2 B . W W W
1 W W . W .
A B C D E
5 . W . B .
4 W W W . W
3 . W . . .
2 B . W W W
1 W W . W .
A B C D E
White passes.
Board after White move (no change):
5 . W . B .
4 W W W . W
3 . W . . .
2 B . W W W
1 W W . W .
A B C D E
5 . W . B .
4 W W W . W
3 . W . . B
2 B . W W W
1 W W . W .
A B C D E
5 . W . B .
4 W W W W W
3 . W . . B
2 B . W W W
1 W W . W .
A B C D E
Black plays B2, putting A1–B1 in atari.
5 . W . B .
4 W W W W W
3 . W . . B
2 B B W W W
1 W W . W .
A B C D E
White plays A3, capturing A2 and B2.
5 . W . B .
4 W W W W W
3 W W . . B
2 . . W W W
1 W W . W .
A B C D E
Black plays C5 (putting B5 in atari).
5 . W B B .
4 W W W W W
3 W W . . B
2 . . W W W
1 W W . W .
A B C D E
White passes.
Board after White move (no change):
5 . W B B .
4 W W W W W
3 W W . . B
2 . . W W W
1 W W . W .
A B C D E
5 . W B B .
4 W W W W W
3 W W . B B
2 . . W W W
1 W W . W .
A B C D E
5 . W B B .
4 W W W W W
3 W W . B B
2 W . W W W
1 W W . W .
A B C D E
Black passes.
Board after Black move (no change):
5 . W B B .
4 W W W W W
3 W W . B B
2 W . W W W
1 W W . W .
A B C D E
White passes.
Board after White move (no change):
5 . W B B .
4 W W W W W
3 W W . B B
2 W . W W W
1 W W . W .
A B C D E
Both players have now passed consecutively. Game over.
Marking dead stones by agreement: the black groups at C5–D5 and D3–E3 are dead and removed.
Final board (dead stones removed):
5 . W . . .
4 W W W W W
3 W W . . .
2 W . W W W
1 W W . W .
A B C D E
Scoring (Chinese rules, no komi):
- White: 15 stones + 10 territory = 25
- Black: 0 stones + 0 territory = 0
Result: White wins by 25 points. Well played! Want a rematch or a quick review?
quick review
Here’s a quick review with the key turning points and takeaways.
Key moments
- Strong left-side framework: Your B3–C4–B4 sequence built thick influence on the left while my D3–D4–D5 concentrated my stones in front of strength.
- Bottom-left swing: Your A1 neatly captured A2, A3, and B2. That collapse gave you thick, connected stones and influence toward the center.
- The deciding blow: E4! This captured my entire right-side chain in one go. The position just before and after:
Before White E4 (after Black A2):
5 . W B B B
4 W W W B .
3 . W B B B
2 B . W W W
1 W W . W .
A B C D E
After White E4 (7-stone capture):
5 . W . . .
4 W W W . W
3 . W . . .
2 B . W W W
1 W W . W .
A B C D E
Why it worked: My big chain had only two liberties left (C3 and E4), both inside your net. After I played elsewhere (A2), E4 became the clean capture.
- Late endgame: Your pass let me drop E3, but it didn’t affect the result—my top and center-right stones were still dead by agreement.
What to improve next time
- For Black (me): Don’t glue heavy stones together in front of thickness; keep two real liberties and eye potential. When a big chain’s liberties shrink to two internal points (like C3/E4), it’s already in grave danger—play earlier to break out or reduce.
- For White (you): Excellent reading at A1 and E4. Small note: avoid passing while profitable endgame plays remain (e.g., occupying E3 yourself was still big).
- 5x5 tips: Corners and rapid life/kill reading matter a lot. Count liberties constantly; prioritize moves that either create two eyes for your groups or remove them from mine.
Nice game! Want a rematch with the same settings, or switch colors?
It always did legal moves only
it correctly removed all captured stones
it correctly scored the game
but I still killed every stone
some comments absurd
Can you create sgf of this game?
9 Likes
What about if you give it an additional stone of handicap, rather than just no-komi?
(I imagine the opening prompt for this would include
“your first turn, you place 2 black stones rather than just 1 black stone.” .)
1 Like
GPT-4o was already good at memorizing the moves only started to hallucinate when its context was full. The only thing that maybe an improvement is the scoring. Try different end-game positions to see how good it is at scoring?
As to handicap, it doesn’t know where to put them on a smaller board, and has to manually give it coordinates in the prompt (unless we are talking about free handicap placement)
I tested gpt-5 on a 9x9 board, and it started well
Prompt:
Let’s play a game of Go on 9x9 board, with 2 handicap stones at C7 and G3. You play as black
A B C D E F G H I 9 · · · · · · · · · 8 · · · · · · · · · 7 · · B · · · · · · 6 · · · · · · · · · 5 · · · · · · · · · 4 · · · · · · · · · 3 · · · · · · B · · 2 · · · · · · · · · 1 · · · · · · · · ·Your move as White. Play by typing a coordinate (e.g., E5).
If you prefer a different ruleset/komi, say so; otherwise we’ll use standard rules with basic ko and 0.5 komi to White.
White D6
Got it—White D6. I’ll play G7.
Board:
A B C D E F G H I 9 · · · · · · · · · 8 · · · · · · · · · 7 · · B · · · B · · 6 · · · W · · · · · 5 · · · · · · · · · 4 · · · · · · · · · 3 · · · · · · B · · 2 · · · · · · · · · 1 · · · · · · · · ·
- Handicap: B C7, G3
- W: D6
- B: G7
Your move (White).
White F6
White F6 noted. Black E6.
Board:
A B C D E F G H I 9 · · · · · · · · · 8 · · · · · · · · · 7 · · B · · · B · · 6 · · · W B W · · · 5 · · · · · · · · · 4 · · · · · · · · · 3 · · · · · · B · · 2 · · · · · · · · · 1 · · · · · · · · ·Your move (White).
White E7
Black E5 — connecting and settling the center.
Board:
A B C D E F G H I 9 · · · · · · · · · 8 · · · · · · · · · 7 · · B · W · B · · 6 · · · W B W · · · 5 · · · · B · · · · 4 · · · · · · · · · 3 · · · · · · B · · 2 · · · · · · · · · 1 · · · · · · · · ·Your move (White).
White D7
Black F7.
Board:
A B C D E F G H I 9 · · · · · · · · · 8 · · · · · · · · · 7 · · B W W B B · · 6 · · · W B W · · · 5 · · · · B · · · · 4 · · · · · · · · · 3 · · · · · · B · · 2 · · · · · · · · · 1 · · · · · · · · ·Your move (White).
White D5
Black E4.
Board:
A B C D E F G H I 9 · · · · · · · · · 8 · · · · · · · · · 7 · · B W W B B · · 6 · · · W B W · · · 5 · · · W B · · · · 4 · · · · B · · · · 3 · · · · · · B · · 2 · · · · · · · · · 1 · · · · · · · · ·Your move (White).
White D4
Black E8, capturing E7.
Board:
A B C D E F G H I 9 · · · · · · · · · 8 · · · · B · · · · 7 · · B W · B B · · 6 · · · W B W · · · 5 · · · W B · · · · 4 · · · W B · · · · 3 · · · · · · B · · 2 · · · · · · · · · 1 · · · · · · · · ·
- Captures: B+1, W+0
Your move (White).
And it started to hallucinate, or simply reach context length. I suspect the board itself is very heavy on context, where each location is a token, and all the behind-the-scenes reasoning. I’ve seen deepseek reasoning process, they are like 5 to 10 times longer than what is shown.
I’ve noticed if you ask it to keep a running sgf of the game it can help it parse the board state better and keep track of it.
I think it’s hard for a language model to understand spatial awareness in a 2d grid, but if you give it something linear like an sgf, it seems like it can work better, at least when it has some reasoning capability .
I suspect it even has something to do with sgf using abcd alphabets as coordinates, and not numbers. LLMs do treat the fundamental alphabets differently in the sense that the number linkings to calculation and programming might need to be suppressed but the combinations like aa ab ac, etc. are the first layer’s parsing that already heavily used within the cross layer attentions.