Use Sora, out of 8 generations, this one, although it looked a bit cartoonish, but has the most plausible board position (the board dimension is still off at 14x13)
You are certainly talking about the overconcentrated black stones in the lower left!
That and the fact that a lot of them have the same or nearly identical faces.
That is the most human-looking LLM-generated board position Ive ever seen!
20 x 18 and 18 x 16!
I’ve tested this as well, and found out that it actually doesn’t “understand” the context that deeply. For example, if I use a Chinese page explanation for it, it would just redraw a Chinese character-look-alike page
If I use the negative image, that is switch the black and white of the original prompt image (black background, white words), it also just redraws the texts
I’ve also tried Sora, and it behaved similarly, and it was only able to draw images of Go games if the “descriptions” are clear English (they cannot have any images or too many foreign languages, like one half English, one half Chinese). And interestingly, it would also create schematic-like images.
And if I put all the texts from the image (From Go wikipage page) as part of the prompt as : draw using the following description : “Go (game) … full texts”. gpt-image-1 actually doesn’t draw a plain Go board, but more or less just like a short prompt with “draw a go game” directly
This led me to think that it must have just extracted the embedding and then recombined that short embedding with the prompts. And after some experiments, I think I have found the equivalent embedding - “a detailed Go game explanation” and I test it with the prompt “draw a plain Go game using the knowledge from a detailed Go game explanation”
very close to the output, and I think there are other smaller embeddings, like 19x19 grids, or background color light (I think this embedding is the cause of the failed generation when I switch to “dark background”), or other textual clues embeddings like shapes etc.
From the “creative” department, I think this generation from gemini 2.0 flash *prompted by the Chinese wiki page screen captures), is the one I think hallucinate the most, but also the most “modern artsy”
GTA game simulation online in browser:
I started with standard image, made a screenshot, then asked gpt to make it more about Go, then uploaded in GTA simulator
This is so exciting, so from “Account” department, here are parts of the brief:
Left Side - Top Block:
- 强 (qiáng) - Strong, powerful
- 联 (lián) - Connect, unite, link
- 接 (jiē) - Connect, receive, join
- 稳 (wěn) - Stable, steady
- 固 (gù) - Solid, firm, strong
- 提 (tí) - Lift, raise
- 劫 (jié) - Tribulation, calamity (in Go, refers to ‘ko’ threat)
- 胜 (shèng) - Win, victory
- 利 (lì) - Advantage, benefit, profit
This top block on the left seems to be related to Go strategy, discussing concepts like “strong connection,” “stability,” “lifting a ko,” and “victory/advantage.”
Amazing, also by Gemini speedy interpretation. The concept of * 子 (zǐ) - Child, son (in Go, it refers to ‘stone’ as in 黑子 - black stone, 白子 - white stone) explained a lot of our emotions involved within a Go match, right?:
Left Side - Middle Block (some larger characters):
- 黑 (hēi) - Black
- 白 (bái) - White
- 子 (zǐ) - Child, son (in Go, it refers to ‘stone’ as in 黑子 - black stone, 白子 - white stone)
- 形 (xíng) - Shape, form
- 势 (shì) - Power, momentum, situation (often ‘形势’ - xíngshì - overall situation/shape of the board)
- 地 (dì) - Ground, territory
- 空 (kōng) - Empty, space (often refers to ‘moyo’ or potential territory)
- 目 (mù) - Eye (in Go, a vital space)
- 死 (sǐ) - Dead
- 活 (huó) - Alive
- 打 (dǎ) - Hit, strike (often ‘打吃’ - dǎchī - atari)
- 吃 (chī) - Eat (capture)
- 点 (diǎn) - Point, spot (often ‘要点’ - yàodiǎn - vital point)
This block clearly lists fundamental Go terminology: black/white stones, shape, momentum/situation, territory, empty space, eyes, dead/alive, capture, and vital points.
And finally, I hope this is the right extract:
Left Side - Bottom Block:
- 腾 (téng) - Soar, gallop (could relate to rapid expansion or development)
- 挪 (nuó) - Move, shift
- 镇 (zhèn) - Suppress, calm, keep down (often in ‘镇压’ - zhènyā - suppress)
- 守 (shǒu) - Guard, defend
- 攻 (gōng) - Attack
- 防 (fáng) - Defend, guard against
- 断 (duàn) - Break, cut (often ‘切断’ - qiēduàn - to cut off)
- 连 (lián) - Connect
- 跨 (kuà) - Stride, straddle (could be a strategic move)
This section seems to describe various Go moves and tactical concepts like shifting, suppressing, defending, attacking, cutting, and connecting.
Other parts of the image: There are many other smaller, less clear characters scattered throughout the image, particularly in the center and right. Due to their distortion and the abstract design, it’s very difficult to make out specific meanings with confidence. They appear to be more fragmented pieces of text, possibly repeating or elaborating on the themes seen on the left.
In summary, the legible characters predominantly relate to Weiqi terminology, strategy, and tactics, reinforcing the theme of the image as an artistic representation inspired by the game of Go.
There is nothing legible in the generated image, not a single Chinese word, they are just blurry dots and lines. The capacity of bullshiting LLM is capable is impressive.
And here is the original prompt image (just a digest of the wiki page for Go)
I can only see the numbers and guess that 17x17 board is available. Could you help (guide Gemini) translate the text into English? Thanks a million.
It’s just this page, and cut the history part out. And it’s content is basically the same as the English page (although differs in some details as the Chinese page would use Chinese references and sources, and some historical references)
Full translation would take way too long for hours (and some would requiring translating the references as well for footnotes otherwise wouldn’t make sesne, like the Chinese legends of five emperors), and LLM already pretrain with them. The problem of LLM translation is never about not enough training date, but much deeper.