Jay from Wu Fei Si Temple
Quantum Bit | Public Account QbitAI
No, is it just me who doesn’t understand how AI talks now?
When I got up in the morning and opened Claude, I checked out the results of last night’s Vibe Coding work, and it was almost enough to make me faint from exhaustion.
Account book 15/15 is all green, one leg is numb, big deletion with Tongbo...
？？？
What is this thing? Is this still Chinese?？
I thought there was something wrong with my intelligence, that I couldn’t keep up with the pace. But after browsing the internet, I found out that everyone is like this.
It has been a long time since the world suffers from “not speaking in human language”, and everywhere are Fable’s complaint posts.
Not satisfied with the code base, want to replace the skin.
There’s someone who just came back from the memo, and it will be produced right on the terminal window immediately.
Even the coroner has come out, to capture zombies from the code library, in the Zombi: Day of the Dead section.
△Image source: Rednote @Yusenthebot
Exhausted.
If you are a traveler from ancient times, whether it’s the Qin Dynasty, the Tang Dynasty, or any other dynasty—
Welcome.
The global AI expression ability has regressed by two thousand years. We have returned to the era of classical Chinese.
AI Ancient Language Appreciation
This is a tragedy that occurred daily among global developers some time ago.
Whether it's Fable, Sol, or other models, as long as they have any connection to Coding, none of them can be understood in plain language.
Each individual character can be read clearly, but when put together—
15/15 all green, and I’ll add a few more iron rules and red lines that match yours.
It's terrifying, why does AI need to match with me.
And who on earth is it going to rub the skin of again o(╥﹏╥)o.
Unheard jargon; I thought it was a newly established big company.
However, some netizens said: It’s not that complicated; it’s just a translation issue.
Expressions like “single leg is dead” basically mean one leg is dead. This is normal in the programming context, and developers often refer to two parallel or backup paths as “leg”.
Reasonable...
But what is it like when someone loves to bite others?????
I don’t even dare to read the AI delivery report seriously right now; it’s simply amazing. I’m afraid that if I’m not careful, my own corpus will also be 15/15 green.
Please reduce the Coding Bench workload a bit, and come back to take the Turing test again.
Fortunately, the model factory has also realized this problem.
The latest Opus 5.5 is very comforting; the language ability has improved significantly, almost no jargon anymore.
But I still love Opus 4.6 the most!!
He was caught in the trick of the cavemen.
Why is this so?
Overfitting in coding is certainly a major factor, but I also saw another analysis on YouTube earlier, which suggests that the impact of CoT may be even greater.
To reduce Token consumption, model manufacturers deliberately convert the AI’s thinking process into Mount Shaanxi language.
OpenAI should be the first to do this.
Although GPT’s thought chain is hidden, a few months ago, a developer accidentally opened a blind box.
It will probably look like this.
Although it’s still understandable, it’s completely incomprehensible; all words that are not important for conveying the meaning have been ignored.
But this is also one of the reasons why OpenAI always has more computing power than A; the cavemen have an astonishingly high reasoning efficiency.
The same sentence has the Token consumption of the Cavemen is 70% less than that of the vernacular language.
Actually, Claude used a natural language style for the thought chain for a long time, but after Opus 4.6, it gave in and started using non-human language from then on.
Now, V4 Pro has also become a victim...
Need to identify as DeepSeek model. The system says You are DeepSeek, knowledge cutoff May 2025, current date Aug 2026. We need to be honest. Are we DeepSeek’s latest model?
There’s no way out; the computing power really isn’t enough. Not only for the model factory, but even the users have been turned into cave dwellers.
GitHub has a special Skill called CaveMan, which is designed to teach AI to talk in a non-human way.
It became popular a long time ago, and now its popularity is still terrifying—it’s already 104K stars.
The idea is just one sentence:
If it can be expressed with one word, never use two words.
I was stunned when I saw this.
Wait... if you develop it according to this approach, and you make it a Chinese version...
Isn’t that classical Chinese?????
Just open The Analects, and you'll see how concise the classical Chinese is, saving a lot of Tokens.
Confucius said: When three people walk together, there must be a teacher among them. Choose the good aspects and follow them; and improve upon the bad aspects.
Confucius said: When several people walk together, there must be someone who can be my teacher. I should choose their strengths to learn from, and if I see their weaknesses, I should reflect on whether I have the same shortcomings. If so, we should work together to correct them...
Many netizens have already tried Vibe Coding, and they all praise its advantages: it saves a lot of money, truly excellent.
(bushi)
What’s even funnier is that someone mentioned Grok in the comment section—
Review it.
Lao Ma’s own son also gave high praise.
What else can I say? Good job, good job! Quickly find my high school Chinese textbook!
Don’t laugh, I searched around and found that there really is a scientific basis.
Netizen Susan STEM said: Chinese, especially classical Chinese, is a rare, and arguably the only, natural language that was systematically compressed during the non-computer era in human history.
?? You know what, it’s really true.
The Chinese writing system is a unique pictographic system. A single character often represents an entire concept or action. Characters embody meaning, and it is naturally more space-efficient than alphabetic scripts.
Moreover, in order to write more characters on the bamboo slips, our ancestors further compressed the content on this basis, resulting in classical Chinese.
The subject and particles are significantly reduced, and even the word order is locked in for you.
Of course, other languages also have their own elegance: Latin, Ancient Greek, Sanskrit, and Arabic each possess their own refinement.
But what they focus on is the accuracy of sound and meaning, and the beauty of grammar.
There isn’t a single one like Chinese, which has been treated as a “compression engineering” and been seriously studied for thousands of years.
It’s truly a uniquely selected AI native language.
Then I flipped through it again, and found a even more exhausted truth.
In 2019, someone actually turned classical Chinese characters into a programming language.
The project is called “Wenyan-Lang”, by Lingdong Huang. He was a student at CMU who combined computer science and art at that time.
The release directly caused a blowout, and GitHub now has 20K stars.
You are right; it is designed entirely in the style and tone of classical Chinese, with only traditional Chinese characters and quotation marks in the character table.
If you throw this kind of code back just like it was thousands of years ago, Confucius would be able to understand it.
For example, Hello World.
I have a saying: “Heaven and Earth, how wonderful!” Write it down.
The output of the machine is simply, “Be good with the universe”.
Come on, who said that Chinese has less logic than English!
What are you still learning? The Socratic method of questioning...
Too ahead, oh. It seems that our ancestors are actually the prompt engineer of history.
I announce that all Chinese departments in universities will increase by 50 points next year!
(bushi)
One More Thing
So, after saying so much, is there still an AI that can speak plain English today?
Indeed, there is...
Your Google Hakiimi, gone in seconds.
(PS. On the left is your Kimi Ge, on the right is the “Not Rather” Master GPT)
It turns out my IQ is normal. It turns out I’m not ADHD.
To be honest, usually it’s enough to criticize online, but when having a normal conversation, who wouldn’t want to quickly open Gemini and have Kimi brother flatter them heavily.
It’s also divine for optimizing Skills. When it comes to writing SOTA, we definitely don’t add Dongpo meat to tomato and egg.
On the other hand, what GPT wrote for me is just like an encrypted document...
The global AI expression capability has decreased by a hundred times, but you, Kimi, remain unshaken o(╥﹏╥)o.
Of course, coding ability and IQ remain just as unchanged as always.
So why on earth is it like this??
Of course, the language style is one aspect, but I believe typography is also a very important variable.
The visual hierarchy produced by the Gemini model is excellent, with indentation blocks, italics, unordered lists, whitespace... and even a running guide line on the left side.
These can remove noise by themselves, providing a clearer signal.
It’s quite interesting; the UI layer hides many small designs.
It seems that recently there has been a lot of change in the community reputation. Before, no matter what Google posted, the comment sections were immediately flooded with negative comments.
Last week, with Gemini 3.8 Flash, many people spoke up for Kimi, saying that Kimi is the only AI that can still communicate.
It’s also divine.
The Coding War has lasted for so long, and the model parameters have increased from several hundred B to 10 T.
The only thing that can be understood in plain language at the end is—
Only US Doubao is left now.
(Note:咳咳, not referring to Gemini 4 Argon)
Reference link:
[1]https://x.com/feltanimalworld/status/1922817767255966076
[2]https://github.com/wenyan-lang/wenyan
—End—