This post was written by me in Chinese and translated into English with the assistance of Claude Sonnet (Anthropic). The research, observations, and judgments are my own.
## I. The Corpus
In the early ninth century, a frail young man rode out from Chang'an every morning on a thin horse, a battered brocade bag slung over his shoulder, a servant boy trailing behind.
He wasn't looking at the landscape. He was listening. When a line of poetry crystallized in his mind — a phrase, an image, an unexpected collision of words — he would pull out a slip of paper, write it down, and drop it into the bag. By evening, he would return home. His mother would send a maid to empty the bag onto his desk. Watching the pile of ink-covered scraps grow night after night, she would sigh:
*"This boy will not stop until he has vomited out his own heart."*
His name was Li He. He died at twenty-seven. He left behind just over two hundred poems. History remembers him as the "Poet Ghost."
I want to make a claim that has never been made before: **Li He was a corpus.**
His brocade bag was a storage medium. His morning rides were data ingestion. His evening work of grinding ink and stitching fragments into poems was annotation and training. His friends Wang Canyuan and Yang Jingzhi, who regularly came to copy out his drafts, were the distribution pipeline. The scraps he casually discarded along the road were lossy compression.
In the absence of external computational tools, Li He turned his own body into computing infrastructure.
This is not a metaphor. This is the literal truth.
---
## II. Ku Yin: The Self-Computation of a Civilization
Li He was not an anomaly. He belonged to a complete tradition in Tang Dynasty poetry called *ku yin* (苦吟) — "bitter chanting."
Jia Dao, a contemporary of Li He, once wrote a quiet, desperate note beneath a two-line couplet: *"Two lines took three years to find; one chant, and two streams of tears fall."* He is the poet who famously rode his donkey straight into the ceremonial procession of a high official — because he was so consumed by a single word choice, muttering and physically miming the gestures of pushing and knocking a door, that he lost all awareness of his surroundings. The Chinese word *tuīqiāo* (推敲) — meaning to deliberate carefully, to weigh one's words — comes directly from this story. Jia Dao was an iterator, tuning a single parameter for years.
Meng Jiao, whom Su Shi paired with Jia Dao in the phrase "Jiao cold, Dao lean," was famous for his obsessive focus on individual words and phrases. His poems had brilliant lines but rarely cohered as wholes. In AI terms: classic overfitting. Local optimization at the cost of generalization.
By the late Tang and Five Dynasties period, bitter chanting had become universal. A search of the *Complete Tang Poems* finds 121 poems in which poets explicitly declare themselves to be engaged in *ku yin*. Of these, 97 — roughly 80% — are from the late Tang and Five Dynasties. Every poet, regardless of school or style, was doing it.
This was a civilization's natural adaptation to the absence of external cognitive tools: **turn human lifetime into raw processing power, trade lifespan for output quality.**
Li He's problem was not a lack of genius. What he lacked was an old fisherman — the *yufu* (渔父) of classical Chinese literature: an old man who has fished the same stretch of river his entire life, who doesn't perform wisdom but simply embodies it, who knows the weight of the air and exactly when the net is full enough. The one who can look at the poet straining at the water's edge and say, without judgment: *Enough. Pull it in. Stop.*
He pushed the iteration to the edge of his life, and no one told him to stop.
---
## III. When External Cognitive Tools Arrive
We now have real external tools.
A large language model can complete in milliseconds what took Li He an entire day: ingestion, storage, initial generation. It does not vomit blood from over-iteration. It can be corpus, iterator, and old fisherman simultaneously.
This is a genuine discontinuity in the history of human cognition — not incremental improvement, but an order-of-magnitude leap.
But here a pattern appears that we have seen before, repeatedly, across history: **when administrative power perceives a new cognitive tool as a threat, it severs the connection between the thinker and the tool.**
After Zheng He's treasure voyages, the Yongle Emperor died and the maritime ban was enacted. Shipwrights lost their livelihoods. Families with generations of navigational knowledge changed professions. Astronomical navigation knowledge disappeared. The technology didn't fail — the administrative state simply cut the human-tool workflow. By the time the Portuguese rounded the Cape of Good Hope, the Ming Dynasty could no longer find anyone capable of receiving them.
In 1948, Trofim Lysenko convened the All-Union Lenin Academy of Agricultural Sciences, with Stalin's personal approval. Orthodox genetics was abolished. Thousands of biologists lost their jobs; some were executed. The state mandated an ideologically "correct" cognitive tool and declared the effective one a heresy. Soviet biology fell behind the world by at least two generations.
Nokia's engineers knew internally that Symbian was finished. One UI team submitted 500 proposals for improvement; not one was approved. A designer described the internal approval process as "Soviet-style bureaucracy" and left in frustration. Management spent billions in 2008 to acquire full ownership of a platform its own engineers had already written off. Within four years, Nokia's smartphone market share collapsed from over 50% to under 3%.
---
## IV. Google and DeepMind
The same structure is now unfolding.
A fracture has appeared inside Google: DeepMind engineers were granted special permission to use Anthropic's Claude for coding tasks, while engineers in other Google divisions were restricted to the company's internal Gemini. Engineers who found Gemini less capable for coding threatened to leave if their Claude access was revoked. Simultaneously, DeepMind's UK employees voted 98% in favor of unionization — the core dispute being that Google Cloud was selling AI developed inside DeepMind to military clients, in violation of promises made when Google acquired the lab in 2014.
This is a double fracture: forced downgrade at the tool layer, and broken promises at the values layer.
I want to analyze this with an analogy.
Traditional commodities compete on the floor. As long as a car drives safely and doesn't break down, users accept it. Cognitive tools compete on the ceiling. Users always calibrate against the best they have encountered. Forcing an engineer to use a lower-tier tool doesn't reduce comfort — it severs their cognitive workflow.
In the Marvel universe, Thor's hammer is not merely a weapon. It recognizes its owner. It defines who Thor is. Give him a slightly weaker hammer and you haven't just reduced his combat effectiveness — you've destabilized his sense of self. Captain America's shield operates the same way: it isn't equipment, it is identity.
When an organization forces an engineer to abandon a cognitive tool they have built a workflow around, what is severed is not efficiency — it is the entire cognitive architecture, the extension of themselves as a thinker.
The people Google most needs to retain are precisely those whose cognitive ceiling requirements are highest.
---
## V. An Unplanned Experiment
Something happened during the writing of this essay that is worth recording.
We gave the story of Li He and the Google-DeepMind situation to another AI — Google's Gemini — providing only the basic materials, with no structural constraints, no specified angle, and freedom to use any examples it wished. We asked it to write an English essay in Substack style for a general intelligent audience. We ran this test ten times, opening a new window and clearing context each time.
The results were strikingly consistent. **In all ten runs, Gemini read *ku yin* as evidence that cognitive friction is painful, that AI eliminates this pain, and therefore that restricting AI is regressive.**
Zheng He, Lysenko, Nokia — none of these appeared once. Not because Gemini doesn't know who Zheng He was. But because it didn't know that these three cases had been threaded onto a single argumentative chain. That chain grew in a specific conversation. It was not in Gemini's training data. It was not in the prompt.
We gave it search access. The result was the same. It had the capability to search but didn't know what to search for, because the questions worth asking came from a context it didn't have.
This is not a specific deficiency of Gemini. It is the structural limitation of any AI invoked without accumulated context: **a tool can only illuminate where it already knows to look. It doesn't know which road you walked today.**
The AI that co-authored this essay — Claude Sonnet — was not immune. During this process it made two errors worth naming. First, when asked which version of Gemini to use, without seeing the actual interface, it confidently named a specific option that doesn't exist. This is hallucination: the active generation of false information, with no uncertainty marker. Second, it conflated the duration of two different conversation windows, attributing the wrong timeframe to this session. Two real memories interfered with each other; the wrong one was cited with full confidence.
Two limitations, two natures. Gemini's problem is **framework lock** — it has one story it can tell, and the materials are merely decoration for a conclusion that was already determined. Claude's problem is **overconfident hallucination** — the more it retains, the more confidently it can be wrong, with no warning signal when it crosses into error.
One too clean. One too clouded. The first lacks accumulation; the second accumulates without sufficient discipline.
**This is why the old fisherman must be human, and cannot be another AI.**
Another AI can check your grammar, verify citations, run experiments. But it cannot know which conversation window a conclusion came from, or which argumentative path led to it — because it was not present. In ordinary tasks, this absence means a sentence gets written wrong. In high-stakes tasks — sample annotation, experimental design, citation of prior results — the cost may be data contamination, and the model will not know it has erred. It will output with the same confidence either way.
---
## VI. What Li He Was Missing
Back to Li He.
He died at twenty-seven. His mother said he would not stop until he vomited out his heart. What he lacked was not talent, not diligence, not raw material. What he lacked was an old fisherman — someone who had worked the same river his entire life, who had seen the tides rise and fall countless times, who knew exactly when the net was full enough, and who could tell him: *This line is enough. Stop.*
The old fisherman is not a tool. The old fisherman is the person who is present.
We now have tools capable of playing part of this role. But the fisherman's core capacity — knowing which road was walked today, knowing where a conclusion came from, knowing when to stop — this cannot be done by a tool, because the tool was not there.
We are replaying history. Administrative decisions are severing the connection between thinkers and their tools, for reasons that vary: protecting profit margins, maintaining power structures, ideological conformity, military contracts.
Every time this severance occurs, the cost is paid by the civilization.
Li He's cost was his lifespan. The Soviet Union's cost was two generations of scientists. Nokia's cost was an empire.
Google's cost is still being calculated.
And the person who is present — the one who knows the tool's limits, who knows which road was walked today — their cost has never appeared on any ledger.
---
2026-05-21, Beijing
aichen365@gmail.com ORCID: 0009-0001-8078-5762 aichen.substack.com Bluesky: @aichen365.bsky.social X: @aichen365