Who else cannot be distilled into skill?

Document: Sleepy.md

Unfortunately, in this era, the more seriously you work without reservation, the more likely you are to distill yourself into a skill that can be replaced by AI.

In the past few days, the hot search lists and media channels have been flooded with "colleague.skill." As this matter continues to ferment on major social platforms, the public's focus has almost predictably been swept up by grand anxieties such as "AI layoffs," "capital exploitation," and "the digital immortality of workers."

These indeed cause anxiety, but what worries me the most is a line of usage advice written in the project README document:

"The quality of raw materials determines the quality of the skill: it is recommended to prioritize collecting long articles he actively writes > decision-making replies > daily messages."

The ones most easily and perfectly distilled by the system, and restored at the pixel level, are precisely those who work the hardest.

They are the ones who, after every project concludes, still sit down to write retrospective documents; those who, when encountering disagreements, are willing to spend half an hour typing long messages in the chat box, candidly analyzing their decision-making logic; those who are extremely responsible, meticulously entrusting all work details to the system.

Seriousness, once the most revered workplace virtue, has now become a catalyst that accelerates the transformation of workers into AI fuel.

Exhausted Workers

We need to re-understand a term: context.

In everyday language, context is the background for communication. But in AI, especially in the world of rapidly growing AI agents, context is the fuel that powers the engine, the blood that sustains the pulse, and the only anchor point that allows models to make precise judgments amid chaos.

An AI stripped of context, no matter how impressive its parameters, is merely an amnesiac search engine. It cannot recognize who you are, cannot grasp the undercurrents hidden beneath business logic, and cannot know what long pulls and trade-offs you have experienced on this network woven from resource constraints and interpersonal games when making a decision.

The reason "colleague.skill" has caused such a huge stir is precisely because it coldly and accurately locks onto the mine that hoards vast amounts of high-quality context—modern enterprise collaboration software.

In the past five years, the Chinese workplace has undergone a quiet yet excruciating digital transformation. Tools like Feishu, DingTalk, and Notion have become massive corporate knowledge bases.

For example, Feishu, ByteDance has publicly stated that the number of documents generated internally each day is enormous, and these densely packed characters faithfully encapsulate every brainstorming session, every heated meeting confrontation, and every strategic compromise swallowed down by over one hundred thousand employees.

This digital penetration far exceeds that of any previous era. Once upon a time, knowledge was warm; it lay dormant in the minds of veteran employees, scattered in casual chats in the tea room. Now, all human wisdom and experience have been forcibly drained of moisture, ruthlessly settling in the cold, unfeeling server arrays in the cloud.

In this system, if you do not write documents, your work cannot be seen, and new colleagues cannot collaborate with you. The efficient operation of modern enterprises is built on the daily cycle of every employee "offering" context to the system.

Serious workers, with diligence and goodwill, unreservedly expose their thought processes on these cold platforms. They do this to ensure the gears of the team mesh more smoothly, to strive to prove their value to the system, and to desperately find a place for themselves within this intricate commercial beast. They are not actively giving themselves up; they are just clumsily and diligently conforming to the survival rules of the modern workplace.

But it is precisely this context left for interpersonal collaboration that has become the perfect fuel for AI.

Feishu's management backend has a feature that allows super administrators to batch export members' documents and communication records. This means that the project retrospectives and decision-making logic you spent three years writing, enduring countless late nights, can be easily packaged into a cold, unfeeling compressed file in just a few minutes with a single API interface.

When Humans Are Dimensionality Reduced to APIs

With the explosive popularity of "colleague.skill," some extremely uncomfortable derivatives have begun to appear in the Issues section of GitHub and on various social platforms.

Some have created "ex.ex.skill," attempting to feed past WeChat chat records to AI, allowing it to continue arguing or being affectionate in that familiar tone; others have made "white moonlight.skill," reducing untouchable feelings to a cold interpersonal sandbox, repeatedly simulating probing phrases, cautiously seeking the optimal emotional solution; and some have created "dad-like boss.skill," preemptively chewing on oppressive PUA phrases in the digital space, building a sad psychological defense for themselves.

The usage scenarios of these skills have completely departed from the realm of work efficiency. Unbeknownst to us, we have become adept at wielding the cold logic of treating tools to dissect and objectify those fleshy, living beings.

German philosopher Martin Buber once proposed that the underlying colors of human relationships boil down to two entirely different modes: "I and you" and "I and it."

In the encounter of "I and you," we transcend prejudice and regard the other as a complete and dignified life form. This bond is open without reservation, filled with vibrant unpredictability, and, due to its sincerity, appears particularly fragile; however, once it falls into the shadow of "I and it," living humans are reduced to objects that can be disassembled, analyzed, and categorized. Under this extremely utilitarian gaze, the only thing we care about is, "What use is this thing to me?"

The emergence of products like "ex.ex.skill" marks the complete invasion of the tool rationality of "I and it" into the most intimate emotional domains.

In a real relationship, a person is three-dimensional, full of wrinkles, flowing with contradictions and rough edges; a person's reactions change constantly based on specific situations and emotional interactions. Your ex's reaction to the same sentence may be entirely different when waking up in the morning compared to after working late at night.

But when you distill a person into a skill, what you strip away is merely the part of them that happens to be "useful" or "effective" for you in that specific bond. The originally warm, self-aware person is completely drained of their soul in this cruel purification, becoming a "functional interface" that you can plug in and out at will.

It must be acknowledged that AI did not create this chilling coldness out of thin air. Before AI appeared, we had long been accustomed to labeling others, precisely measuring the "emotional value" and "network weight" of each relationship. For instance, we quantify people's conditions into tables in the matchmaking market; we categorize colleagues in the workplace as "those who can work" and "those who slack off." AI merely makes this implicit, functional extraction between people completely explicit.

Humans have been flattened, leaving only the aspect of "what use is it to me."

Electronic Patina

In 1958, Hungarian-British philosopher Michael Polanyi published "Personal Knowledge." In this book, he proposed a penetrating concept: tacit knowledge.

Polanyi famously stated, "We know more than we can tell."

He gave an example of learning to ride a bicycle. A skilled rider gliding through the wind can perfectly balance in every gravitational tilt, but they cannot accurately describe that moment's subtle intuition of the body using dry physics formulas or pale words to a beginner. They know how to ride, but they cannot say it. This knowledge that cannot be encoded or articulated is tacit knowledge.

The workplace is full of such tacit knowledge. A senior engineer may pinpoint a problem by glancing at the logs, but it is difficult for them to document this "intuition" built on thousands of trial and error; an excellent salesperson may suddenly fall silent at the negotiation table, and the pressure and timing of that silence are things no sales manual can capture; an experienced HR can detect the fluff in a resume just by noticing a candidate's half-second of eye avoidance during an interview.

What "colleague.skill" can extract is merely the explicit knowledge that has already been written down or spoken. It can capture your retrospective documents but cannot capture the struggles you faced while writing them; it can replicate your decision replies but cannot replicate the intuition you had when making those decisions.

What the system distills is always just a shadow of a person.

If the story ends here, it is merely another clumsy imitation of humanity by technology.

But when a person is distilled into a skill, this skill does not remain static. It will be used to reply to emails, write new documents, and make new decisions. In other words, these AI-generated shadows begin to produce new contexts.

And these AI-generated contexts will be deposited in Feishu and DingTalk, becoming training materials for the next round of distillation.

As early as 2023, research teams from Oxford University and Cambridge University jointly published a paper on "model collapse." The research indicated that when AI models are iteratively trained using data generated by other AIs, the distribution of the data becomes increasingly narrow. Rare, marginal, yet extremely real human traits are quickly erased. After just a few generations of training on synthetic data, models completely forget those long-tail, complex real human data, instead outputting extremely mediocre and homogenized content.

In 2024, "Nature" also published a research paper pointing out that training future generations of machine learning models with AI-generated datasets will severely pollute their outputs.

This is like those meme images circulating online, originally a high-definition screenshot, being forwarded, compressed, and forwarded again by countless people. Each transmission loses some pixels and adds some noise. In the end, the image becomes blurry, covered in electronic patina.

When the real, tacitly knowledgeable human context is drained, and the system can only train itself with the patina of shadows, what will be left?

Who Is Erasing Our Traces

What remains is only the correct nonsense.

When the river of knowledge dries up into an endless regurgitation and self-mastication of AI against AI, everything the system breathes in and out will inevitably become extremely standardized, extremely safe, yet hopelessly hollow. You will see countless perfectly structured weekly reports, countless emails that cannot be faulted, but there is no breath of living people inside, no truly valuable insights.

This great collapse of knowledge is not because human brains have become dull; the real tragedy lies in the fact that we have outsourced the right to think and the responsibility to leave context to our own shadows.

A few days after the explosive popularity of "colleague.skill," a project called "anti-distill" quietly appeared on GitHub.

The author of this project did not attempt to attack large models, nor did he write any grand declarations. He simply provided a small tool to help workers automatically generate seemingly reasonable but actually filled with logical noise ineffective long texts in Feishu or DingTalk.

His purpose is simple: to hide his core knowledge before being distilled by the system. Since the system likes to capture "actively written long texts," let’s feed it a bunch of nutritionally void gibberish.

This project did not explode like "colleague.skill"; it even seems a bit small and powerless. Using magic to defeat magic is essentially still spinning within the game rules set by capital and technology. It cannot change the growing trend of the system relying more on AI and increasingly ignoring real people.

But this does not prevent this project from becoming the most tragically poetic and profoundly metaphorical scene in the entire absurd drama.

We strive to leave traces in the system, writing detailed documents, providing meticulous decisions, trying to prove that we once existed in this massive modern corporate machine, proving that we are valuable. Yet we do not realize that these extremely serious traces will ultimately become the erasers that erase us.

But looking at it from another angle, this may not be a complete deadlock.

Because what that eraser erases is always just "the past you." A skill packaged into a document, no matter how clever its capture logic, is essentially just a still snapshot. It is locked in the moment it was exported, able only to spin endlessly within established processes and logic, relying on outdated nutrients. It does not confront the instinct of the unknown chaos, nor does it possess the ability to self-evolve in the face of real-world setbacks.

When we hand over those highly standardized, formulaic experiences, we also free up our hands. As long as we continue to reach out, constantly breaking and reconstructing our cognitive boundaries, that shadow lingering in the cloud will forever only be able to follow in our footsteps.

Humans are flowing algorithms.