Author: Liu Mo (People Magazine)

Translation: Deep Tide TechFlow

Deep Tide Introduction: This is one of the most in-depth internal reports on AI companies ever by People Magazine, where journalists were allowed to stay inside Moonshot AI for 100 hours, closely documenting this AI company with a valuation of over 120 billion RMB and only over 300 employees. From the collective shock after the DeepSeek impact, to the extreme flat management of "no departments, no KPIs, no ranks," and the "genius swarm" style of organizational evolution—this feature article unveils the true core of China's most关注 AI startup companies.

The spring of 2026 is particularly friendly to Kimi.

In just a few months, the company behind Kimi seems to have continuously broken milestones—revenue, financing, and valuation have all set new records. A research paper involving a 17-year-old high school intern received accolades from Silicon Valley, including Elon Musk. American programming tool Cursor, valued at about 50 billion USD, has been pointed out by Chinese observers as heavily relying on Kimi's model for product experience. In other words, Kimi seems to be winning simultaneously in capital, technology, and commercialization.

This company was founded only three years ago, and its valuation has already exceeded 120 billion RMB, approximately 16 billion USD. It can no longer be ignored in the global AI narrative.

But Moonshot AI remains deeply mysterious.

I was allowed to observe inside the company for 100 hours. As an independent contributor, I could interview any employees willing to talk, listen in on any meetings that did not involve business secrets, and no one would edit what I wrote afterward, nor was anyone paying me for my articles. This fits the style of this company perfectly.

Walking into the office feels like standing in the eye of a storm.

The central area is eerily quiet. There are only sporadic sounds of keyboards, and occasionally someone laughs. But the noise outside—rumors, debates, hype, imitation, and endless comments—seems to leave no trace here.

The company has only over 300 people, with an average age of less than 30. Calculated on a per-head basis, each person shouldered nearly 400 million RMB in corporate value.

About 80% of the employees are "I people" in the online context—inward, borrowing MBTI terminology. People sit together, but typing feels more comfortable than talking. Here, introversion is not a flaw; it’s almost a mode of operation.

I recall that night when I first visited in 2024, just when the storm was brewing. I did not leave a particularly good first impression.

"DeepSeek Saved Us"

On December 24, 2024, Christmas Eve. For most people in China, this is not much of a holiday. But for Julian, it became one of the darkest nights of her life.

She is 26, only graduated from Peking University two years ago, has no industry experience, yet she is already one of Kimi's earliest employees. That night, this very young but already considered "senior" employee sat at a long table in a meeting room called "Radiohead," facing over 30 colleagues, and cried.

She was unable to deliver a holiday marketing plan that satisfied the co-founders.

Only a month left until the Spring Festival. The latest plan has already been revised six times and now needs to be upgraded, possibly scrapped and started over. Rebuilding the plan from scratch and coordinating the product and engineering teams for execution, time is almost insufficient. However, the company holds high hopes for growth during the Spring Festival of 2025.

This is significant because last year's Spring Festival was the breakout point for Kimi. With the brand positioning of "2 million characters long text input," Kimi once went viral in China. The surge in end-user demand led to the emergence of "Kimi concept stocks" in the A-share market.

That weekly meeting was long and brutal.

About 20 young employees took turns reporting everything: social media advertising, user operations, domestic public relations, and overseas marketing, down to the smallest detail. Everyone discussed collectively, with the co-founders making final decisions.

At that time, Kimi resembled a pubescent teenager: talented, with potential, but not yet in full control of itself. Despite having an ad budget of tens of millions of RMB each month, it still appeared clumsy in the face of rapidly rising competitors.

The meeting ended around 4 a.m.

No one knew whether Julian's final proposal would succeed. A month later, that no longer mattered.

At that moment, the world first heard the name DeepSeek.

Hayley, responsible for growth, went back to Wenzhou for the New Year and found relatives and friends all asking the same question: "Have you heard of DeepSeek?" Kimi seemed to have turned into old news overnight.

She said it was the most challenging Spring Festival of her life. The silence inside the company was deafening.

The annual all-hands meeting usually takes place in March after the New Year, where employees can directly ask management questions. That year, almost every question revolved around DeepSeek.

The sharpest question came from the HR team. They spoke candidly, stating the uncomfortable truth:

"Candidates ask us: DeepSeek also gave me an offer, why should I come to Kimi?—How do we respond?"

But not everyone's response was the same.

Alex from the algorithm team said if he felt any intense emotion during the "DeepSeek moment," it was not fear, but excitement.

This sentiment was not just personal. It reflected the mindset of many people in the algorithm team. DeepSeek proved that another path might exist: lower-cost strategies, open-source pathways, and a fact many didn’t previously dare to believe—that an unknown Chinese startup could earn global respect as long as the technology was strong and the models were good.

The product team was not particularly anxious either. Kevin, one of the earliest product employees, believed that DeepSeek's impact relied on models. Once Kimi's own model capabilities caught up, the product team would actually have more room to develop valuable features.

No outsider knows what the co-founders discussed in detail. But the company's actions were swift. Adjusting strategies, narrowing focus, achieving close to complete internal consensus.

Now, if you ask almost anyone in the company what the most important thing is, they will unhesitatingly respond: models.

Since then, respect for DeepSeek within Kimi has grown stronger. Part of it is professional admiration, and part of it is something else.

Alex put it this way:

"In a sense, DeepSeek saved us."

Taste Is Everything

"How do you wear these shoes?"

After Ezra asked me, I was more surprised than she was. In her office section, almost everyone had a pair of slippers under their desks. Comfortable clothes and shoes are believed to make people more relaxed, focused, and creative.

That’s the dressing code of smart people.

I have seen many top students in my life. But the "good students" here are an entirely different species.

Ezra once tried to crack her family’s computer password in elementary school because her parents refused to tell her. In middle school, she started paying attention to Bitcoin, which then cost only a few hundred RMB. She asked her mother for pocket money to invest, but her mother said it was a scam. In high school, when she took her first taxi, she mentally sketched a prototype of a ride-hailing product. She said if there had been today’s AI tools back then, she might have really made it. Finally, when she went to university and had her own money, she invested in A-shares and lost 90%.

That painful experience made her realize the limitations of human judgment and pushed her toward AI.

Her understanding of AGI (Artificial General Intelligence) is straightforward: to create "N Einsteins" and use them to solve humanity's hardest problems. Since then, she was determined to find a company truly pushing the limits of AGI, even though she had already recouped her losses in the stock market by then.

Because of her outstanding academic background, she received offers from many companies. She chose Kimi for one reason only: during the interview, founder Yang Zhilin's understanding of technology and attention to detail deeply moved her. She felt he genuinely cared about models. He lacked the restlessness common among intelligent people and the utilitarian mindset prevalent among businessmen. In fact, by the end of the interview, she did not even know he was the founder.

Karen's personality is different, but she ended up in the same place.

He was rebellious from a young age, arguing with teachers and ignoring his parents. In school, he insisted on going abroad, and after graduation, he was adamant about starting his own business. The stable and comfortable life offered by a big company depressed him—he did not want to live a life with a beginning that was already predictable.

I asked him: if given the choice between definitely getting 60 points (out of 100) or a 1% chance of getting 100 points, which would you choose?

He chose the latter without hesitation.

It’s not that he couldn’t accept 60 points; he just couldn’t tolerate the 100% certainty route.

This "founder-style DNA" constitutes the underlying texture of the company. According to rough internal statistics, at least 50 people at Moonshot AI had previously founded companies or joined startups.

Some say Kimi likes to recruit CEOs.

More accurately, this company shelters a group of flowing genius drifters. Genius does not necessarily mean being a top student or exemplary employee. What matters is that in some dimension, they can see through time.

At a company where about 80% of employees come from 985 and 211 universities, Yannis's resume is not particularly outstanding. But as early as 2023, he predicted in the engineering community that both DeepSeek and Kimi would rise—at that time, model companies didn’t even have products. Another post-00s employee noticed his judgment and brought him into the company.

Karen said too many smart people are trapped by the system. First family, then school, and then the workplace. They unconsciously obey group expectations, forgetting what they truly want. Only a few attempt to escape, and they are often unnoticed.

One of Kimi’s missions, he said, is to see them.

Without this intuition, a 17-year-old high school student could not possibly be introduced as an intern at Kimi, collaborating with the team to publish papers that later received accolades from Elon Musk. The person who placed the student’s name in the first author position was Bob, his mentor, and the one who discovered him first.

There is only a thin line between genius and madness. When a "misunderstood madman" arrives at Moonshot AI, he may suddenly transform into a world-changing genius. Or perhaps some yet-unrevealed geniuses can only truly shine in such places.

Bob told me that to some extent, a big ego is not a problem; it might even be a good thing. If that ego serves as an inner drive, and if a person believes they must partake in a great mission, then they are likely the person the company cannot afford to miss.

Geniuses are paranoid.

In this team, training top AI models is jokingly referred to as "alchemy"—a common term in China's tech circle, describing the semi-scientific, semi-metaphysical process of model training. But in practice, alchemy means constantly fixing bugs.

Every time flagship training kicks off, Bob and his teammates enter the same set of rituals. The first thing every morning is to refresh the company’s massive internal monitoring dashboard. Hundreds of thousands of metrics. Even if a single curve exhibits abnormal fluctuations, alarms go off in their minds: Are there optimization issues? Is there a defect in the architecture? Are the numerical precisions mismatched?

Their reactions are almost animalistically sensitive.

Some even check the training data token by token, printing out tokens that produce extreme gradients, interrogating them like suspects: Why did you jump so wildly?

Every person who has truly participated in "delivering" a model has experienced this sleepless tension. It is not anxiety but an obsession driven by curiosity. It is this paranoid vigilance that pushes models to peak levels.

Geniuses gather.

In the past year, Kimi has had over 100 employees join through internal referrals—friends of friends, friends of friends of friends. The company internally refers to this as "person-to-person."

Trust, because of these dense networks, has become a natural organizational asset.

Essentially, Kimi has shifted the hardest part of management to recruitment. If people are referred in by trusted colleagues, they are more likely to share the same intuition. That’s why a particular word appears repeatedly within the company:

Taste.

On a September evening in 2025, several engineers casually launched an internal small project named Ensoul. They wanted to bring the dormant code in files "to life," turning it into a conversational assistant in the command line.

This sensitivity to naming is not accidental.

They once had a framework called YAMAHA, which stands for "Yet Another Moonshot Agent." The underlying infrastructure was called Kosong, meaning "empty" in Malay, inspired by the Buddhist concept of "form is emptiness." It hints at a blank slate, with no preset functions but infinite possibilities.

Taste, in other words, shapes the product itself.

When many companies stuff chat windows into command lines, Kimi’s engineers find it ugly. Real programmers open terminals to issue commands, not to chat. So Kimi CLI was designed to resemble a smart shell more than a chat interface. It understands commands but does not force itself to take the shape of a dialogue box.

This simplicity is also visible in the code. The core logic is only about 400 lines of Python, stripped of all unnecessary embellishments. Modules are cleanly decoupled. Users can customize functionalities or dismantle Kimi to reassemble into their applications.

Kimi Agent was internally associated with the phrase "OK Computer"—another Radiohead reference—but was later renamed because it was too obscure for broader users. Those who chose these names seemed not very interested in maximizing traffic. They follow their own musical taste and language standards.

Someone joked that if measured by the proportion of employees who can play instruments, Kimi might rank first among AI companies.

Taste has become the highest standard for recruitment and the most difficult to define.

It cannot be quantified, but it is omnipresent.

Generalize First, Then Evolve

You might never figure out exactly what everyone at Kimi is doing.

The company likes to use "team" rather than "department." From a high-level perspective, the main directions are clear enough: algorithms, products and engineering, growth, strategy, operations. But once you attempt to zoom in to see actual departmental divisions or fixed responsibilities, everything begins to blur.

Because this is an organization with no formal departments, no hierarchy, no titles, no OKRs, and no KPIs. Reporting relationships are so simple they feel fake.

For Brandon, this is completely unreasonable.

He graduated from Tsinghua University and has worked in management for Silicon Valley giants and Chinese industry leaders, helping a startup valued at about 1 billion dollars. After years in the industry, he is adept in technical management and has led teams of nearly 1000 people. He hopes to enter the AI field and showcase his talents.

But co-founder Zhang Yutao told him, this company does not operate that way. If he joins, he might have just two direct reports.

Yet something about the future drew him in, making him want to talk again.

So in January 2025, during a time when suspicion and unease permeated the company, Brandon met Yang Zhilin—his Tsinghua alumnus.

At that time, Brandon did not know that Yang Zhilin's name would later appear in the same reports as Elon Musk and Jensen Huang. What he remembers most clearly is Yang Zhilin’s first line after the usual small talk:

"Reinforcement learning is the future."

The subsequent dialogue felt almost like Yang Zhilin was speaking to himself. He was immersed in his thoughts, and Brandon could hardly understand what he was saying, even though it was all in Chinese.

But one thing was crystal clear: This was the first time in Brandon's life he felt the knowledge system and thinking model he had built over the past 20 years beginning to collapse. Along with it collapsed his arrogance.

I asked him why he eventually joined, and he answered rather esoterically: Yang Zhilin might become a great prophet because he has both vision and enough purity.

Later, when the company hesitated about how to define his role in this system that almost had no titles, Brandon firmly replied:

"Even if you put me to sweep the toilets, I will come, and I will sweep it cleaner than anyone else."

Not every former management from major corporations can survive in this environment.

Post-00 Phoebe transitioned from the growth team to the product and engineering team. She humorously labeled herself as "a little girl who knows nothing," but she said something important: in this company, profound experience and shiny resumes might become burdens.

AI is too new, and changes too rapid. An experienced expert may not learn and adapt faster than a younger person with fewer assumptions.

She witnessed at least three mid-level managers from large firms fail to integrate. One eventually chose to leave the industry, stating that the people around him were simply too young and too smart. After being repeatedly surpassed, he gave up. He concluded that this was no longer his era, nor his industry.

After the DeepSeek impact, Phoebe also felt a deep sense of crisis. She decided to quit her advertising job and instead assist the company through product and engineering. She began an intense self-study routine, even live-streaming her learning process on Bilibili, accumulating hundreds of hours.

What surprised her the most was that the company hesitated not at all to offer her a chance to shift roles from the very beginning.

In fact, among the 30 employees I interviewed, more than half had switched responsibilities multiple times. Compared to their previous work, about 80% of the people are now doing completely different things.

Kimi likes people with "generalization ability."

In AI, generalization means that models perform well in new scenarios outside of training data. It’s not just about rote memorization of answers but about learning underlying structures.

The company applies this concept to people as well.

Mid-level managers from large companies may have optimized within a certain KPI system, using a specific reporting language, or playing a certain internal political game for too long. Their "algorithms" became overfitted to a local optimum. When the environment changes completely, they may be unable to adapt.

If traditional large corporation employees are like specialized models, then Moonshot AI's desired people resemble foundational models. First, learn basic rules through supervised fine-tuning, and then gain cross-domain transferability through reinforcement learning and repeated self-play across tasks.

James, who returned from Silicon Valley at 26, said his dream is to "give money to young people."

As a devout believer in AI, he views his body as a sensor for the Agent to collect information. While playing "League of Legends," he records and collects physiological data such as heart rate and pulse, then analyzes which teammate's comments affected his emotional state and game performance.

His views are sharp to the point of extremism. He said: if a person starts learning a truly new language after the age of 14, they will never reach the native level. He believes AI follows a similar logic.

Dan, who joined the company right after graduating, said he felt genuine knowledge anxiety for the first time in his life.

In school, he only trained "toy models"—around 7 billion parameters, running on 32 GPUs for a few days. Now he handles MoE (Mixture of Experts) models with hundreds of billions of parameters, with training data measured in trillions of tokens. It feels like jumping directly from a small pond into the Pacific Ocean.

To keep up, he has almost entered a self-destructive learning state. His schedule has completely collapsed; Beijing's daytime has turned into Silicon Valley's night and vice versa. He has stared at training dashboards for hundreds of hours, like a stock trader watching the market, leaving no room for blinking.

The real challenge is not just the volume of work, but that he has to do three jobs simultaneously.

He must act as an algorithm architect, designing the optimal方案 in a maze of model choices. He must be a systems engineer, debugging distributed computing issues like repairing a pipeline spanning the globe. He has to act as a data alchemist, performing "alchemy" on massive datasets, ensuring the model scores well on benchmarks while being natural and smooth in actual conversations.

Sometimes this means performing emergency surgery during training. Once, a key parameter stored in bf16 precision began to show dangerous behavior. The team made a decisive switch to fp32 precision halfway through the training just to stabilize this round of training. Dan said if you only know how to write algorithms, or only know systems, or only clean data, you will never create a top model. There is no excuse here for "I only handle this part."

The company expects you to integrate the work of algorithms, engineering, and data while navigating across multiple worlds. It feels like holding multiple jobs simultaneously. But this high-intensity cross-training can bring years of growth in a very short time.

Thus, anyone wanting to join Kimi faces a brutal test.

No OKRs, no KPIs, no office politics, no PUA-style management, not even clocking in. But if you are not AI-native, if you cannot generalize, if you cannot keep reinforcing and adapting, you may find it hard to find meaning in your existence here.

"There’s No Bureaucratic Flavor Here"

Most brands want a story.

But almost every Kimi employee gently reminds me: don’t write about Pink Floyd, and don’t write about that piano at the office entrance.

Their view is that those who understand will naturally understand, and those who do not do not need to. The names Moonshot and Kimi have no direct relation to AI or technology. But if the company talks too much about its connection to rock music and art, it may begin to appear overly self-aware and affected. They seem to feel that beauty should require no explanation.

Win, a post-00 who "escaped" from a big company, told me that this place is strange because people can really finish tasks without meetings.

In his former workplace, meetings during the day followed by work at night were the norm. He learned a simple truth: if your energy is mainly spent on coordinating production relationships, there’s almost no room left to improve actual productivity.

This is part of what an AI-native organization looks like.

More than ten employees clearly told me they increasingly prefer dealing with AI rather than people. AI is more reliable and simpler. This tendency also fits the company’s overall introverted temperament. Some used a gentler word: shy.

In group chats, everyone can be lively and expressive. However, many become quiet when face-to-face. Kimi doesn’t organize cultural activities very often. Besides the annual meeting, the last collective activity was getting massages in the office.

Being introverted does not mean a lack of communication or vitality.

Even though no one is required to talk to me, not a single person refused. In group chats, information flies back and forth, mixed with various abstract emojis. No one’s message would be left hanging.

If you need someone else’s cooperation to complete a task, the process is simple: just go find them.

No need to go through supervisors, no need for approvals, no coordination meetings, no breaking down department walls.

Kimi has no department walls. In a sense, it doesn’t even have departments.

Yang Zhilin's signature only has four words:

Direct Communication.

Despite this, everyone admits the company has been constantly changing since its founding.

Some changes are proactive, some are reactive, and some even seem like reversals. The company has shifted from mass investment to focusing on models, from insisting on closed-source to embracing open-source, from chatbot products to Kimi Agent, Kimi Code, and Kimi Claw, from C-end to B-end and back to C-end. Not every pivot withstands perfect scrutiny.

But according to Ezra, one thing has remained unchanged: respect for facts.

She believes all those changes have only one reason and one goal: to align the company better with objective reality.

The company tolerates arrogance but does not like to recruit those who place themselves above facts.

From the co-founders downwards, people can be relatively easily persuaded—as long as the facts are clear enough. Employees say this willingness stems from an intense obsession with the truth, reality, and "what is real." Truly intelligent people will not be hurt by honest feedback.

This honesty comes with a condition: the company has no horse-race system, no zero-sum competition, no significant internal conflicts of interest. People are willing to share their research findings and technical details without expecting rewards or credits. Early on, the company had its own community and still advocates for community culture today. Sharing information and knowledge accelerates learning for everyone, ultimately benefiting all.

Win says toxic culture can be contagious, and good culture can also spread.

Some described the atmosphere here as "unity"—a term that sounds almost old-fashioned when applied to startups. But the environment the company operates in is harsh: outside are giant competitors, and internally there’s pressure from large companies, with limited computing resources. These constraints, if they serve any purpose, seem to enhance cohesion.

Ultimately, people are the only truly important asset in an organization.

Recently, Florence was poached by a competitor for double the salary. She immediately refused. The reason was simple:

"There’s no bureaucratic flavor here."

The company’s new office.

"I Don't Know How She Managed to Hold On"

Before the interview began, I was very nervous. I was about to interview some of the smartest AI practitioners in the world, and I come from a liberal arts background, having never worked in the tech industry, with limited knowledge of AI.

But when I actually started talking to the young experts from the algorithm and product engineering teams, I found that they were the ones feeling anxious. They were worried it would be awkward if I didn’t understand the terminology.

So they first translated English into Chinese, then took that Chinese and translated it into a simpler version that I could understand.

That protective instinct was touching.

Before the interview started, the company only gave me one directive: protect everyone.

So I tried to avoid asking sensitive or potentially hurtful questions.

Even so, Ty couldn’t completely hide a tremor of emotion during our phone interview. He was going through a difficult adjustment period when he first joined the company and once felt he couldn’t go on, even considering quitting.

Then at a weekly meeting, he saw Annie—a girl who had only graduated two years ago—finally push through a challenging project after experiencing numerous setbacks and self-doubt. Seeing this, he felt he could not give up. He was older, had more experience, but in pure endurance and willpower, he felt weaker.

He said:

"I don’t know how she managed to hold on."

In fact, Ty was not the only one who thought of leaving.

Annie thought about it too.

For a long time, she had been building the overseas business line from scratch but had never truly broken through. Worse still, colleagues from other teams, out of good intentions, would directly tell her to give up on what they considered a pointless effort.

She said she cried more times at Kimi than at any other company, and more than for any ex.

It wasn't that she didn't have a way out. She had already received a higher-paying offer. But she said she just couldn’t convince herself to work for someone else. She wanted to speak with Zhang Yutao one more time.

Afterward, she decided to stay.

She didn’t tell me what that conversation was about. She only said: Yutao is the strongest boss I’ve ever seen, the fastest in iteration, with the highest ceiling. Following him is my way to elevate my own limits.

Then Annie repeated the same phrase:

"I don’t know how she managed to hold on."

When you accumulate enough material, you start to notice certain phrases that keep reappearing. The most frequently repeated words often reveal the deepest common qualities of a team.

Bob, who was pulled back to China by Yang Zhilin, giving up his opportunity to pursue a doctorate in the United States, joined the company on its very first day. If anyone deeply understands this company, it is him.

When I asked him the question everyone asks—what is the team’s most important quality—he pondered for about two minutes before answering with one word:

Resilience.

For a company only three years old, talking about resilience might sound a bit extravagant. But he is serious. He said that intelligence and courage are sometimes opposing forces. The smarter a person is, the clearer the risks they see, making them easier to choose to leave. Blind perseverance will also not succeed. Only those who see the truth, calculate the failure probabilities but still continue, deserve to be called resilient.

A story circulates within the company called "Three Times at the Cliffs."

In May 2023, Freddie and colleagues received what seemed an impossible task: get the AI to read and understand 128K contexts at once—equivalent to hundreds of pages of text—while the industry standard was still around 4K.

He quickly designed a plan called MoBA v0.5, but this required rewriting the underlying training framework halfway through the main model training. The cost was too high, and the proposal was shelved. This was the first "cliff experience."

Half a year later, he returned with v1, this version designed to continue training from the existing model. It passed on small models, but when tested on large models, it encountered loss spikes and failed repeatedly. The project had to be returned for the second time—another six months. It even missed the milestone of the company’s 200,000-character product. But the team wasn’t disbanded; the company initiated a "saturation rescue"—mobilizing technical experts from various areas for concentrated effort. They rewrote the core logic, and v2 finally passed the classic long-text "finding a needle in a haystack" test.

Just when it seemed to be about to go live, the third setback arrived. In the supervised fine-tuning stage, the model performed poorly on long text summarization tasks due to sparse training signals. By this point, a significant amount of resources had already been invested. Engineers again returned to the "cliff," seeking solutions, ultimately resolving the issues by adjusting the attention mechanisms in the last few layers.

Three returns and three comebacks.

At the end of the interview, I asked Freddie the ultimate question: how would you describe this company?

He answered with two words:

Moon Landing.

Why moon landing?

He quoted Kennedy’s famous saying:

"We choose to go to the moon in this decade not because it is easy, but because it is hard."

All the conference rooms in the company are named after bands.

Genius Swarm

In the end, I did not disturb or attempt to delve into the co-founders themselves.

Outwardly, they are almost invisible. They dislike interviews and have no interest in personal fame. But internally, they are omnipresent.

In a highly flat organization, you need a superbrain at the center. Otherwise, energy can turn into chaos. With almost no mid-level management, each co-founder connects directly with about 40 to 50 employees, staying at the forefront of technology and business. This is how the company ensures alignment between decision-making and execution.

All five co-founders come from Tsinghua University. But physiological limits still exist. Human attention bandwidth is limited, and management radius is also limited. After the company’s valuation surpassed 120 billion and the number of employees exceeded 300, even these superbrains started to feel the pressure.

Not just the founders.

This is an infinite game driven by self-motivation. If everyone realistically carries a valuation of 400 million, then the expected value created by each person is also extraordinary.

The revolutionary variable is tools.

Kimi does not actually operate on extreme overtime. Employees can wake up naturally, not required to stay in the office until dawn every night. Leo from the product team says he now commands "an army"—the AI Agents.

Imagine this scenario:

Leo wakes up at 10 a.m. and walks into the office. His task is to analyze user feedback from five global markets over the past 24 hours and decide the product priorities for the week. In the past, this would require three people two days.

Now, he activates three Agents.

One strategy Agent scans 3000 feedback pieces, filtering out high-priority needs related to long text interruptions. One translation Agent interprets Japanese dialects and Korean honorifics in real-time, annotating true emotional intensity. One competitor Agent monitors updates from Cursor and ChatGPT, generating technical comparisons.

Leo himself only does three things: veto a comment wrongly identified by the system as a sincere sarcastic remark; mark a screenshot containing unread UI; confirm the top three needs recommended by the Agents.

By 11:30 a.m., the product requirement document has been completed. Meanwhile, a programming Agent has generated about 70% of the foundational implementation, leaving only the more creative design parts for discussion with human engineers in the afternoon.

Humans set the rules; silicon-based systems execute. The organization has become a container for algorithms.

In an AI-native company, skillfully using Agents and deeply embedding them into workflows is not a plus but a basic requirement.

Models are not just targets but also tools.

Whether directly improving productivity or fundamentally changing management structures, the logic of AI has entered the bones of this company. Just as the company builds Agent Swarms, the team itself has begun to resemble a Genius Swarm: many independent geniuses working in parallel, seamlessly collaborating.

However, this flat structure has inherent vulnerabilities.

When I asked if the model could sustain itself if the company expanded from 300 people to 3000, most responses were cautious. History is not very optimistic. Similar extreme flattening experiments—like holacracy or Haier’s model—often encounter decision bottlenecks when surpassed around 500 people. Too many information nodes, and "direct communication" can transform into information overload.

A closer pain point is individual disorientation.

Without hierarchical buffers for uncertainty, confusion about direction is directly transmitted to everyone. A former employee who eventually returned to a large corporation bluntly stated: without top-down OKRs and KPIs, some mornings when you walk into the office, you don’t know what to do. No one will necessarily tell you how well you are doing. This lack of feedback creates insecurity. It makes one nostalgic for the clear reporting lines, assessment nodes, and quantifiable outputs of large companies.

Those cumbersome structures, after all, provide something: a bottom line of certainty.

Where is the goal? What counts as completion? How is performance judged? In big companies, these are visible.

That person said it is not Stockholm syndrome; it is basic organizational physics.

If Alibaba is like a precisely calibrated promotion conveyor belt, ByteDance resembles a fiercely targeted combat army, and Tencent acts like a more lenient vocational college, then Moonshot AI is like a primeval forest.

Geniuses might find paths for hunting. Ordinary people may just wander in the fog.

The Necessary "Two-Dimensional Foil"

No departments, no titles, no evaluations.

This AI-native organizational model is anti-bureaucratic and deliberately unstructured. Large companies find it challenging to shift toward this model, while small companies often miss the window due to premature expansion into traditional structures. This is an asymmetric warfare.

Here, the author cites a classic concept from The Three-Body Problem. In that story, an advanced civilization casually uses a weapon called "two-dimensional foil" to collapse the solar system from three dimensions into two. Planets, stars, and humanity all become flat images without thickness.

The author believes that Moonshot AI is actively throwing itself such a "two-dimensional foil."

Not to eliminate competitors, but to flatten the organization, pursuing maximum efficiency.

No vertical depth of hierarchy, no horizontal walls between departments, no three-dimensional entanglement of office politics. Only "models" and "intelligence" face each other in the most straightforward form.

In the AI era, every startup is forced to throw such a piece of two-dimensional foil at itself. The rise of one-person companies reflects the generational explosion of the same AI-native talent. If technology can compress organizational capabilities onto individuals, then the intermediate management layer will evaporate in large numbers. Organizations are flattened. No convoluted depths. Everyone is forced to directly confront the problem itself.

This may be a hard rule of organizational evolution in the business world.

Everyone, in the end, will be folded.

Once people are exposed on the same plane, one person’s influence over fifty people will no longer be a managerial miracle but become the norm. The distance from center to edge is redefined. Those who rely on titles and OKRs as coordinates might suffocate in an instant. But geniuses on this exposed plane can fiercely dismantle intelligence itself; while the "guardians" clear noise and increase entropy, seeing themselves as vanguards broadening the boundaries of human civilization, not without humility.

However, the transition from three dimensions to two is irreversible.

This means that Kimi cannot take a backward step.

Every strategic adjustment becomes a high-risk chaotic iteration. Competitors can still turn slowly in the maze, but if Moonshot AI attempts reckless expansion, it may structurally tear itself apart. This self-dimension reduction is only acceptable as it serves a more radical goal.

The endpoint of reducing organizational dimensions is elevating the dimensions of intelligence.

Only when model intelligence crosses the critical threshold, high enough to escape the gravitational well of all carbon-based organizations, can Moonshot AI truly crush the organizational advantages of competitors and prove that this irreversible bet was the right one.

At that point, discussions about management radius or organizational structure will no longer matter. It's like asking which dimension a three-body civilization occupies—the real point is that its dimension-reducing weapon has rewritten the rules of war.

By then, "Moonshot AI" will no longer be a metaphor.

It will become a high-dimensional light source, illuminating the dark side of the intelligent universe. All previous organizational growing pains were merely the insulation burned off the lunar module as it passed through the atmosphere.

Either ascend to immortality,

or seal it in collapse.

There is no third way.

All English names mentioned in the text are pseudonyms.