Part IThe Story
The teaser video for Kimi K3 cuts on the beat. It moves through fifty-six source clips, and each cut is motion-matched so that a gesture begun in one shot completes in the next. The edits are frame-accurate to the music. Someone selected the clips, processed the audio, and revised the sequence over several rounds. That someone was the model. In its launch materials, Moonshot AI said the system it released on July 16, 2026 "edited its own teaser video from 56 source clips, handling clip selection, motion-matched cuts, frame-accurate beat synchronization, audio processing, and multiple rounds of revision." The promotional film for the largest open-weights model yet built was assembled by the model itself.
The man who runs the company used to play the drums.
The next day was a Friday. The Nasdaq fell more than 1.4 percent in early trading. Nvidia dropped far enough that Apple, for a short while, passed it as the world's most valuable company. The PHLX Semiconductor Index slid into bear-market territory. MoneyWeek estimated that hundreds of billions of dollars had come off AI and chip valuations, which made it the largest shock to Western markets since DeepSeek. Moonshot paused new subscriptions because demand had overwhelmed its servers. Greg Brockman, OpenAI's president, gave the kind of grudging praise that only a direct competitor can: "It's a pretty good model. There's no question about it." David Sacks, until recently the White House's AI lead, said he had moved much of his own work from Claude to Kimi. A few months earlier, Anthropic's Dario Amodei had put Chinese models six to twelve months behind his own. That estimate now looked out of date.
The conversation turned almost at once from the model to its maker, and it began with a question. A post on X asking why a researcher of Yang Zhilin's caliber had left the United States reached more than seven million views by July 20. Vinod Khosla blamed immigration policy. The tech journalist Tae Kim called the alternative explanation "ludicrous" and wanted green cards stapled "to the top 10 AI computer science PhD degrees NOW." The entrepreneur Ankit Gupta went further, describing America's "moronic visa policy" and suggesting, apparently only half joking, that it "would be more logical to seize their passports and force them to stay."
The question rested on a premise nobody had checked, which was that Yang had been pushed out. He had not been pushed. He had been offered nearly everything an American career could offer, including, at one point, a job with Apple that he could have done from Beijing. He turned it down. The real puzzle of Yang Zhilin was never why America lost him. It was what he went home to get that America had no way of giving him, and what that cost him.
The name of his company is a clue, though only half of one. In Chinese it is Yuè Zhī Ànmiàn, the dark side of the moon. In English it is Moonshot. One name is about an object you can't see. The other is about going there.
By the Numbers
The Moonshot Ledger
2.8TParameters in Kimi K3, the largest open-weights model ever released (July 16, 2026)
$35BPost-money valuation after Moonshot's July 2026 round, oversubscribed threefold
~300Employees at Moonshot AI in 2026
200,000Chinese characters Kimi could process per conversation at its October 2023 launch
~$4.6MReported training cost of Kimi K2 Thinking (November 2025)
3Separate admissions Yang earned to Tsinghua University
4 yearsTime Yang took to finish his Carnegie Mellon PhD, two fewer than the U.S. norm
Three Doors Into the Same Building
Shantou sits on the coast of eastern Guangdong. It was one of the original Special Economic Zones, the places where the Chinese state first tried out the market, and by the time Yang Zhilin was born there in 1992 it had become a global center for plastic toys. His family, in the Chenghai district, is described in Chinese coverage as ordinary. The region, Chaoshan, has a particular reputation inside China. Its merchants are proverbially cash-flow people: rice traders, grocers, restaurant owners, property developers, families who built commercial networks across Southeast Asia one shop at a time. When a Chinese magazine went looking in August 2026 for the Chaoshan in Yang's company, it found almost none. His backers were not a hometown association. His founding team was not kin. The business did not throw off cash. It had "no sign of the cash flow that Chaoshan merchants cared most about." It was, the magazine wrote, "purely burning money to bet on the future."
The trader's instinct shows up in him anyway, in a different form. He bargained with the examination system as if it were a market where one sale was never enough.
At Shantou Jinshan Middle School, the city's best, he was put into the informatics olympiad training class without knowing how to program. Most of his classmates had started coding in junior school. A year later he won first prize in the Guangdong division of the National Olympiad in Informatics, which guaranteed him a place at Tsinghua University. Most students would have stopped there. Yang then sat Tsinghua's own independent entrance examination and passed it comfortably. Then, in 2011, he sat the gaokao, China's national college entrance exam, the one-shot ordeal that decides the lives of millions of teenagers a year, as an ordinary candidate. He scored 667, the top science score in Shantou. Tencent News later said he had "entered Tsinghua three times."
There are generous and less generous readings of this. The generous one is that he liked a challenge for its own sake, and that a mind which finds the problem more interesting than the prize keeps solving after the prize is won. The less generous one is that he did not trust a single result, so he checked his answer twice against independent tests. Both readings point to the same habit, and the habit never left him: getting more evidence than the decision strictly needs, then making the decision alone. "I made almost every important decision in my life myself," he told the Chinese outlet Sohu in 2020. Western profiles have tended to cast him as the product of a punishing system and demanding parents. The Spanish newspaper El País rejected that picture. Its version was a bright young man "given enough freedom to develop his independence and dream of becoming either a rock star or a wandering poet."
A rock star or a wandering poet. Both are vagrant ambitions, the kinds of life that don't come with an address. He was admitted to the most selective institution in China three separate times.
The Book He Cannot Remember
Tsinghua placed him in thermal engineering. Students joke that the major amounts to "boiler stoking." In his second year he transferred to computer science, and the stated reason has been quoted so often that it now reads like legend. He had been reading Haruki Murakami, and one of the novels had a gifted programmer among its characters.
He can't remember which novel it was. He can't remember the plot. What he kept, according to El País, was the programmer and "the desire it sparked to bring technology to life and change the world." Other accounts say the image that stayed with him was of programmers working late at night and turning code into real products. The book that redirected his life left behind one figure and one feeling. Everything else is gone.
This matters because the work that made Yang's name, and later the product that made his company's, was about memory. It was about how a machine could hold onto what came earlier in a long text, so that the beginning of a story still shapes the end. The man who would spend a decade fighting the fixed-length context window has a gap in his own past at exactly the point where his career turned. He recalls the effect and not the cause, the character and not the story. He has never seemed bothered by it. "Everything that happened afterwards," he said later, "convinced me that transferring to the Computer Science Department was the most important and correct decision I've ever made."
In computer science he did what he did everywhere else. Over 90 percent of his major courses came back above 95. He ranked first in his year. His adviser was Tang Jie, who headed the Knowledge Engineering Group in Tsinghua's computer science department. Tang would later become academic vice president of the Beijing Academy of Artificial Intelligence, lead its Wu Dao project, and co-found Zhipu AI, now called Z.ai, out of his own lab. More than a decade later the Wall Street Journal described the two men as "mere teacher and pupil" who had once shared "dreams of building a machine that could think like a human," and who were now both Chinese AI tycoons with companies valued in the tens of billions. Tang was the senior figure. He was also, by every account, Yang's loudest early advocate.
In 2014 Yang defended his candidacy for Tsinghua's Special Class Scholarship, the university's highest undergraduate honor. Tang introduced him to the committee as one of the most outstanding students he had met in years, and one person who later watched the footage noted how "passionate" the professor was. Yang won 13 of the 14 votes. His closing words had the earnest symmetry of a twenty-two-year-old who believes them completely: "This is the best of times for computer science, because it is the age of data. This is the best of times in my life, because I have a dream to do the best thing: to change the world with data."
The rock star had not gone anywhere. He had only been put on a schedule. At Tsinghua, Yang started a band with two friends, played drums, wrote songs, and named the group Splay, after the splay tree, a self-adjusting binary search tree that moves recently used elements to the root so they are quicker to reach next time. It is a data structure with a working memory. Music, El País reported, became the band's way of expressing "the pressures they faced and the absurdity they perceived in the world around them." Splay entered the university's original-song competition. It did not win. The prize it took home was best original track, and the regret Chinese profiles attribute to him is phrased with straight-faced precision: the band "didn't manage to win the original song competition, taking home only an award for best original track."
He graduated in 2015 having won almost everything, and the loss that people still bring up is the drummer's.
Beyond a Fixed-Length Context
Pittsburgh in 2015 was one of the places where modern machine learning was being worked out in earnest. Yang went to Carnegie Mellon's Language Technologies Institute to study under two advisers. Ruslan Salakhutdinov was a deep-learning researcher who would later direct AI research at Apple. William W. Cohen later became a principal scientist at Google DeepMind. The co-advising set the pattern: Yang would always have a foot in the scientific world and a foot in the engineering world, and the people who trained him were each partway out the door toward industry.
The pace of his output from those years is easy to underrate. His archive page reads like a ledger. February 2017: two papers accepted at ICLR. March 2017: two at ACL. May to August 2017: an internship at Facebook AI Research under Jason Weston. September 2017: two more at NIPS. January 2018: an oral at ICLR, which accepted 2 percent of submissions as orals. May 2018: an internship at Google Brain under Quoc V. Le. September 2018: HotpotQA, a multi-hop question-answering dataset that became a standard benchmark. Along the way came a 2017 Facebook ParlAI Research Award, a 2018 Nvidia Fellowship, a 2019 Siebel Scholarship, and papers co-written with Yoshua Bengio and Yann LeCun, both later Turing Award winners.
Two papers are the ones that count. In January 2019 he released an updated Transformer-XL, accepted at ACL that year. Its full title states the problem directly: Attentive Language Models Beyond a Fixed-Length Context. The Transformer, the architecture behind every large language model since, could only attend to a fixed window of tokens at a time. Transformer-XL found a way to carry information forward across segments, so a model could remember more of what it had read and produce longer, more coherent text. In June 2019 came XLNet, a new autoregressive pretraining method that beat Google's BERT on many language tasks. NeurIPS accepted it as an oral. It has been cited more than 10,000 times.
— Ruslan Salakhutdinov, Yang's doctoral adviser at Carnegie Mellon, to The Wire ChinaHis unique characteristic is that he was very strong scientifically but also very strong in terms of coding and implementation.
Salakhutdinov called him an "absolutely brilliant student." Seven years after graduation, in the week K3 came out, he posted on X with the tone of a proud and slightly dazed parent: "It feels like just yesterday Zhilin was graduating from my lab at CMU. Not only did he complete his PhD in just four years, but he also made truly fundamental contributions to machine learning during his time at Carnegie Mellon University." Four years, when the American norm is six. The thesis was titled Advances in Generative Feature Learning.
There is a third thread from these years that looks like a side project and turned out to be the whole story. In 2016, still a doctoral student, Yang co-founded Recurrent AI, which built AI tools for analyzing sales conversations. The company later said it spent 2016 and 2017 working out a direction and only began real operations in 2018. In 2021 it raised $38 million. It ran on Transformer-XL. Yang, then, was publishing the research and selling it at the same time, applying his best work on long context to the most ordinary corporate need there is: a record of what a salesperson said to a customer and what came of it. One Chinese founder-profile site put it plainly: Recurrent AI taught him the actual constraints of enterprise AI, which go well beyond model quality to data, delivery, sales cycles, and customer ROI. He was a co-founder and a technical core of that company, "not a number one who held all power." That last part would matter later, in arbitration.
The Job and the Seat
In 2019 the market valued Yang the way markets value rare assets. Postdoctoral positions at Stanford and MIT were open to him. Apple tried to recruit him, and Salakhutdinov later described the exchange with a specificity that suggests he still found it a little remarkable. He had been on an email thread with an Apple executive in a reporting line to Tim Cook. The executive asked whether Zhilin would consider joining. Salakhutdinov said his student wanted to go home. The executive said that was fine, since Apple had an office in Beijing.
Yang said no to that too.
Salakhutdinov's recollection, translated from his posts and retold across Chinese media in July 2026, is the closest thing we have to Yang's reasoning: if he didn't try at least once to start his own company, he would regret it for the rest of his life. The Wire China recorded the line in Yang's own fragmentary words: "If I don't do a start up, I will always regret." The Wall Street Journal added something the adviser's later warmth had covered over. At the time, Salakhutdinov thought it was a bad idea.
A Chinese blogger, Jason Zhang, worked through the debate after the K3 frenzy and offered a framing that cuts through it. America could give Yang a very good job, good enough to do in Beijing for the most valuable company in the world. China could give him a founder's seat. The choice wasn't between Silicon Valley and Zhongguancun. It was between two identities. The viral summary, which Zhang was careful to say was not a quote from Yang, ran: America gives you a good job; China gives you the chance to be a founder.
That framing is cleaner than what actually happened, which looks less like a launch and more like a long approach. Yang went back in 2019 and kept moving between institutions for nearly four years. In 2020 he worked with Huawei on an early version of its PanGu model. In 2021 he led a team on Wu Dao at the Beijing Academy of Artificial Intelligence, the large-model project his old adviser Tang Jie oversaw. From July 2021 to June 2024 he was a principal investigator at the Shanghai Qi Zhi Institute. He took an assistant professorship at Tsinghua's Institute for Interdisciplinary Information Sciences, and a speakers' bureau listing from that period has him recruiting doctoral students from the IIIS summer camp each June and July. In 2022 his lab released a multilingual code generation model, CodeGeeX, and published papers on prompt tuning and few-shot learning, some with Tang Jie as co-author.
His personal website, hosted at a GitHub address under the handle kimiyoung, appears to have stopped being updated in 2022. Its last status line says: "I am working on a start-up." His GitHub bio offers something closer to a creed: "The ultimate goal of all my work, including both research and business, is to maximize the value of artificial intelligence."
Researcher, professor, co-founder, national-project lead. He held all of these at once without fully settling into any of them. The pieces were on the board. What brought them together was not his own plan but a product release in San Francisco.
A Hundred Million Dollars, Calculated in Silicon Valley
OpenAI released ChatGPT on November 30, 2022. A month later Yang flew to the United States to see what was going on. He has said the release is what led him into generative AI, and in a 2025 interview he described the effect in almost statistical terms: "AI has developed over 70 years, but never before has there been such a breakthrough. Previously, AI companies focused on B2B. But since last year, applications like GPT have emerged, reaching non-technical users. Hitting 100 million users was a clear signal."
What he did in Silicon Valley is described in the opening of Zhang Xiaojun's first profile of him, and it is the most useful paragraph anyone has written about him. "Just one year ago," she wrote in March 2024, "AI scientist Yang Zhilin did a precise calculation in Silicon Valley. He realized that if he decided to launch a foundation-model startup aimed at AGI, he would need to raise more than $100 million within the next few months. Yet that was merely a ticket to the game. A year later, that figure had multiplied thirteenfold."
A precise calculation. Not a vision or a calling. An entry fee, worked out to an order of magnitude before he had written a pitch deck. Zhang is a former print journalist and the host of one of China's most influential long-form business podcasts. She has interviewed Yang at length twice, a year apart, and her framing of his first year has stuck: the competition among model companies, she wrote, "is less a scientific contest than, first and foremost, a brutal contest of money." Yang put it in his own words: "It needs the concentration of talent and the concentration of capital."
He brought talent first, and he brought it from the band. His co-founders, Zhou Xinyu and Wu Yuxin, had been Tsinghua classmates and, by Wikipedia's account, members of Splay. Their résumés fit together the way a rhythm section does. Zhou had worked at Hulu, Tencent, and Megvii on running deep neural networks on hardware with little computing power, which is the craft of making a lot of model fit into a small space. Wu had worked at Google Brain on foundation models and at Meta AI Research on computer vision. The company would later list Group Normalization and ShuffleNet among the techniques its founding team had helped author. Between them, the three had the long-context theorist, the efficiency specialist, and the vision researcher. They were the three capabilities Moonshot would spend three years combining.
The money came quickly. "It was alright, the process didn't take much time," Yang said later. He also said he couldn't tell many people what he was doing. HongShan, the firm formerly known as Sequoia China, and ZhenFund came in at a $300 million valuation. Wang Xinyu of Meituan's Longhill Capital heard Yang explain his view of where large models were going and came away thinking, as he later put it, "There is no reason not to invest." At that point the company had a little over twenty people, no formal structure, and no model.
It incorporated on March 1, 2023. That was fifty years to the day after The Dark Side of the Moon came out, the Pink Floyd album Yang has named as his favorite of all time. The Telegraph would later point out, with some amusement, that it was released almost twenty years before he was born, and that he also loves Led Zeppelin. Cao Xi, founder of Monolith Management and a former Sequoia China partner, named his own fund after the black slab in 2001: A Space Odyssey and invested in three consecutive Moonshot rounds. In his fund's podcast he described what the album sounds like: "panting footsteps, airport boarding announcements, cash register drawers opening and closing, and throbbing bass lines," a record about "conflict, greed, time, and death."
Cash register drawers opening and closing. The company was named after a record that sets money to music.
— Yang Zhilin, in conversation with GeekPark founder Jack Zhang, November 2023When you look at the moon, you only ever see the lit side. You never see the back, and that gives you a strong urge to explore it. Large models feel similar.
Yang explained the name more than once, and his explanations shift a little with the audience. To GeekPark in November 2023 he talked about rock's spirit of challenging "the existing shape of things." To a Western interviewer in 2025: "you usually only see the illuminated side of the moon, but the dark side remains mysterious. It's challenging, yet full of potential." To a Chinese technology journalist around 2024, quoted later by the Telegraph: "We may be the earliest to explore the dark side of the Moon." The company's own website records the founding in a single entry under the heading Moonrise: "Driven by the dream of AGI, we embark on a journey to the far side of the moon."
The first chatbot came out that October. Moonshot gave it Yang's English name. It was called Kimi.
The White Piano and the Fourth Founder
When Zhang Xiaojun visited in early 2024, Moonshot had eighty people working in its first office, which she described as somewhat run-down. It was not in Beijing's Sohu Internet Building, where model companies usually clustered. "There was no logo at the entrance," she wrote. "Only a white piano standing guard by the door."
The company's most valuable possession was a product named after its founder, and the company itself had no sign on the door. Visitors met an instrument first. Two years later, after the money, the fame, and the move into the JD Technology Building in Haidian, the Financial Times found the piano had come along. The paper described the "mythology" Yang had built as something that "helped distinguish Moonshot from China's otherwise austere AI industry." It described an intense culture of long hours, and then the other thing: "Yang has also infused the company with his own... obsession with rock music... A white piano stands prominently in the office."
The piano was the side of the company everyone could see. Behind it was a figure most people couldn't.
In late February 2024 Moonshot raised roughly $1 billion in a round led by Alibaba, which took about a 36 percent stake. The pre-money valuation was $1.5 billion and the post-money about $2.5 billion, which made it, per TechCrunch, the largest single funding round on record for a Chinese large-language-model developer. Investors who tried to get into Kimi's rounds in that period told the Chinese outlet Waves something odd: their first meeting was often not with Yang. It was with Zhang Yutong.
Zhang had been a managing partner at GSR Ventures, a rare woman with what Waves called "genuine tech investment chops." She had backed Xiaohongshu, the chip company DeePhi Tech, and Recurrent AI, Yang's first company. By early 2024 she was so deeply involved in Kimi's operations that people described her as "like the company's chief steward." Moonshot's official position for a long time was that she had "not yet joined." Inside the company she used an alias, Aether, the name of the protagonist in miHoYo's game Genshin Impact, a traveler "forced onto a journey of wandering and searching." A source close to the Alibaba deal said negotiations "continued through the New Year holiday, forcing Alibaba's strategic investment team to work overtime." Other reports said she had arranged the alliance from Singapore.
So the company called the dark side of the moon had a founder its own investors met before they met the CEO, and that founder went by the name of a wanderer. You don't need to invent anything to see the irony. It was already there.
The bill arrived in the second half of 2024. Recurrent AI's investors, five of them including GSR Ventures, objected to how Yang had left his first company to start a second, and to how much of the second's value had moved away from the first. On November 11, Waves reported that the dispute had gone to arbitration. The South China Morning Post confirmed in December that the case was before Hong Kong arbitrators. Zhu Xiaohu, managing partner at GSR, publicly and repeatedly attacked Zhang, his firm's former managing partner. The arbitration was under seal. Waves talked to about a dozen people near the dispute and came to a verdict about the period rather than the people: "It is a story of human nature being instantly magnified in a rare arena of reckless acceleration."
Yang's answer, posted on social media one night in early December 2024, read like something a lawyer had drafted and a scientist had approved. When he left Recurrent AI, he wrote, every necessary procedure "had received the signed consent of every board member." His decisions were independent of those of his former company and of certain shareholders. Then, for the first time in public, he addressed Zhang. He had invited her to join Moonshot as a co-founder. Her shares depended on years of service and performance, on the condition that she kept working at the company and delivered results. The grant, he said, had nothing to do with his departure from Recurrent AI and nothing to do with some shareholders' decision not to invest in Moonshot.
A settlement was eventually reached through the Hong Kong International Arbitration Centre, according to the Telegraph, which drew a moral that was hard to dispute: "Recurrent AI taught him how to woo investors and scale a business. But he learned painful lessons too." He had once been "not a number one who held all power." Now he was the number one, and the past had sent its invoice.
Traffic Spending Maniac
Kimi's first pitch was a number, and the number was a claim about memory. At launch it could handle 200,000 Chinese characters in one conversation. TechCrunch, passing along the company's claim, said that was roughly eight times what GPT-4-32K could do. Yang explained it as plainly as he had explained the founding arithmetic. In the GeekPark conversation that November, his first long public interview, he described the Transformer as a new kind of computer. Parameter count was its CPU. Context length was its memory. "Forty years ago people thought 500K of memory was plenty. The same fallacy is being repeated about context." The goal at the end of that road was intimate: a lifelong AI companion, because trust and complex emotion only show their strength over decades, and "an AI that has to reset its context window every day cannot do that."
Kimi was Transformer-XL turned into a product, the doctoral student's problem sold back to the public. In March 2024 Moonshot said the limit had risen to two million characters. On March 21 the traffic took Kimi down for two days, and the company apologized. By May there were six paid tiers, from 5.2 yuan for four days to 399 yuan for a year of priority use. By August 2024, Kimi ranked third in China by monthly active users. That month Tencent and Gaorong Capital joined a $300 million round that valued Moonshot at $3.3 billion, making it the highest-valued of China's new AI "tigers." Cao Xi, the Monolith founder, told the South China Morning Post that Yang was "an entrepreneur with potential," a description that sounds modest only if you haven't sat across a table from a venture capitalist.
The strategy in this period had three clear labels, which GeekPark later listed when it republished that first interview: long context, proprietary and closed-source, consumer-first. Yang argued for the second label with full confidence. The super-app entry point, he said in 2023, would most likely be closed-source, because owning how the model evolved gave the product a difference from the first day. Moonshot built "one and only one consumer product," Zhang Xiaojun noted, while its rivals took safer business-to-business routes or verticals like healthcare and gaming. Yang described the company he wanted as a fusion of "the technology idealism of OpenAI with the business philosophy of ByteDance." In the GeekPark interview he called it Silicon Valley's technical idealism as the drive and the Chinese focus on usefulness and business models as the fuel.
The fuel was expensive. Kimi spent heavily on user acquisition, and by the end of 2024 the Chinese press had a name for Moonshot: traffic spending maniac.
The problem with buying users is that you never fully own them. On January 20, 2025, Moonshot released Kimi K1.5 and claimed it matched OpenAI's o1 in math, coding, and multimodal reasoning. The technical report said the gains came from long-context scaling and better policy optimization, without Monte Carlo tree search, value functions, or process reward models. It was a serious result. That same month DeepSeek released R1, and the story of Chinese AI moved somewhere else overnight.
The man it moved to grew up in the same province. Liang Wenfeng was born in Wuchuan, in Zhanjiang, Guangdong, seven years before Yang. He spent fifteen years in quantitative trading and funded his AI research from trading profits and compute he had built himself, which is the cash-first logic Chaoshan merchants would recognize and Moonshot had skipped. The 36Kr outlet AlphaRank wrote that Liang "replaced Yang Zhilin as the new protagonist of the 'technical genius' narrative overnight." What Yang faced, the same writers said, was "the collapse of the commercialization path, the loss of public opinion momentum, and the stagnation of financing." By June 2025 Kimi had fallen to seventh in monthly active users. In 2026 a Chinese headline would compare the two Guangdong founders this way: Liang Wenfeng has no personal life, Yang Zhilin has no way back.
The phrase "no way back" is accurate, and it doesn't mean what the headline seems to intend. Yang had left Apple's Beijing office and the Stanford postdoc behind years earlier. What he had left in 2025 was a direction.
Problems Are Soluble
In September 2024, a few months before the worst of it, Yang gave a talk at Tianjin University's Xuanyuan College. He ended by quoting Daniel Kahneman, author of Thinking, Fast and Slow: "Many times, you have the courage to attempt something unknown because you don't realize how much you don't know. Once you try, you discover many new problems, and perhaps that's the essence of innovation." In the same talk he summed up the old paradigm of AI in a farmer's phrase: "You reap what you sow. If you want a watermelon, you plant a watermelon." General models had broken that rule. You trained one thing and got many.
What Yang did in the first half of 2025 also broke his own earlier rules, and he did it quickly. According to AlphaRank, he made two strategic choices back to back. The first was to "completely abandon the C-end traffic acquisition strategy that had made him famous, and go all in on foundational model R&D." The second was to take the agent ground after DeepSeek had claimed deep reasoning, which made Moonshot "one of the domestic players closest to Anthropic's technical paradigm." Later Chinese commentary added that DeepSeek's success had given Moonshot cover: R1 had shown that open source and algorithmic efficiency could work in China, which made it easier to justify stopping the spending on users.
The founder who had argued in 2023 that the winning entry point would be closed-source now gave his models away.
The technical groundwork had been laid in quieter places. In February 2025 Moonshot and UCLA published Muon is Scalable for LLM Training. It showed that Muon, an optimizer known to work well on small models, could train a 16-billion-parameter mixture-of-experts model at roughly twice the computational efficiency of AdamW, the standard. Yang later explained the point in a way any merchant would understand: a model that "eats one portion of data and learns two portions of knowledge." In March 2025, Moonshot's paper on Mooncake, the serving platform that handled 100 billion tokens a day for Kimi, won the Erik Riedel Best Paper Award at the USENIX FAST conference. The South China Morning Post headlined it as an "AI booster that may help counter US chip ban." Moonshot did not have the most chips. It worked hard on getting more out of the chips it had.
In July 2025 the company released the weights of Kimi K2: one trillion total parameters, 32 billion active at inference, a mixture-of-experts design trained on 15.5 trillion tokens, under a modified MIT license. The day after release it was the most downloaded model on Hugging Face. Nature ran a news piece headlined "'Another DeepSeek moment.'" Moonshot's own engineers told The Wire China that K2 had drawn heavily on DeepSeek's architecture, an admission about lineage that the Chinese open-source ecosystem treats as normal and that Washington would later read as suspicious. Andrew Carr, a former OpenAI researcher, put his finger on something benchmarks couldn't capture: "K2 is noticeably different in how it communicates. It is fresh, interesting, and most importantly not sycophantic. It will reason with you about a problem and push back in ways other models don't."
A model that pushes back. Its maker was a man who made almost every important decision himself.
In August 2025 Yang sat down with Zhang Xiaojun for their second long conversation, about a year after the first. His first reaction, he said, was that it felt like a long time had passed. "One day in AI is like one year in the human world. I am not sure how many days in the human world equal one year in AI, though." The snow mountain from the earlier interview was still there. In 2024 he had described the work as driving toward a range of peaks "you don't know what's inside," moving forward "step by step." Now he had a book for it. He cited David Deutsch's The Beginning of Infinity repeatedly, and especially its core claim, which he took as a working philosophy: problems are inevitable, but problems are soluble. Every solution opens a harder problem, so the mountain has no top. That is a hopeful idea and also a tiring one. The episode was titled Standing at the Beginning of Infinity.
He gave Zhang an image for the shift he was betting on. A pure reasoning model, he said, is a brain in a vat. "It does not need to interact with the outside world. It is like a fish tank. You put a brain in it. It has no connection with the outside world." An agent breaks the glass. It searches, writes code, reads the output, and adjusts. In the same conversation he made an observation about a competitor that was blunt and generous at once: "Claude's reasoning performance is not very high, but its performance as an agent is very high." Anthropic had bet on acting over thinking. Yang thought the bet was right and said so, then set out to make a better version of it.
In November 2025 came K2 Thinking. It had the same trillion-parameter architecture, a 256,000-token context, native INT4 quantization, and the ability to make 200 to 300 sequential tool calls without a human in the loop. Moonshot reported scores above GPT-5 and Claude Sonnet 4.5 on Humanity's Last Exam (44.9 percent), BrowseComp (60.2 percent), and SWE-Bench Verified (71.3 percent). The reported training cost was about $4.6 million, less than the price of a mid-sized office floor in Haidian. The license now required attribution from any product with more than 100 million monthly users or $20 million in monthly revenue. By the end of the year Moonshot had closed a $500 million Series C and held more than 10 billion yuan in cash. The traffic spending maniac had become a lab that published papers and kept a large cash balance.
What the Lit Side Shows
In January 2026 Moonshot released K2.5. It added native vision through a 400-million-parameter encoder called MoonViT, introduced an Agent Swarm feature that let the model coordinate up to 100 sub-agents, and launched Kimi Code, a command-line coding agent. In March, Yang gave a keynote at Nvidia's GTC conference in San Jose. It ran about forty minutes, in fluent but heavily accented English. The Telegraph noted, approvingly or warily depending on how you read it, that after one broad statement of purpose he "sticks safely to the technical realm." The statement was this: "We believe in democratising intelligence, with open models you can deploy anywhere … to make intelligence more accessible to anybody in the world." The rest of the talk was about scaling gains along three axes at once, token efficiency, long context, and agent swarms, and multiplying them.
That same month Cursor shipped its Composer 2 coding model. Business Insider later reported, under the headline "Cursor acknowledges its new low-cost coding model has Chinese bones," that it was built on Kimi. Thinking Machines Lab's first model, Inkling, released in July, was based on K2.5. Cognition said in September that Devin had also been built on Kimi. The most interesting American AI companies of the year were, in part, quietly running on Moonshot's weights.
Then the lit side of the moon started casting shadows.
In February 2026, Anthropic accused Moonshot, DeepSeek, and MiniMax of a "distillation attack": using thousands of fraudulent accounts to generate millions of conversations with Claude in order to train their own models. In April, the House Committee on Homeland Security began pressing DoorDash and Anysphere, Cursor's maker, about their use of Kimi. On September 10, Anthropic published a second, more detailed account. It said Moonshot had used 5,380 accounts through proxy servers in Singapore and Japan, had made more than 23 million exchanges, mostly with Claude Opus, and had extracted Claude's chain of thought, which is meant to reach users only as a summarized "thinking signature." The Cyberspace Administration of China opened an investigation into the claims against Moonshot and DeepSeek. Observers read the probe as a cooperative gesture ahead of Xi Jinping's state visit to the United States later that month.
At the end of July, Bloomberg reported that Alibaba Cloud had given Moonshot access to 20,000 Nvidia H200 GPUs, and that Moonshot could reach Nvidia's newest Blackwell chips through Southeast Asia. Alibaba denied supplying the H200s.
None of this has been decided, and this profile doesn't decide it. What the record does show is a pattern that follows Yang through his career: what the public sees of his company is always only part of it. A sign-less office with a piano at the door. A fourth founder who went by a game character's name. Compute that may or may not have come through a third country. An open model whose openness includes, as of K3, a license requiring any host making more than $20 million a year from serving it to sign a commercial agreement and share up to 30 percent of revenue. The weights are free. The business around them is not.
The Telegraph pointed out a second, older risk. "Other tech billionaires in China have found, to their cost, that the Communist Party likes its economic champions on a short leash." Jack Ma was barely seen for three years after a 2020 speech criticizing financial regulators. Earlier in 2026, Beijing had blocked the attempted sale of the agent startup Manus to Meta. Yang keeps his public speech technical. In March 2026 he told Chinese media something the state would have been happy to hear: "China has built the world's best talent system over several decades. From compulsory education to high school, university, and doctoral programs, we are able to cultivate people with the strongest investigative capabilities in the world." In Washington, the same talent system was being described as a funnel for extracting American work.
— David Sacks, on switching much of his work from Claude to KimiIt's just more fun because it gets straight to work instead of lecturing you.
Swarm
K3 came out on July 16, 2026. It had 2.8 trillion parameters, a one-million-token context window, native multimodality, and an Agent Swarm that now coordinated up to 300 sub-agents. Moonshot said it beat Claude Opus 4.8 max and GPT-5.5 high on benchmarks and trailed only Claude Fable 5 and GPT-5.6 Sol, at prices comparable to Anthropic's mid-tier Sonnet. On an independent private evaluation of long-horizon knowledge work it scored an Elo of 1547, 732 points above K2.6 and behind only Fable 5. It could optimize GPU kernels, produce research results in frontier physics, and edit video, including, as noted, its own trailer.
Six days before the release, Moonshot, the Agricultural Bank of China, and American Express announced two co-branded credit cards that pay out Kimi memberships based on spending. That is a very Chaoshan idea, loyalty points redeemable in intelligence, and it came from a company that had given up on consumer acquisition eighteen months before. When K3 sold out every subscription package, Moonshot chose not to cap usage for existing users, even at the cost of short-term revenue. It paused new sign-ups instead.
On July 30, Moonshot closed a round at a $35 billion valuation. The Series F was oversubscribed threefold and closed early. The pre-IPO round planned for August was pulled forward at a pre-money valuation reported at $50 billion. Unverified offers of secondary shares, supposedly from Yang's own holdings, circulated widely enough that the company had to say publicly that any transfer without its approval was void. On September 2, LatePost reported that Moonshot had confidentially filed its A1 application with the Hong Kong Stock Exchange. The company told AlphaRank: "we have no relevant information or response to release."
In public, Yang now talked less about models than about organizations. That had always been his deeper interest. In the November 2023 GeekPark interview, before Moonshot had any reputation, he had already said that OpenAI's success was at its core an organizational innovation: OpenAI "invented nothing new," yet an "industrial masterpiece" had come from combining the Transformer, about 10^25 FLOPs of compute, and twenty years of internet data. Google's environment, he said, could produce a scientific result like the Transformer but not a product like ChatGPT. "Only if you get the organization right can you actually walk the AGI road." The headline of that interview was his own sentence: Different organizations allow different things to emerge.
By 2025 he was describing management in the vocabulary of his own training runs. Supervised fine-tuning is top-down instruction: efficient, but it can kill initiative. Reinforcement learning sets a reward and lets the team find the path: more creative, but exposed to reward hacking, where people optimize the metric and miss the goal. "The risk of managing with SFT is that people lose creativity. The risk of managing with RL is that the reward gets hacked. It can look like you achieved a good result, while in reality you still missed the final goal." He said he was experimenting with an RL-centered company, with SFT as a supporting scaffold to give direction.
So the swarm grew. A hundred sub-agents in January, three hundred in July. At GTC he predicted that intelligence would move from single agents to swarms generated on the fly. His company had about 300 people, roughly one employee per sub-agent, and a valuation that by one Chinese estimate put his own shares in the billions. Its founder had once described the dream of a lifelong AI companion that wouldn't have to forget each day.
Pittsburgh's reaction had changed too. Salakhutdinov, who had thought going home was a bad idea, now posted about it with pride. "He was right," he wrote of Yang's decision. Zhang Xiaojun republished her 2024 interview that summer in English, now as a historical document, and added a note at the top: This translation was generated by Kimi K3. The journalist who had once worried that "a young AI scientist" might not "survive in a realistic China" had let his machine translate her doubts about him.
There is one more detail, and it is the one Zhang put first. Before the logo, before the Alibaba round, before anyone outside Haidian could say his name, a visitor walking into Moonshot's first office would find no sign telling them where they were. They would find a white piano by the door.
It was still there in 2026, in the JD Technology Building, standing prominently in the office. No profile has reported anyone playing it.
Part IIThe Playbook
The principles below are drawn from Yang Zhilin's record, not his rhetoric. Several of them sit in tension with each other, which reflects how he actually operates. He is a long-term absolutist who reverses strategy within a quarter. He is an open-weights evangelist who writes a 30 percent toll into his license. He is a researcher who prices capital before he picks a problem. The useful lessons come from the reversals, the arithmetic, and the moments when Moonshot's public story and its actual behavior diverged.
The principles
- 1.Run the regret calculation, not the risk calculation.
- 2.Price the ticket before you choose the game.
- 3.Make your research thesis legible as a single number.
- 4.Recruit the rhythm section you already trust.
- 5.Treat strategy as a checkpoint, not a vow.
- 6.When you lose the narrative, retreat to the layer you own.
- 7.Make efficiency the strategy under constraint.
- 8.Open the weights, write the toll into the license.
- 9.Break the glass: build for action, not demonstration.
- 10.Manage like a reward designer, and audit for reward hacking.
- 11.Your last company is a liability on a balance sheet you can't see.
- 12.Build a mythology, then keep your mouth technical.
Principle 1
Run the regret calculation, not the risk calculation
In 2019 the expected-value math clearly favored staying in America: postdocs at Stanford and MIT, an Apple executive offering a Beijing seat, an adviser who thought leaving was a mistake. Yang didn't argue with any of that. He changed the objective function. The question was not which path had the best risk-adjusted payoff but which one he could live with not having taken, and "If I don't do a start up, I will always regret" answers that directly.
What makes this more than a slogan is that he did it without drama. He didn't jump straight into a venture-backed moonshot. He spent four years moving between Huawei, BAAI, Qi Zhi, Tsinghua, and Recurrent AI while keeping a website that said, simply, "I am working on a start-up." Minimizing regret is not the same as being impatient. It sets a direction and leaves the timing open.
The 2026 immigration debate shows the cost of misreading this kind of decision. Commentators assumed a constraint because they couldn't imagine a preference. For founders the lesson runs the other way: the best people turning down your offer may not be rejecting its terms. They may be choosing a different identity.
Tactic: Before a major career fork, write down which option you'd regret not trying at seventy, and treat that answer as a constraint on the decision, not as one more input.
Principle 2
Price the ticket before you choose the game
The most revealing thing in Yang's founding story is not the Pink Floyd anniversary. It is the "precise calculation in Silicon Valley" that a foundation-model company aimed at AGI needed more than $100 million within a few months just to exist. He worked out the entry fee before he worked out the product. A year later that figure had grown thirteenfold, and Moonshot had raised it.
This runs against the usual startup advice: start small, find product-market fit, raise as traction allows. Yang's reading was that some markets are decided by capital concentration before they are decided by product. "It needs the concentration of talent and the concentration of capital." If you raise a seed round into a game whose table stakes are nine figures, you have joined a different, smaller game without noticing.
Knowing the price ahead of time also explains how fast he moved. He told Zhang Xiaojun the raise "didn't take much time," and that he couldn't tell many people he was doing it. When the number is clear, you stop debating whether to raise and focus on raising before the window closes.
Tactic: Before committing to a market, estimate the minimum capital needed to be a credible competitor in eighteen months, and if you can't raise that number, choose a different market or a different layer of it.
Principle 3
Make your research thesis legible as a single number
Kimi launched in October 2023 with a single headline claim: 200,000 Chinese characters per conversation, about eight times GPT-4-32K. That number was Transformer-XL, Yang's 2019 paper on getting past a fixed-length context, turned into something a lawyer, an analyst, or a student could understand at once. Five months later it became two million.
Yang supported the number with an analogy anyone could follow. The Transformer is a computer, parameters are the CPU, context is the memory, and people once thought 500K was plenty. The analogy did two jobs. It told users what they were getting, and it told investors that the advantage would compound, because memory demands always grow.
The lesson isn't that long context was the right bet forever. Yang himself predicted it would become standard. The lesson is that a company founded on a technical insight should be able to state that insight as a measurable, comparable figure. Other labs at the time were selling "a Chinese ChatGPT." Moonshot sold a specification.
Tactic: Reduce your core technical advantage to one number a customer can compare against a competitor's, and lead with that number at launch.
Principle 4
Recruit the rhythm section you already trust
Moonshot's co-founders, Zhou Xinyu and Wu Yuxin, were Yang's Tsinghua classmates and, by most accounts, his Splay bandmates. Their skills fit together precisely. Zhou knew how to run neural networks on constrained hardware. Wu knew foundation models from Google Brain and vision from Meta. Yang brought long context and generative pretraining. Within three years, the company's model line had combined exactly those three: efficiency, scale, and multimodality.
A band is a useful model for a founding team because a band can't work with resentment between its members. Timing is shared. The drummer can't rush. Teams formed by recruiting strangers into vacancies take years to develop that kind of shared timing. Teams formed from people who have already played together start with it.
The risk is insularity, and Yang guarded against it by bringing in people who were not from the band. Zhang Yutong was an outsider with capital-markets skill, and the Recurrent AI dispute showed how complicated adding someone with outside loyalties can become. The core stays trusted. The edges have to be managed.
Tactic: Build your founding core from people whose working rhythm you've already tested under pressure, and assemble it so their skills match the capabilities your product roadmap will need to combine.
◐
The Reversals
Moonshot's 2023 positions versus its 2026 practice. Few founders have publicly reversed this many core bets this quickly and come out stronger.
| Dimension | 2023 position | 2026 practice | Status |
|---|---|---|---|
| Model access | Super-app entry point will most likely be closed-source | Open weights from K2 onward; K3 the largest open model ever | Reversed |
| Go-to-market | "One and only one consumer product" | Abandoned C-end traffic acquisition; API, Kimi Code, and open ecosystem | Reversed |
| Core capability | Long context as headline differentiator | Long context as one of three multiplying axes, with token efficiency and agent swarms | Absorbed |
| Paradigm | Chat and reading assistant | Agents executing hundreds of sequential tool calls | Extended |
| Monetization | Consumer tiers from 5.2 yuan for four days | License revenue share up to 30%; co-branded credit cards; enterprise | Diversified |
| Organizing thesis | "Only if you get the organization right…" | RL-centered management with SFT scaffolding | Constant |
Principle 5
Treat strategy as a checkpoint, not a vow
In November 2023 Yang argued, carefully and convincingly, that the winning AI product would be closed-source because owning the model's evolution differentiated the product from day one. By July 2025 he was releasing a trillion-parameter model's weights under a modified MIT license. When GeekPark republished the 2023 interview in English, its editors added a note saying this was "the bet Moonshot has since reversed."
The phrase that best explains how he can do this comes from his own technical vocabulary. A checkpoint is a saved state of a model in training. You can evaluate it, restore from it, or throw it away. The commitment is to the training objective, AGI and maximizing the value of AI in his GitHub bio's words, not to any particular checkpoint. That is how Yang can say "AI is not about finding a PMF in the next year or two; it's about how to change the world over the next ten to twenty" and also make a hard strategic turn within one quarter. The long horizon is what makes the short-term reversal affordable.
Most founders do the reverse. They are flexible about the mission and stubborn about the strategy, because the strategy is what they announced in public. Yang has been stubborn about the mission and has dropped strategies without ceremony.
Tactic: Separate your company's objective from its current strategy in writing, and review the strategy against the objective each quarter as if it were a model checkpoint you can discard.
Principle 6
When you lose the narrative, retreat to the layer you own
In early 2025, by AlphaRank's account, Moonshot faced "the collapse of the commercialization path, the loss of public opinion momentum, and the stagnation of financing." DeepSeek had taken the genius story. Kimi was falling toward seventh place. The obvious move was to spend more on users to defend share. Yang did the opposite. He stopped buying traffic and put everything into the foundation model, the one layer where his team's credentials, Transformer-XL, XLNet, Muon, Mooncake, were hard to replicate.
This is a retreat in the military sense: you give up ground you can't hold to concentrate on ground you can. Paid user acquisition was a contest any well-funded rival could win. Model research was a contest decided by the specific people Moonshot had. Within eighteen months, the model layer brought back everything the consumer layer had lost, including the users, who now arrived for K3 in numbers that forced Moonshot to pause sign-ups.
Yang's second move in the same period mattered just as much. He picked the next frontier, agents, rather than chasing the one DeepSeek had just won, reasoning. Moving to the layer you own is half of it. The other half is choosing a direction where the leader isn't already standing.
Tactic: When a rival takes your story, list the capabilities only your team can credibly build, cut spending on everything a better-capitalized competitor could copy, and point the freed resources at the next frontier, not the one just won.
Principle 7
Make efficiency the strategy under constraint
Moonshot had fewer chips than its American rivals, and probably fewer than it wanted, which is why reports of H200 access through Alibaba Cloud became news. The company's answer was to make efficiency its main product. Muon doubled compute efficiency over AdamW. Mooncake served 100 billion tokens a day and won a best-paper award. K2 Thinking used native INT4 quantization and reportedly cost about $4.6 million to train. At GTC, Yang described scaling as no longer "simply about resource accumulation" but as finding scale effects in "computational efficiency, long-range memory, and automated collaboration" at the same time.
His shorthand for token efficiency, eating one portion of data and learning two portions of knowledge, also describes the business. The constraint doesn't only limit Moonshot. It shapes the company's research identity, and that identity is now an asset: K3 offered near-frontier performance at prices comparable to a mid-tier American model, and that price gap is what shook the markets.
This works only if the efficiency research is published. Moonshot's papers on Muon and Mooncake did as much as its benchmarks to establish it as a serious lab rather than a fast follower. Efficiency you keep hidden is a cost advantage. Efficiency you publish becomes a reputation.
Tactic: Identify your scarcest input, make reducing its use a named research program with published results, and present the resulting cost advantage as a capability, not as thrift.
Principle 8
Open the weights, write the toll into the license
Kimi's licenses show a business model in their fine print. K2's modified MIT license required attribution for products over 100 million monthly users or $20 million in monthly revenue. K3's custom license goes further: any host selling access as a service and earning more than $20 million a year must sign a commercial agreement with up to 30 percent revenue sharing.
The design lets almost everyone use the model for free, and that distribution is enormous. Kimi became the base for Cursor's Composer 2, Thinking Machines' Inkling, and Cognition's Devin, which is a kind of enterprise reach that no sales team in Beijing could have bought, especially under American political scrutiny. The toll applies only to the few who are making a lot of money from it. Openness acquires the customers. The license collects from the largest ones.
This is a more developed position than the old open-versus-closed argument. It treats openness as a pricing tier, not as an ideology. Yang's public language, "democratising intelligence, with open models you can deploy anywhere," is sincere, and it also describes a funnel.
Tactic: If you open-source a core asset, set a revenue threshold above which commercial terms apply, so that the ecosystem funds your distribution and its biggest beneficiaries fund your company.
Principle 9
Break the glass: build for action, not demonstration
Yang's brain-in-a-vat image is the clearest version of a product principle many AI companies are still learning. A pure reasoning model "does not need to interact with the outside world. It is like a fish tank." It looks impressive in a demo and fails in production because it never checks its work against reality. An agent searches, runs code, reads the result, and tries again.
His willingness to praise a competitor for exactly this is instructive: "Claude's reasoning performance is not very high, but its performance as an agent is very high." He saw that Anthropic had chosen action over thinking, decided the choice was right, and moved Moonshot to compete on that axis instead of the one where DeepSeek led. The results show up in the product line: 200 to 300 sequential tool calls in K2 Thinking, a 100-agent swarm in K2.5, 300 in K3, and a model that edited its own launch video.
The general point goes beyond AI. Products tuned for the demo improve on the measures people look at. Products tuned for the loop improve on the measures that decide whether customers keep using them.
Tactic: For every product capability, define the feedback loop through which it touches the real world, and measure the loop's success rate, not the demo's impressiveness.
Principle 10
Manage like a reward designer, and audit for reward hacking
Yang's management framework comes straight from his training runs. SFT, or top-down instruction, is efficient and suppresses creativity. RL, or setting a reward and letting the team find the path, maximizes creativity and invites reward hacking: "It can look like you achieved a good result, while in reality you still missed the final goal." He runs Moonshot as an RL-centered organization with SFT as a supporting scaffold for direction.
The analogy works because it names the failure mode. Every goal-setting system eventually gets gamed. Benchmarks are the obvious case in AI. Yang himself has warned that agent training overfits to tests like SWE-Bench and fails on real tasks. A CEO who thinks in terms of reward design will assume the team's metrics are being optimized away from their purpose and will build ways to detect it.
It also explains why he believes organization matters more than any single model. "Different organizations allow different things to emerge." The model is a product of the organization's reward structure, so the reward structure is what the CEO is actually responsible for.
Tactic: For each team goal you set, write down in advance the most plausible way it could be met without achieving the real objective, and assign someone to check for exactly that.
↺
The Arc
Key decisions and inflection points, 2011–2026.
2011
Enters Tsinghua thermal engineering after earning admission three separate ways; transfers to computer science the following year.
2014
Wins Tsinghua's Special Class Scholarship with 13 of 14 votes, nominated by Tang Jie.
2016
Co-founds Recurrent AI while a CMU doctoral student.
2019
Publishes Transformer-XL and XLNet; finishes PhD in four years; declines Apple, Stanford, MIT; returns to China.
2020–21
Works on Huawei PanGu, then leads a Wu Dao team at BAAI.
2023
Founds Moonshot AI on March 1; launches Kimi in October with a 200,000-character context.
2024
Alibaba-led ~$1B round at $2.5B; Tencent joins at $3.3B; Recurrent AI investors take dispute to Hong Kong arbitration.
2025
K1.5 overshadowed by DeepSeek R1; abandons traffic acquisition; releases open-weight K2 and K2 Thinking.
2026
K2.5 and GTC keynote; Anthropic distillation accusations; K3 released July 16; $35B round; confidential Hong Kong IPO filing.
Principle 11
Your last company is a liability on a balance sheet you can't see
Recurrent AI gave Yang his first experience of enterprise AI, his first investors, and an early commercial use for Transformer-XL. It also produced the most damaging episode of Moonshot's early years. Five of its investors took him to arbitration in Hong Kong over how he had left and how value had moved to the new company. A partner at one of those funds attacked a Moonshot insider in public.
Yang's defense rested on documentation: every necessary procedure "had received the signed consent of every board member." That is a strong position, and it apparently held up well enough to reach a settlement. Even so, the dispute took executive attention and reputation during the most competitive stretch of the company's life. Recurrent AI was the reason Moonshot could move quickly. It was also the reason Moonshot had a liability that never appeared on its own books.
For serial founders the lesson is structural. Prior investors, prior co-founders, and prior equity arrangements are all claims on your attention and, sometimes, on your new company's value. Deal with them deliberately before the new company's success makes them worth fighting over.
Tactic: Before you launch a second venture, get written, board-level sign-off on your departure from the first and document how any shared people, IP, or investors will be treated, because the dispute will come when the new company is worth fighting over.
Principle 12
Build a mythology, then keep your mouth technical
Yang built the most distinctive founder story in Chinese AI. The company was named for The Dark Side of the Moon and founded on its fiftieth anniversary. There's a white piano at the door, a flagship product named after him, and a career change prompted by Murakami. The Financial Times credited that mythology with setting Moonshot apart from "China's otherwise austere AI industry." It helped with recruiting and with fundraising, and when K3 arrived, global media had a story ready to tell.
In public, though, Yang stays almost entirely technical. The Telegraph noticed that after one sentence about democratizing intelligence, his forty-minute GTC keynote "sticks safely to the technical realm." Given the fates of Jack Ma and Manus, that discipline looks deliberate. The founder story is built from things that can't hurt him, like rock music, novels, and the moon. The opinions that could hurt him stay private.
The broader lesson applies anywhere a founder answers to powerful stakeholders, whether regulators, governments, or a dominant investor. Make your personal brand out of taste and craft, not commentary. Taste attracts people, and commentary makes enemies.
Tactic: Build your public identity from aesthetic and technical commitments that energize your team, and keep political or regulatory opinions out of anything that will be recorded.
Free playbook
Get The Business Model Playbook
58 business models, one visual page each: how the money flows, the metrics that matter, and who runs it. Free when you join the Faster Than Normal email.
Free. No spam. Unsubscribe anytime.
Part IIIQuotes and Maxims
In their words
— Yang Zhilin, interviewed by Zhang Xiaojun, published March 1, 2024If everyone thinks you are normal—if your dream is one that anyone could have—it adds nothing to the sum total of humanity's dreams.
— Yang Zhilin, interviewed by Zhang Xiaojun, March 2024AI is not about finding a PMF in the next year or two; it's about how to change the world over the next ten to twenty.
— Yang Zhilin, on managing a company like training a model, 2025The risk of managing with SFT is that people lose creativity. The risk of managing with RL is that the reward gets hacked. It can look like you achieved a good result, while in reality you still missed the final goal.
— Yang Zhilin, on transferring out of thermal engineering at TsinghuaEverything that happened afterwards convinced me that transferring to the Computer Science Department was the most important and correct decision I've ever made.
— Yang Zhilin, quoting Daniel Kahneman at Tianjin University, September 14, 2024Many times, you have the courage to attempt something unknown because you don't realize how much you don't know. Once you try, you discover many new problems, and perhaps that's the essence of innovation.
Maxims
- Price the ticket first. Work out the capital a credible entry requires before you fall in love with the game.
- Regret is a valid objective function. When every option looks good, choose the one you couldn't bear not to have tried.
- One number beats a paragraph. Turn your deepest technical insight into a figure a customer can compare.
- Hold the mission tightly and the strategy loosely. A long horizon is what makes a fast reversal affordable.
- Lose the story, keep the layer. When a rival takes the narrative, retreat to the ground only your team can hold.
- Constraint makes good research. Publish your efficiency gains and scarcity becomes a reputation.
- Open the gate, keep the tollbooth. Free distribution for the many and commercial terms for the few who profit most.
- Break the glass. A brain in a vat can impress you, but only an agent that acts gets judged against reality.
- Every reward will be hacked. Design the incentive, then design the check for how it will be gamed.
- Let taste speak for you. Build the founder myth from music and craft, and keep opinions out of the record.
