Yang Zhilin Went Back

The day Kimi K3 dropped, a post appeared on arXiv with a title of just four Chinese characters: “He went back.” No signature. But everyone knew who it meant.

2.8 trillion parameters. Open‑source. Topped every global code benchmark. The tech circles picked apart those numbers, debated them, compared them. But the thread beneath that post veered in another direction: Why?

Yang Zhilin. Top scorer of the Gaokao in Shantou, Guangdong. Tsinghua. CMU Ph.D. His advisor is the head of AI at Apple. If he wanted, he could have an O‑1 visa by the next morning. The U.S. immigration records still list his OPT status as “active.”

But he is in Beijing.

Americans on forums tried to find an external reason. Visa denial? No. Insufficient pay? His advisor’s position at Apple means any American AI company would hand him a blank check. So what was it, then?

In Silicon Valley, Chinese engineers chat in parent groups about “the kids can’t speak Chinese anymore,” and in church groups about “our parents are getting old and we can’t bring them over.” In Yang’s office at CMU, he had pinned a photo of his old home in Shantou. Beside it, taped to the wall, was his father’s phone number, written by hand. It stayed there for three years.

Someone brought up the saying that “for Chinese in America, the family dies out in three generations.” The phrase isn’t precise, but it points at something concrete: the first generation grinds their teeth and takes root; the second works hard to assimilate; the third no longer remembers their grandparents’ mother tongue. It has nothing to do with money. The wealthiest Chinese‑American families in Silicon Valley still sit in Sunday Chinese schools, watching their kids doze off in the back row.

Part of why Yang came back was that. But he never mentions it. In interviews, he only talks about “compute costs” and “team dynamics.”

Moonshot AI’s office is in Beijing. Yang’s desk is by the window. On the windowsill sits an old Casio calculator, its screen crossed by two cracks. That was the one he used in his senior year of high school. Someone asked him why he didn’t replace it. He said, “It still computes.”

In the physics lab at Shantou Jinshan Middle School, a line of text was scratched into a desk with the tip of a compass. Later, when asked about it, Yang said it was a line from Schrödinger’s equation; he carved it to remember a moment, the moment when “it suddenly clicked.” China’s education system got one thing right: it let that moment happen in Shantou.

The year Yang took the Gaokao, he was the top science scorer in Guangdong. That same year, in the same province, another top science scorer came out of Zhanjiang: Liang Wenfeng. Twenty years later, the two sat next to each other on the same row of chairs at a closed‑door meeting in Guangzhou. In the 2000s, group photos of Chinese tech billionaires showed real‑estate developers. In the 2020s, they show people in hoodies. The two men weren’t talking about land prices; they were talking about parameter efficiency. Someone nearby took a picture: Liang was drawing a curve on a piece of paper, and Yang reached over to correct it. The sound of the pen tip hitting the paper was very soft.

The Economist ran a piece saying that this generation of Chinese tech billionaires no longer builds on real estate. The editor probably never knew about that pen.

On the American side, a Silicon Valley VC wrote in an internal memo: “They did the same thing with a fraction of our budget.” He appended a number: one‑sixth of OpenAI’s per‑training cost. He didn’t explain why. The reason, in America’s tech industry, can’t be spoken aloud. The big companies have deep compute pools, but every layer has a sign‑off. And every person signing off is paving the way for their next position. This isn’t a technical problem.

Someone posted a long thread on X, arguing that “Yang’s achievements are a product of American education.” The first reply came from an account with the handle “cmu_2017”: just one sentence: “I shared an office with him at CMU. He didn’t go to a bar in Pittsburgh once in three years. His creativity wasn’t inspired; it was starved out.” That reply got retweeted over ten thousand times.

In the acknowledgements of his Tsinghua undergraduate thesis, Yang wrote seven lines. The last line read: “I thank my parents. They can’t read what I write, but they never once asked, ‘What’s the use of this?’”

After he came back, he said something at an internal meeting that an employee later wrote down. His exact words: “We don’t lack smart people. We lack the ones who don’t need encouragement to keep going.” That sentence was later taped to the wall of Moonshot AI’s break room. The person who put it up used tape on all four corners, very neatly.

After Kimi K3 launched, a reporter asked him: “Why didn’t you stay in the U.S.?”

He looked down at the old Casio calculator. The cracks were at the lower right of the screen; it didn’t affect the reading.

He said: “I can compute faster here.”

The reporter waited, expecting more. There was none.

Outside the window, Beijing winter light fell on the keyboard mat, throwing a small patch of white. Beside the desk, a pot of pothos was half wilted. No one had watered it.

That night, someone opened an issue on GitHub, asking about a certain function in Kimi K3. An unverified account replied. The reply was short: just a line of mathematical derivation. Beneath the derivation, no signature, only two Chinese characters:

“Tomorrow.”

The account’s avatar was a photo of an old house in Shantou. In the corner of the photo, a phone number was written by hand.