A chatbot that can read an entire book in one go did not begin with a lab full of servers. It began with a researcher annoyed that language models forget the first page by the time they reach the last.
Yang Zhilin (杨植麟) is the founder and CEO of Moonshot AI (月之暗面), the Beijing company behind Kimi, the assistant that made long context (长上下文) a household phrase among Chinese AI users. At 32, he is one of the few China-born model builders whose research papers predate his own company.
From thermal engineering to the top of Tsinghua
Yang was born in 1992 in Shantou (汕头), a coastal city in Guangdong province. His path looked straight — until he bent it. He entered Tsinghua University in the thermal-energy engineering department, then switched to computer science, where he graduated first in his class in 2015.
He has said the turn was almost accidental. Accounts of his student years describe a detour through a novel about a programmer and a campus band — he played drums in a group called Splay. The musical detail is more than color. Moonshot's Chinese name, Yue Zhi An Mian (月之暗面), is a direct nod to Pink Floyd's "The Dark Side of the Moon," and Kimi carries his own English name.
The papers before the product
Before founding anything, Yang built the intellectual toolkit that Kimi later turned into a product. At Carnegie Mellon University he earned a PhD in 2019 — in four years, against a norm closer to six — working with Ruslan Salakhutdinov and William Cohen.
He is a co-author of two papers that shaped modern language modeling:
- Transformer-XL — a method for letting models reuse representations from earlier text instead of forgetting it, directly attacking the fixed-context-window limit.
- XLNet — an alternative pretraining approach that influenced how later models learn from unlabeled text.
He also contributed to HotpotQA, a multi-hop reasoning dataset, and interned at Google Brain and Meta AI. Salakhutdinov has publicly described him as "absolutely brilliant" and recalled that Yang turned down senior industry offers to start his own company.
Recurrent AI, then Moonshot
Yang's first company was Recurrent AI (循环智能), an enterprise AI firm focused on analyzing sales conversations. He co-founded it before Moonshot. That earlier venture's investor disputes later became a footnote to Moonshot's rise, but the throughline was consistent: a fixation on how much context a machine can actually use.
In early 2023, as ChatGPT reset global expectations, Yang launched Moonshot AI in Beijing with the explicit goal of building toward artificial general intelligence (AGI, 通用人工智能). The company's first product, Kimi, leaned into a single differentiator — an unusually long context window that could ingest hundreds of thousands of Chinese characters at once, letting a user "talk to" a full report or contract. International reporting describes Kimi's early context capability reaching roughly two million characters.
Why "open" became the strategy
Moonshot's public reputation rested first on consumer reach, then on model capability. Yang has argued that scale, not clever new algorithms, is the primary lever — "if you can solve it with scale, don't solve it with a new algorithm," he told an interviewer — and that open ecosystems will ultimately outweigh closed ones in total value.
That conviction pushed Moonshot toward open weights (开源权重). The company drew backing from major Chinese internet firms, including Alibaba and Tencent, according to international reporting. The open-release move placed Kimi among a small set of Chinese models that developers outside China could actually download and run.
What makes the profile matter
Yang is not the only young China-born AI founder, but his arc is unusually complete:
- A researcher whose work (Transformer-XL, XLNet) is citation infrastructure for the field.
- A founder who returned to China rather than staying in the US, at a moment when US immigration friction was pushing talent outward.
- A builder betting that context length and open weights, not just benchmark scores, define usefulness.
His story is also a window into a broader shift: Chinese model labs increasingly treat open release as a strategy, not a concession.
Honest limitations
Biographical facts — birth year and city (1992, Shantou), Tsinghua (thermal engineering to computer science, graduated top of class 2015), CMU PhD 2019 under Salakhutdinov and Cohen, co-authorship of Transformer-XL and XLNet, the Recurrent AI precedent, and Moonshot's founding in early 2023 with the Kimi product — are drawn from CNN, Business Insider, ThinkChina/Caixin, and YesPress profiles. The claim that Kimi's long-context capability reached roughly two million characters is reported by Caixin (via ThinkChina). Company valuation and precise user counts are intentionally omitted: they appear only in secondary or self-reported sources and could not be confirmed against a primary filing. Quotes attributed to Yang (the "scale over algorithm" line; the Pink Floyd naming story) come from media interviews, not documents we independently verified. This profile contains no RMB amounts, so no currency conversion applies. Analysis is current to 2 October 2026.
What readers can do now
- If you study AI, read Transformer-XL and XLNet not as history but as the reasoning lineage behind long-context products like Kimi — the "memory" problem is still unsolved.
- If you build products on Chinese models, test Moonshot's open-weight releases directly; weigh context length against your document-heavy workloads.
- If you follow the China–US talent flow, treat Yang's return as a data point: capability is now being built in both places, and "where the founder lives" no longer predicts "where the lab wins."