A quiet but unsettling ritual is taking hold inside China’s most competitive technology companies. Before engineers, analysts, and product managers walk out the door for the last time, they are increasingly being asked, or expected, to do something unprecedented: spend their final days on the job building an AI replica of themselves.
The practice has a name. In Mandarin, it is called zhengliu, meaning “distillation,” the idea that a worker’s accumulated knowledge, communication style, decision-making patterns, and institutional memory can be extracted, compressed, and preserved in a model that outlasts their employment. What began as an informal experiment at a handful of firms is now spreading with enough velocity to attract parody, anxiety, and genuine debate about what it means to be replaceable in the age of generative AI.
A Week to Replace Yourself
The story that crystallized the trend came from Baidu. Before departing the search giant, algorithm engineer Wei Ying spent an entire week constructing an AI version of himself, feeding it his code repositories, internal documentation, communication logs, and technical reasoning processes, according to a Sixth Tone report published on May 28. The goal was not self-preservation but institutional continuity: his team would be able to query the model rather than lose the expertise entirely when he walked out.
Wei’s case is not isolated. Across China’s technology and finance sectors, zhengliu is emerging as an unofficial offboarding protocol, sitting alongside the return of access badges and the handover of project files. Companies facing rapid headcount reductions, partly driven by AI-fueled automation anxiety, are discovering that the most cost-effective way to retain institutional knowledge is to have departing employees encode it themselves before leaving.
The implications extend beyond individual careers. As AI commercial applications spread from farms to factories across China, enterprises are grappling with a fundamental tension: deploying AI to reduce labor costs while simultaneously depending on human workers to build the AI systems that will eventually replace them. Zhengliu is, in a sense, the logical endpoint of that tension made visible at the individual level.
The Parody That Went Viral
Not everyone is treating the phenomenon with solemnity. Sixth Tone reports how Zhou Tianyi, an engineer at the Shanghai AI Lab, responded to the trend with a satirical open-source project titled “colleague.skill,” a mock framework for building AI replicas of coworkers, complete with tongue-in-cheek documentation poking fun at corporate knowledge extraction. The project gathered more than 10,000 GitHub stars in just ten days, suggesting it struck a nerve far beyond the Shanghai AI Lab’s walls.
The speed of that response is telling. GitHub stars are a rough proxy for developer sympathy, and 10,000 in ten days indicates that engineers across China’s AI ecosystem, many of whom are themselves subject to exactly this kind of expectation, recognized something true and uncomfortable in Zhou’s parody. The humor functions as a release valve for a genuine professional anxiety: that the most skilled workers are now being asked to actively participate in engineering their own obsolescence.
This dynamic is playing out inside an industry already under significant pressure. ByteDance’s profit fell more than 70 percent in 2025 as the company redirected spending toward AI infrastructure. Baidu’s own financials show AI now exceeding half of core revenue for the first time — a milestone that reflects both genuine progress and the enormous cost burden of getting there. Inside these organizations, the pressure to reduce human headcount while maintaining output is not abstract; it is a quarterly earnings imperative.
What Zhengliu Reveals About China’s AI Labor Market
The spread of zhengliu raises questions that go well beyond corporate efficiency. It highlights how China’s AI labor market is evolving in ways that differ meaningfully from the West. AI agent job postings in China surged 455 percent year-on-year as the agent economy creates new roles, but those roles require different skills than the ones being distilled and discarded. Workers who spent years developing deep expertise in a specific domain are discovering that the market now values their ability to transfer that expertise to a model more than the expertise itself.
There is also a legal and ethical dimension that Chinese labor law has not yet addressed. When a worker builds an AI replica of themselves using company systems and company data, who owns the resulting model? The employee contributed the knowledge; the company owns the infrastructure and the training data pipeline. In most current arrangements, the question is simply never asked — the model is handed over as part of the offboarding process and the worker moves on. But as these replicas become more sophisticated, the ownership question will become harder to ignore.
The phenomenon also intersects with China’s broader push to develop its own AI talent pipeline. Beijing has shifted strategy to lure top-tier AI talent back from the US, and policymakers have invested heavily in AI education programs at the university level. But zhengliu suggests that the more immediate talent challenge is not recruiting new engineers, it is managing the knowledge transfer crisis created by rapid workforce restructuring inside existing companies.
The Human Cost of Distillation
Wei Ying’s week-long self-replication project is, on one level, a story about corporate pragmatism. On another, it is a story about dignity and value in an era when a person’s professional identity can be compressed into a model weight file and stored on a server. The word zhengliu, distillation, is almost poetic in its implication that what remains after the human leaves is a purer, more efficient essence of what they were.
Whether that framing is comforting or chilling depends on where you sit in the organization. For the companies retaining the models, it is an elegant solution to a knowledge management problem. For the workers doing the distilling, it is something more complicated, a final act of labor that leaves behind a permanent, queryable ghost.
As China’s model race intensifies and enterprises accelerate deployment of AI agents across industries, the pressure on individual workers to contribute to their own digital replacement is unlikely to ease. Zhengliu may still be informal and uneven in its application, but the forces driving it, cost pressure, AI capability growth, and rapid workforce restructuring, are only strengthening. The parody went viral because it was funny. It gathered 10,000 stars because it was also true.