On September 14, Liu Shengyu, a kernel engineer at DeepSeek, published a long personal essay titled “I Have to Bury My Talent Yesterday” on his WeChat account. It passed 100,000 reads, topped Zhihu's hot list, and spread on X. The essay does not reach for grand narratives. It describes one thing: a young man who has spent years hand-writing high-performance GPU kernels has concluded that AI will likely match his craft within six months to a year—and that he is helping to speed up that day himself.
[1][2]Liu is a 2021 graduate of the Turing Class at Peking University's School of Electronics Engineering and Computer Science. He joined DeepSeek in 2025 to work on machine learning systems and kernel design and optimization, and he captained Peking University's supercomputing team. When DeepSeek released V4.1, the model's main attention kernel—a head-dim-512 MQA attention that underpins the model's inference speed—was delivered by him.
His assessment is specific rather than rhetorical. A year ago, AI could help him look up documentation, read code, and find bugs. Now, he writes, AI can read CUDA, PTX, and SASS on its own, use professional tools to analyze instruction stalls, and optimize kernels by itself. “AI can think 300 tokens per second, type a command in half a second, write a piece of code in twenty seconds. I can't.” From this he draws a cruel corollary, which he calls the technical paradox: “The better I write kernels, the faster our new models train and infer, the faster model capability improves, and the sooner I am replaced.”
He does not plan to stop. The reason is blunt: even if he coasted or quietly sabotaged his own work, other companies' models would improve anyway and eventually replace him all the same. “Of course I hope I won't be revolutionized. But if revolution is unavoidable, I hope the person who revolutionizes me is myself.”
The essay's emotional center is “career change.” He expects not to be unemployed but to be forced out of his craft: from hand-writing kernels to becoming what he calls a “mecha pilot” directing agents that write kernels. His metaphor is knitting. A knitter of twenty years, loved for her patterns, gets a machine that knits faster and just as well. The livelihood survives; the pleasure of sitting by the window, threading needle and yarn through a slow afternoon, is crushed by the machine's roar. “I have to bury my talent yesterday and become a mecha pilot. My hands hold more gears, but my heart holds fewer beats.”
There is a governance argument folded into the essay. The future, he writes, may split toward two extremes: a communist-style liberation of productive forces, or a Cyberpunk-2077 world in which a few companies control the best intelligence. He does not trust Anthropic or OpenAI to distribute frontier intelligence openly and cheaply—“if Anthropic permanently controlled the world's most advanced AI, the severity would be no less than Hitler getting the atomic bomb before the Allies.” This, he says, is why he stays at DeepSeek: open-source, accessible AI is the rope that might pull the world back from the 2077 end.
On the evening of September 15, Liu posted a follow-up. The essay's viral spread surprised him; its purpose, he said, was neither to express anxiety about unemployment nor to set DeepSeek against Anthropic, “but to say goodbye to the time of hand-writing kernels.” He noted that the passage in question was his personal view, not his employer's. The follow-up matters because it clarifies what the essay is: not a fear story, not a victory lap, but one specific person recording, with unusual candor, how it feels to watch the craft he loves become history while he is still holding the pen.
[1][2]