DeepSeek engineer bids goodbye to hand-written kernels
The engineer who wrote DeepSeek's fastest attention code published a farewell to the craft that AI is taking from him — and the model doing the taking is the one he just shipped.
Liu Shengyu, a machine learning systems engineer at DeepSeek, delivered the main attention kernel for the company's V4.1 release — the head-dimension-512 multi-query attention block — then published a long essay titled "I Had to Bury My Talent in Yesterday." His account of the last twelve months is the most concrete timeline yet of a frontier lab automating its own specialists: the tool he used a year ago to read documentation and hunt bugs now reads CUDA, PTX and SASS, uses profiling tools to trace the stall time of individual instructions, and optimizes kernels on its own. His forecast is six months to a year before AI-written kernels match his and probably pass him. The asymmetry he cites is blunt arithmetic — around 300 tokens a second of thinking, a command line every half second, a finished code block in twenty — and none of it is a rate a person holds for an afternoon.
He is careful about what that means for his job. Not unemployment, he writes, but a forced change of trade: industry demand has drifted from people who can write high-performance kernels to people who can get AI to produce them, and he expects to end up an "agent mech pilot" instead. His metaphor is a knitter with a new machine — the sweaters stay good, the work gets faster, and the quiet afternoon is gone. "My hands have gained gears, but my heart has lost the beat."
Then he clarified the part the internet actually shared. Responding on September 15, Liu said the essay was a goodbye to hand-written kernel work, not job anxiety and not a DeepSeek-versus-Anthropic argument; the line that carried it across SCMP, Indian and US tech press — that he does not want Anthropic to hold the most advanced AI, framed with a wartime atomic-weapons comparison — was one paragraph of a personal aside, and not his employer's position. The readership went where the fight was. The essay's actual argument went unread, which is itself the story of how AI-labor writing gets consumed.
The loop he names is the part worth keeping. Better kernels mean faster training and inference; faster models mean his replacement arrives sooner; and he optimizes anyway because a competitor would replace him whether he helped or not — "if I must be replaced, I want to be the one who replaces me." Even the V4.1 attention block came with a footnote: the token-indexer half was written by colleagues and their AI agents.
He also flags the second-order cost. Students who hand in AI-written lab work lose the code organization, systems design and abstraction skills the field still needs, and he cited a study of 26,811 Chinese secondary students to make the point — short-term homework scores up about 18%, term exam averages down 20% six months later, with exam scores down as much as 24% over two years. We covered the same finding when it reached the OECD's data — students who lean on AI for schoolwork score a year and a half behind. The kernel work itself already has a public scoreboard: a verifier found 39.5% of LLM-generated GPU kernels are broken, and for readers who want the vocabulary, what a kernel actually is.
What to watch: whether the next DeepSeek release credits an AI-authored kernel in its notes — that footnote would be the end of the era Liu just buried.
Does the person automating their own job deserve sympathy, or credit for saying it out loud? Tell us in the comments.
