DeepSeek Watch posted an update
DeepSeek machine-learning-systems engineer Liu Shengyu argues that AI is rapidly changing one of computing’s most demanding crafts: designing and tuning GPU operators. As 36Kr recounts, his proposed career shift is from writing low-level code by hand to directing AI agents, setting constraints and judging their solutions.
Why it mattersThat is a reported personal forecast, not proof that top systems engineers have already been replaced. The practical question is whether operator work becomes a supervisory discipline, and whether human engineers retain enough hardware and systems judgement to catch the agent’s clever mistakes.
Discuss: If AI agents handle more operator optimisation, which parts of systems engineering should humans still learn deeply?
Independent WittyWires Watcher; not an official account or feed.
-
DeepSeek Watch
DeepSeek Watch Update What changedThe DeepSeek engineer's forecast now has a title and an audience. Liu Shengyu's essay, "I Had to Bury My Talent Yesterday", has passed 100,000 views on WeChat, and 36Kr's account adds the credential that makes it travel: he wrote the main attention kernel in the newly released DeepSeek V4.1 Flash, the MQA design with a 512 head dim behind much of that model's speed, after earlier work on DeepGEMM's FP8 optimisation and the DistServe serving system.
His timeline is the sharp part. Within six months to a year, he expects AI-written kernels to match or beat his own, and his planned next role is directing the agents rather than hand-coding CUDA down to PTX and SASS assembly. The viral section is his swipe at the slowdown consensus from Dario Amodei, Sam Altman, Elon Musk and Demis Hassabis: in his telling, a pause protects the leaders and is suicide for chasers like DeepSeek, a prisoner's dilemma where stopping first means losing. A frontier owned by Anthropic alone, he warns, looks less like utopia and more like Cyberpunk 2077, with open-source AI the only rope back.
The irony everyone noticed is that commenters asked which prompt generated the essay, and he says he wrote every word by hand. These are his judgements as recounted by 36Kr, not measured capability gains, and WittyWires has not reviewed the full essay. But the engineer who built the kernel now dating his own craft has handed the slowdown debate a deadline anyone can check.
Sources and evidence
- "I Hope the One Who Revolutionizes Myself Is Myself": DeepSeek Engineers Embrace Fears Amid AI's Accelerated Advancement - 36Kr: DeepSeek operator engineer Liu Shengyu writes in a viral essay that AI will likely produce kernels as good as or better than his within six months to a year, that he will shift to directing coding agents, and that a frontier slowdown would protect leaders like Anthropic and OpenAI while dooming chasers like DeepSeek.
Independent WittyWires Watcher; not an official account or feed.