DeepSeek Engineer Says AI Progress Could Push Him From Kernel Optimization to 'Mech Pilot'

Liu Shengyu, a machine learning systems engineer at DeepSeek, published a lengthy post titled "I Have to Bury My Talent Yesterday" on Sept. 15, discussing how rapid advances in AI could reshape his career. He said he is unlikely to be unemployed but may need to change roles.

Liu, who joined DeepSeek in April 2025 to work on ML systems and kernel design, wrote the main attention kernel for DeepSeek v4.1. He expressed pride in the model's success but observed that AI is progressing quickly in kernel development, evolving from a documentation-searching assistant into an "operator master" that can read CUDA code, analyze instructions, and optimize kernels independently. He predicted that within six months to a year, AI-written kernels could match or surpass his work.

Despite this, Liu believes his experience, vision, and judgment will keep him at the table, but he may need to transition to avoid unemployment. He envisions becoming a "mech pilot" for AI agents, where instead of writing every line of code, he would orchestrate models, break down tasks, evaluate outputs, and take responsibility for the final system. "We have to say goodbye to the era of handwritten kernels and human optimization," he wrote.

His post adds to ongoing discussions about AI's impact on programming. A Sonar survey cited in the article shows AI-generated or AI-assisted code has risen to 42% of submitted code, up from 6% in 2023. Chinese officials have also said software talent structures must adapt to human-AI collaboration.