(Yicai) Sept. 16 -- Shengyu Liu, who developed a core component of DeepSeek’s newly released V4.1 artificial intelligence model, expects AI to catch up with and possibly surpass the quality of his own work in a matter of months, highlighting how AI is moving beyond routine software tasks and into the specialized engineering work that powers the technology itself.

Liu, a machine learning systems engineer at the Hangzhou-based AI unicorn, said in an article posted on his personal WeChat account recently that he probably will not become unemployed because of AI, but he may well be forced to "change professions.”

Liu was a member of Peking University’s 2021 Turing Class, an elite undergraduate cohort in computer science, and previously captained the university’s supercomputing team, leading it to top-two finishes in international university supercomputing competitions.

Just days ago, Liu completed the main attention operator for DeepSeek’s V4.1, a core component that determines the model’s inference efficiency. But within six months to a year, AI-written operators will probably match or even outstrip his own abilities, he admitted.

Liu expects that he may eventually have to give up the operator design and manual optimization he loves to become a “mech pilot” of AI agents, directing models, breaking down tasks, evaluating results, and taking responsibility for the final system.

Liu described the shift as a “technological paradox.” “The better operators I write, the faster our new models train and run inference, and the more rapidly model capabilities advance, meaning I get replaced sooner,” Liu pointed out.

“AI can think of 300 tokens in one second, write a command in half a second, and finish writing code in 20 seconds, which I can’t," Liu said.

Unlike humans, AI can also continue to scale its reasoning, tool use, and ability to perform tasks in parallel, he said. It has evolved from an assistant that merely searches documentation and reads code into one capable of independently analyzing and optimizing operators using Compute Unified Device Architecture, Nvidia’s programming platform for graphics processing units, Liu said.

Liu’s concern reflect a broader shift as AI programming tools become embedded in developers’ daily work. According to the 2026 Developer Survey report released by code quality platform Sonar, 72 percent of developers who have used AI programming tools use them every day, while 42 percent of code submitted by respondents is AI-generated or heavily AI-assisted, up from just 6 percent in 2023.

AI is also changing the skills employers look for in programmers, according to the China Software Industry Association. At a press conference held by China’s industry and information technology ministry last week, CSIA Secretary-General Lv Weifeng addressed the debate over whether AI will replace programmers.

Pure coding skills will become less important, while understanding requirements, designing architectures, integrating systems, and ensuring quality and safety will become more important, he said. Lv added that collaborating with AI is becoming a core skill for programmers, requiring them to clearly communicate tasks, evaluate AI-generated output, and coordinate multiple AI agents.

Editor: Emmi Laine