DeepSeek Engineer Says AI May Soon Replace the Work He Does, Creating a “Self-Destructive Loop”

A DeepSeek engineer has described a strange problem at the heart of the AI race. Liu Shengyu, who worked on key operators for the company’s V4.1 models, says AI has advanced so quickly that the technology he is helping build could soon perform parts of his own job as well as, or better than, he can.

DeepSeek engineer says AI is coming for his own job

In a public WeChat eassy titled “I Have No Choice but to Bury My Talent in Yesterday,” Liu reflected on how dramatically his work has changed in roughly a year. He said AI previously helped him search documentation and identify bugs, but has progressed to reading GPU code and optimizing software operators.

Liu said he is proud of his contribution to DeepSeek V4.1, but believes AI-generated operators could match or surpass his own work within six months to a year. His concern is not simply that he could lose his job. Instead, he expects the hands-on programming work he enjoys could increasingly be handled by AI agents.

This creates the “self-destructive loop” at the center of his argument. Engineers use AI to make AI development faster, which produces more capable models that can handle increasingly difficult engineering tasks. Those models can then accelerate the development of the next generation even further.

DeepSeek’s rapid model development provides a useful backdrop. The company released DeepSeek-V4.1-Flash on September 10, describing it as a more capable and efficient model with native multimodal understanding. The model has 552 billion parameters, although only a fraction are active for individual inputs and outputs.

Why Liu says he keeps building AI

Despite his concerns, Liu does not appear to believe that simply stopping his own work would solve the problem. His argument is that competing AI labs will continue advancing regardless, leaving individual engineers with little ability to halt the broader race.

He also raised concerns about who ultimately controls the most capable AI. Liu argued that frontier intelligence should remain broadly accessible rather than being controlled by a small number of companies. He specifically criticized proprietary approaches from OpenAI and Anthropic while pointing to DeepSeek’s open-source strategy.

His comments arrive as other AI researchers and executives are also debating recursive self-improvement, the idea that increasingly capable AI could contribute to improving future AI systems. Reuters reported this week that several industry figures have raised concerns about the possibility, although fully autonomous recursive self-improvement remains an unresolved question rather than an established reality.

For Liu, however, the immediate issue is more personal. He is helping build technology that could make the specialized programming skills he spent years developing less important. The question now is whether AI will simply become another tool for engineers, or whether the people building these systems will increasingly find themselves supervising machines that can do their own jobs.