Interview: When AI learns to write GPU operators, what is left for engineers

On September 14, Liu Sheng, who leads the main Attention operator on DeepSeek V4.1, published an article titled "I Had to Bury My Talent Yesterday." It drew more than 100,000 reads on WeChat, climbed to No. 1 on Zhihu's trending list, and sparked discussion on X. We did not conduct a separate interview with him. However, his public writing reveals the reality of being a front-line engineer.

In the article, he describes the shift over the past year. AI began as a small helper that could look up documentation, read code, and hunt bugs. It has since become an "operator master" capable of independently reading CUDA, PTX, and SASS code, analyzing the stall time of every instruction using professional tools, and optimizing operators on its own. He predicts that soon AI will also design operator schedules, evaluate the performance of different options, and handle their implementation and optimization.

"Of course I hope I won't be made obsolete, but if it has to happen, I hope the person who does it is me."

He is proud of what V4.1 achieved, yet he is aware that the times are changing. He says that in another six months or a year, operators written by AI will likely be as good as his, or even better. He does not think he will become unemployed, but he will have to pivot. The job may remain, but the work he once loved may be gone.

Changes are also appearing at the application level. On September 15, a user on X posted that he used the DeepSeek V4.1 Flash model combined with the Hermes Agent tool to compress texture files in Resident Evil 7. This raised the frame rate on a OnePlus 12R from about 20FPS to about 30FPS. The AI identified I/O and memory bandwidth as the main performance bottlenecks and recommended compressing textures from 4,096 pixels to 1,024 or 512 pixels.

Viewed together, these two events point to the same conclusion: AI's capability in programming and engineering tasks is moving from assistance to primary execution. For developers, the question is no longer whether they can use AI, but how to position judgment, taste, and system design, the skills that are harder to replace, at the critical points of the workflow.