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 set off discussion on X. We did not get a separate interview with him. His public writing still shows what the job looks like for an engineer on the front line.

In the article he describes one year of change. AI started out as a small helper that could look up documentation, read code and hunt bugs. It became an "operator master" that reads CUDA, PTX and SASS on its own, uses professional tools to analyze the stall time of every instruction, and then optimizes operators by itself. He expects that before long AI will also design operator schedules, weigh the performance of different options, and implement and optimize them.

"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, and he also knows which way the wheel is turning. In another six months or a year, he says, operators written by AI will probably be as good as his, or better. He does not think he will lose his job, but he will have to change what he does. The paycheck may survive; the work he once loved may not.

Change is showing up in applications too. On September 15, a user on X posted that he called the DeepSeek V4.1 Flash model together with the Hermes Agent tool to compress the texture files of Resident Evil 7, raising the frame rate on a OnePlus 12R from about 20FPS to about 30FPS. The AI found that I/O and memory bandwidth were the bottleneck and recommended compressing textures from 4096 pixels to 1024 or 512 pixels.

Put the two together and they point the same way: AI's ability on programming and engineering tasks is moving from assistant to main force. For developers the question is no longer whether they can use AI, but how to place judgment, taste and system design, the skills that are harder to replace, at the key points of the process.