DeepSeek Engineer Says AI May Match His Kernel Optimization Skills Within a Year

DeepSeek engineer Liu Sheng said AI is advancing fast enough that it could match or surpass his ability to write kernels within six months to a year. According to ChainCatcher, he wrote that DeepSeek V4.1's main Attention kernel was written by him, but AI has already progressed from helping him search documentation and debug code to reading CUDA, PTX, and SASS, analyzing instruction stalls, and independently optimizing kernels. Liu said he expects AI to next design scheduling plans, evaluate performance, and complete implementations on its own. He added that improving kernels at DeepSeek speeds up both training and inference, which in turn helps AI catch up faster, but said he will keep optimizing as long as other companies' models continue advancing. He said he is unlikely to lose his job, but may be forced to shift from writing kernels by hand to operating agents. Liu also said his main concern is not losing his livelihood, but possibly never again having the chance to do work he enjoys. He said he remains at DeepSeek because he believes frontier AI should be open and affordable for everyone, and said he does not believe Anthropic or OpenAI can achieve that. He warned that if the most powerful AI is controlled by only a few companies, the technology gap could widen into a gap in power and social class.