Liu Shengyu
A researcher at DeepSeek specializing in low-level GPU kernel optimization. He participated in the development of core Attention operators for DeepSeek large language models and holds deep expertise in low-level instruction set optimization, including CUDA PTX and SASS assembly.
Timeline (last 90 days)
Liu Shengyu graduated from Peking University's Turing Class and led his team to win the championship at the World Supercomputing Competition.
Liu Shengyu has publicly acknowledged that he is being pushed to the brink of obsolescence by the AI he personally trained.
AI is currently capable of reading low-level assembly code such as CUDA PTX and SASS, and can independently optimize operators.
Liu Shengyu believes AI may push the world toward two extremes: one where productivity is completely liberated and the technological dividend benefits everyone, and the other where a small number of tech companies controlling the most advanced AI form class barriers.
Liu Shengyu is a DeepSeek researcher who participated in writing the main Attention operator for DeepSeek V4.1.
Liu Shengyu believes he will in all likelihood not lose his job, but may lose his former passion, transitioning from someone who writes operators to a coordinator who directs agents.
Liu Shengyu chose to stay at DeepSeek to research powerful, fast, and universally accessible artificial intelligence and open-source it.
Liu Shengyu predicts that in another six months to a year, the level at which AI writes operators will in all likelihood surpass his own.
All Facts (8)
Liu Shengyu is a DeepSeek researcher who participated in writing the main Attention operator for DeepSeek V4.1.
60%UnverifiedAI is currently capable of reading low-level assembly code such as CUDA PTX and SASS, and can independently optimize operators.
60%UnverifiedLiu Shengyu believes AI may push the world toward two extremes: one where productivity is completely liberated and the technological dividend benefits everyone, and the other where a small number of tech companies controlling the most advanced AI form class barriers.
60%UnverifiedLiu Shengyu chose to stay at DeepSeek to research powerful, fast, and universally accessible artificial intelligence and open-source it.
60%UnverifiedLiu Shengyu has publicly acknowledged that he is being pushed to the brink of obsolescence by the AI he personally trained.
50%UnverifiedLiu Shengyu graduated from Peking University's Turing Class and led his team to win the championship at the World Supercomputing Competition.
50%UnverifiedLiu Shengyu believes he will in all likelihood not lose his job, but may lose his former passion, transitioning from someone who writes operators to a coordinator who directs agents.
50%UnverifiedLiu Shengyu predicts that in another six months to a year, the level at which AI writes operators will in all likelihood surpass his own.
60%