Shengyu Liu, a DeepSeek engineer who wrote attention kernels for its v4.1 model, predicted that AI would probably match or surpass his kernel coding within six months to a year. In an essay, Liu described AI progressing in just a year from helping him read documentation and find bugs to independently analyzing and optimizing kernels, code whose performance affects the speed of model training and inference.

Liu anticipates shifting from writing kernels himself to directing AI agents rather than losing his livelihood. “I probably will not lose my job, but I will have to change what I do,” he wrote. He said he continues optimizing kernels despite helping accelerate his own replacement because competing companies would keep advancing even if he stopped.

Liu also argued that the most advanced AI should be openly and affordably available to everyone, warning that concentrated access could entrench inequality. “I do not trust Anthropic or OpenAI to do that,” he wrote. He singled out the prospect of Anthropic controlling the most advanced AI, saying, “To put it dramatically, I think the stakes would be comparable to Hitler obtaining the atomic bomb before the Allies did.” Liu cited DeepSeek’s work on accessible AI and its open sourcing of models as his reason for staying at the company.