DeepSeek kernel engineer says AI he built is retiring his craft — and he's speeding it up
On Sept 15, DeepSeek kernel engineer Liu Shengyu (interestingLSY/intlsy), who says he wrote core attention kernels for DeepSeek v4.1, published the essay "I Have to Bury My Talent in Yesterday," widely shared across the AI community and translated into English by teortaxesTex. He writes from the frontline that the AI he builds is retiring the craft he loves — and he chooses to embrace and personally accelerate that process.
Confirmed
- Identity: former captain of Peking University's supercomputing team, which took first and second place globally; per TheTuringPost, he is the main author of the attention kernel in the newly released DeepSeek-V4.1-Flash and heavily involved in the DeepEP V2 expert-parallel communication library.
- The essay has two parts: a personal confession that his beloved hand-crafted CUDA/kernel optimization is becoming obsolete due to AI progress, and his endorsement of the trend, arguing he should keep accelerating it. Per TheTuringPost, he also explains he stayed away from Anthropic because he "doesn't want AGI to fall into their hands."
- He discusses recursive self-improvement (RSI), citing AI's pace: about two years from ChatGPT to o1/R1 reasoning models, and even less from reasoning models to fluent tool-using agents.
- Per minzhu0305's retelling, in just one year AI went from an assistant that could only look up docs, read code and find bugs to one that can independently read CUDA/PTX.
- Another concrete prediction: within a year, AI-written kernels will beat his own — and if upfront compute is no object, this may already be true today. He observes this kind of work is nearing the end of its abstraction level.
Unconfirmed
- Multiple reposts mention "DeepSeek v4.1" with the poster noting "sic"; the version number's accuracy rests with the original blog.
- Personal details such as the Anthropic remark come via TheTuringPost's secondhand retelling, not directly from the original post.
Why it matters
- This is a rare first-person confession from a frontline AI infrastructure engineer that the system he built is displacing his own craft — combining emotional impact with technical judgment — making it a landmark text in debates over RSI and AI's impact on highly skilled labor.
2026-09-15 ~ 2026-09-17 · 9 related posts
- Episode 1: DeepSeek kernel engineer says AI he built is retiring his craft — and he's speeding it up(2026-09-15, 9 posts)
- Episode 2: DeepSeek Kernel Engineer's Essay on AI Risk and Staying On(2026-09-17, 2 posts)
Primary sources
- [source] DeepSeek kernel engineer blogs on his craft becoming obsolete — teortaxesTex · 2026-09-15
- [source] DeepSeek kernel engineer's poignant essay on AI making his own craft obsolete — teortaxesTex · 2026-09-15
- DeepSeek kernel engineer pens striking post on AI making his craft obsolete — StewartalsopIII · 2026-09-15
- DeepSeek engineer who wrote v4.1's core Attention operator reflects on building his own replacement — WebAssemblyMan · 2026-09-15
- Former PKU Supercomputing Captain, Twice Global Top-2, Joins DeepSeek Ranks — teortaxesTex · 2026-09-15
- Peking University supercomputing ex-captain: AI will write kernels better than him within a year — teortaxesTex · 2026-09-15
- DeepSeek kernel engineer's blog: why he stayed instead of joining Anthropic — TheTuringPost · 2026-09-17
2 near-duplicate retellings: teortaxesTex · minzhu0305