A viral essay from a DeepSeek machine learning systems engineer is reframing the AI-and-jobs debate in a way worth taking seriously: not "will AI take my job," but "will AI end the specific craft I built my identity around, even if my job survives."
Liu Shengyu, credited with delivering DeepSeek's flagship MQA attention kernel and a former Peking University Turing Class student and supercomputing team captain, published an essay — widely translated and shared under the title roughly "I Have to Bury My Talent in Yesterday" — predicting that AI will match or surpass top human kernel engineers within six to twelve months. His stated conclusion isn't fear of unemployment. It's grief for a specific kind of hands-on, creative optimization work that he expects to no longer exist as a human occupation in any meaningful sense.
This adds a sharply specific data point to a debate explainx.ai has tracked closely all year — not whether AI displaces developers broadly, but what happens to expertise and judgment when the hardest, most technically demanding layer of the work gets automated first.
TL;DR
| Question | Answer |
|---|---|
| Who wrote the essay? | Liu Shengyu, a DeepSeek machine learning systems engineer, credited with the MQA attention kernel |
| What's the core claim? | AI will match or surpass top human kernel engineers within 6-12 months |
| Is this about job loss? | No — explicitly not. Liu expects to keep working, likely shifting toward directing AI agents |
| What is he actually mourning? | The hands-on, creative craft of kernel optimization itself — not the paycheck |
| What's kernel engineering? | Writing low-level code controlling exactly how model computations run on GPU hardware — extremely specialized, performance-critical work |
| Why does this matter beyond DeepSeek? | If AI can automate the most technically demanding layer of ML engineering, it's a stronger signal than automating higher-level application code |
Why kernel engineering specifically is the canary here
Kernel engineering sits near the bottom of the software stack that makes large models run efficiently — writing the code that directly controls memory access patterns, compute scheduling, and hardware-specific optimizations on GPUs. It's widely regarded as one of the most demanding, craft-intensive specializations in ML engineering: correctness matters, but so does an almost artisanal feel for squeezing extra performance out of hardware through techniques that don't reduce to a formula. Liu's own credential — delivering the MQA attention kernel that likely shaped DeepSeek's real-world efficiency — puts him at the top tier of people qualified to make this specific claim credibly.
That's precisely why this essay lands differently than the broader "AI will replace junior developers" narrative that's circulated for years. If the argument is that AI automates a senior kernel engineer's most specialized, hardest-won expertise within a year, it's a much stronger claim about the trajectory of AI capability than automating boilerplate application code — the layer of software work that's always been most exposed to automation first.
The distinction that matters: job loss vs. craft erosion
Liu's essay is explicit that he doesn't expect to lose his job — he expects to change what his job is. His framing draws an analogy to a traditional hand-knitter forced to abandon hand-crafting in favor of industrial machinery: the knitter doesn't necessarily become unemployed, but the specific skill, judgment, and creative satisfaction that defined the craft gets displaced by a machine that does the mechanical part faster and, eventually, better. The knitter's remaining value shifts toward operating, directing, or designing for the machine — not competing with it at the craft itself.
Applied to his own field, Liu's suggested path forward is pivoting toward orchestrating AI agents rather than hand-optimizing kernels directly — supervising, directing, and catching the mistakes of automated systems rather than doing the underlying creative work personally. That's a specific, credible version of the "move up a layer of abstraction" advice that's become common across the software industry, but coming from someone whose stated expertise is precisely the layer being automated gives it more weight than the same advice from a commentator with no comparable specialization.
Why this resonates beyond one engineer's essay
This essay went viral for a reason: it names something a lot of highly skilled technical workers have been feeling without a clean way to articulate it. The broader anxiety in software engineering this year hasn't primarily been "will I be unemployed" — unemployment fears are real but often overstated relative to how fast demand for software itself is growing. The sharper, quieter anxiety has been closer to Liu's framing: will the specific expertise I spent years building still matter, or will it become a historical curiosity the way hand-calculating logarithm tables became after calculators. explainx.ai's coverage of AI-driven de-skilling has tracked a related but distinct version of this concern — junior developers who never build debugging intuition because AI handles the hard parts from day one. Liu's essay is the senior-engineer mirror image: someone who already has deep expertise, watching the ground shift under a skill they've already mastered.
What this means if you're building a technical career right now
- Don't assume "senior" or "specialized" makes a skill automation-proof. Liu's essay is a direct counterexample to the assumption that only junior, repetitive work is exposed — the most specialized, craft-intensive work can be exactly what AI systems are optimized hardest to replicate, because it's high-value and well-documented in training data and benchmarks.
- The orchestration/judgment layer is where Liu (and much of the industry) expects durable value to concentrate. Directing AI agents, catching their failures, and making judgment calls about what to build — rather than personally executing the most automatable parts of the craft — is a consistent theme across multiple fields facing similar disruption this year.
- Grief is a legitimate response worth naming, not just optimizing past. Liu's essay resonates because it doesn't pretend the emotional cost of watching a mastered craft become obsolete is irrelevant just because a career survives it. That's worth acknowledging honestly rather than only offering "just learn to prompt better" as a response.
How China's discourse on this differs from the US framing
One detail worth surfacing from the broader coverage of this essay: Chinese commentary on AI-driven job disruption has tended to frame the conversation less around dystopian "Matrix"-style displacement narratives common in US discourse, and more around practical, immediate questions of skill adaptation and career pivoting — closer to Liu's own framing than to the more apocalyptic tone that sometimes surrounds similar discussions in Western tech commentary. That's a generalization with obvious exceptions, but it's consistent with how Liu's essay itself is structured: methodical, specific about timelines (6-12 months), and focused on a concrete adaptation strategy (shift toward orchestrating agents) rather than open-ended alarm. Whether that reflects a genuinely different cultural posture toward automation or simply this particular essay's tone is hard to say from one data point, but it's a useful reminder that the global conversation about AI and jobs isn't monolithic — different technical communities are processing the same underlying shift in noticeably different registers.
The kernel-engineering timeline is itself a claim worth tracking
Liu's 6-12 month prediction is specific enough to be checked against reality relatively soon, which makes it a good test case for how seriously to weight predictions from practitioners embedded in frontier labs versus more general industry commentary. If AI genuinely matches or exceeds top human kernel engineers within that window, it would be a concrete, checkable milestone — unlike vaguer claims about AI "transforming" software engineering broadly, which are harder to falsify. Worth revisiting this specific claim around mid-2027 to see whether it held.
FAQ
Who is Liu Shengyu? A machine learning systems engineer at DeepSeek, former Peking University Turing Class student and supercomputing team captain, credited with the flagship MQA attention kernel.
What did his essay actually argue? That AI will match or surpass top human kernel engineers within 6-12 months, ending the hands-on craft of kernel optimization — not that he'll lose his job.
Is this the same as saying AI will cause mass job loss? No — Liu explicitly frames it as skill erosion and identity loss, expecting to shift toward directing AI agents rather than becoming unemployed.
What is kernel engineering, and why does automating it matter? Writing low-level, hardware-specific code that controls model computation performance — automating it signals AI has reached even the most technically demanding layers of ML engineering.
How does this compare to broader "de-skilling" concerns in software? It's a senior-engineer version of the same concern already playing out with junior "vibe coding" developers — skill erosion from over-reliance on AI for the hardest parts of a job.
What does Liu suggest engineers do in response? Pivot toward orchestrating and directing AI agents rather than competing with automation at the craft itself.
Related reading
- AI-driven de-skilling: why vibe coding is producing developers who cannot debug
- Cognitive debt: what happens when you let LLMs retype your code
- Does AI make you dumb? The research, explained
- Your job in 2027: AI transformation in every domain
- What is an agent harness? Complete guide
- TypeSafe AI's Jev: a "System One Model" for structured decisions
- Official: coverage via HyperAI and ChinaTalk
Details in this piece reflect translated coverage of Liu Shengyu's essay available as of September 16, 2026. Direct English-language sourcing is limited; treat direct quotes in secondary coverage as paraphrased translations rather than verified original wording.