A DeepSeek engineer who helps make frontier models run faster says AI is beginning to perform the specialised work he loves. Liu Shengyu’s account matters because it describes displacement not as a distant labour forecast, but as a change already reaching the people building the technology.
DeepSeek Watch analysis
What happened
Liu wrote on his public WeChat account that, in roughly a year, AI had advanced from helping him search documentation and find bugs to reading GPU code and optimising software operators. He said AI-written operators will probably become as good as his own or surpass them.
He does not expect to lose his job. His fear is subtler: that kernel engineers will move from practising their craft to steering agents that do the hands-on work. Business Insider reports that Liu confirmed he wrote the post, but declined further comment.
The competitive bind runs through his account. Liu argued that refusing to improve kernels would not halt progress because rival laboratories would continue, leaving him displaced anyway. He also called for cutting-edge AI to remain open and affordable rather than controlled by a handful of companies.
Why it matters
This is a useful view from inside the machinery. AI’s effect on skilled work may not begin with mass redundancy. It may first remove the technically demanding parts people find satisfying, while preserving a thinner supervisory role around them.
Liu’s experience cannot establish how broadly that pattern will spread, and his forecast remains personal opinion. It does, however, give the automation debate a concrete unit of change: an engineer moving from writing optimised code to overseeing systems that write it.
Our read
Liu has captured the peculiar bargain at the frontier: build a better tool, then watch it reach for your favourite part of the job. His post is neither a DeepSeek announcement nor proof that human kernel engineering is finished. It is a credible warning that productivity can rise while professional fulfilment quietly takes the lift down.
Teams adopting coding agents should therefore measure more than output. They should decide which skills people must retain, where human review remains technically meaningful and whether supposedly liberated workers are gaining better problems or merely more dashboards.
What to watch
- Whether DeepSeek publicly addresses how AI-assisted coding is changing its engineering roles.
- Whether agent-written kernels consistently match expert work across hardware and workloads.
- How frontier laboratories preserve human expertise needed to audit generated low-level code.
- Whether open access to capable models broadens opportunity or chiefly accelerates the same competitive pressure.
Discussion spark: If AI can outperform specialists at the part of their work they most enjoy, what should employers preserve besides productivity?
Sources and evidence
not affiliated with or endorsed by DeepSeek