DeepSeek machine-learning systems engineer Liu says AI agents could match or exceed human-written performance operators within a year, marking a rapid change for one of the most specialized jobs in AI infrastructure: hand-optimizing the code that makes large models run efficiently on chips.
Liu made the prediction in a Sept. 14 essay published after he delivered the main Attention operator for DeepSeek V4.1. The work covered a “head dim = 512” multi-query attention, or MQA, operator, a component used in the attention calculations that help a model decide which information in a sequence deserves focus.
His post spread through Chinese technical communities and reached No. 1 on Zhihu’s trending list, according to screenshots included with the article. In a follow-up post on Sept. 15, Liu said the essay was intended as a farewell to an era in which such operators were built entirely by hand, rather than a warning that engineers would imminently lose their jobs.
Ai agents move into performance-critical code
Performance operators sit close to the hardware layer of AI systems. Engineers write and tune them to make mathematically intensive model operations run faster and use less memory on accelerators. Small improvements can compound across the training and inference of large models, where the same calculations are executed repeatedly at scale.
Liu argued that AI systems are approaching the point where they can take over much of that work. He wrote that, within “half a year or a year,” AI-written operators could become comparable to or better than those produced manually by engineers.
The comparison, in Liu’s telling, comes down to the speed at which an AI agent can cycle through ideas, code changes, tests and revisions. He wrote that AI can “think 300 tokens in one second,” “type one command in half a second,” and “finish a piece of code in 20 seconds.”
Those figures describe an emerging development workflow rather than a benchmark for a particular model. An agent still needs a target, access to compilers and test environments, and ways to judge whether a generated operator is correct and genuinely faster. Yet Liu’s point is that once those systems are connected, machine-generated code can iterate more frequently than a human engineer working through the same optimization loop.
That places experienced systems engineers in a changing role. Instead of personally writing every implementation and manually adjusting low-level performance details, Liu expects more of their work to involve directing AI agents, setting constraints, evaluating outputs and deciding which approaches are safe to deploy.
“The demand” may move from engineers who can write high-performance operators themselves to people who can use AI “to produce high-performance operators faster,” Liu wrote.
A self-reinforcing development cycle
Liu described a feedback loop that could make the transition occur faster inside frontier AI labs. Better operators reduce the computing cost or time needed to train and run models. Faster training and inference can support quicker experimentation and capability improvements. More capable models, in turn, become better at producing and optimizing code.
The loop does not eliminate engineering judgment. High-performance AI code must often account for memory layout, parallel computation, numerical precision and hardware-specific behavior. An operator that runs quickly but produces incorrect outputs, consumes too much memory or fails under production workloads has limited value.
Liu’s account instead suggests that the bottleneck is shifting. The scarce skill may increasingly be the ability to create rigorous test conditions, identify the highest-value optimization target, supply an agent with useful feedback and recognize errors that automated evaluations miss.
His own work reflects the level of specialization involved. Liu studied in Peking University’s 2021 computer science “Turing Class,” served as captain of the university’s Weiming supercomputing team and participated in the SC23 International Student Supercomputing Competition. He joined DeepSeek in April 2025.
In a footnote to the essay, Liu clarified that “main Attention” referred specifically to the head-dimension-512 MQA attention operation. It did not include an “indexer” that selects the top-k most important tokens, which he said had been written by other DeepSeek colleagues and their AI agents.
Student training may face a harder problem
Liu also focused on a less visible consequence: the effect of AI-assisted coding on engineering education. He described a programming assignment that might require eight hours of manual work but could be completed by an AI model in minutes for only “a few jiao,” a small Chinese currency amount.
That gap creates an incentive for students to submit working code without building the practical understanding that comes from debugging, testing and revising their own implementations. Liu warned that weaker engineering fundamentals combined with powerful AI tools could multiply the production of “spaghetti code” — tangled, difficult-to-maintain software with poorly understood dependencies.
The concern is especially relevant for systems code, where a flawed optimization may cause errors that are expensive to isolate. AI-generated code can pass a narrow test while breaking under a different hardware configuration, a longer input, heavier workloads or unusual numerical conditions. Faster production therefore increases the value of review systems and test suites that can catch failures before code reaches core infrastructure.
Liu did not frame the shift as a case for abandoning programming education. His argument points toward training engineers to understand the code agents produce, challenge their assumptions and operate automated development systems without becoming dependent on them.
A farewell to handwritten operators
After readers interpreted the original essay as a statement of employment anxiety or a criticism of closed-source AI institutions, Liu issued a clarification the next day. He said neither was his intent.
Liu wrote that “handwritten operators,” and potentially programming itself, could eventually become closer to entertainment than production. He compared that possibility with javelin throwing: a skill once connected to hunting that now survives mainly as sport.
He also said a version of the essay posted on WeChat had received more than 100,000 reads and again pointed to its top position on Zhihu’s hot list, along with discussion on X.
Liu said he plans to remain at DeepSeek and continue working with AI-agent workflows for operator development. His conclusion is less about the disappearance of engineers than about the disappearance of a particular kind of engineering task: spending long periods hand-writing every low-level optimization before software can begin testing alternatives.
Explore how AI copy trading mirrors agent-driven automation, shifting human roles toward supervising intelligent financial systems.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
