Last night, Liu Shengyu, an operator engineer at DeepSeek, posted a lengthy article that quickly became a trending topic today, accompanied by tags such as 'Even Top Engineers Feel AI Anxiety' and 'The Imperative for Career Transition'.

The top-rated comment in the discussion section was a self-deprecating quip: 'Are we joining the losing team now?'

However, the most striking sentence from his original article, which received little attention in reposts, was: 'An individual with subpar engineering skills, when paired with AI, can churn out shoddy code several times more efficiently than before, introducing various hidden risks into the system and making the world even more chaotic.'

While everyone is discussing how AI is breaking barriers, what Liu Shengyu truly worries about is the rising floor. AI elevates not skill levels but output volumes.

These are two fundamentally different phenomena.

First, let's delve into what he is relinquishing.

To grasp the significance of this article, we must first understand his background.

Liu Shengyu hails from the 2021 Peking University Turing Class, served as the captain of the Weiming Supercomputing Team, represented his university in the international supercomputing competition SC23, and joined DeepSeek during his undergraduate studies. In February 2025, during the open-source release of the DeepGEMM operator library, he played a pivotal role in optimizing the underlying FP8 GEMM operators. This April, during the upgrade of the DeepEP V2 expert parallel communication library, he participated in reconstructing the communication operators. This month, he independently oversaw and implemented the main Attention operator for V4.1, specifically the MQA attention with a head dimension of 512.

There is a consensus in the industry that this represents one of the most challenging skills for human programmers to be replaced. It requires not only a deep understanding of high-performance computing logic but also a profound grasp of GPU microarchitecture, memory hierarchy, and even down to the PTX assembly and SASS machine code levels to analyze hardware pipeline stalls and perform register allocation and shared memory scheduling.

This cannot be compensated for by mere hard work; it demands innate talent and years of experience.

Liu Shengyu predicts that within the next six months to a year, AI's ability to autonomously design and implement end-to-end extreme operators will likely catch up with or even surpass that of top human engineers.

His reasoning is not based on the model's intelligence but on a set of simple comparisons: AI can process 300 tokens per second, type out a command in half a second, and write a piece of code in twenty seconds, whereas he cannot. AI can continuously enhance its thinking intensity, tool invocation frequency, and parallelism, whereas he cannot.

This is not a competition of abilities but of physical limitations. Humans simply cannot keep up.

He acknowledges a vicious cycle: the more he optimizes operators, the more efficient model training and inference become, leading to faster iterations and quicker obsolescence of his skills.

Yet, he continues to write. His original words were: 'Of course, I hope I won't be revolutionized, but if I must be, I hope the one revolutionizing me is myself.'

By the way, after the article was published, some suspected it was written by a large language model. He responded succinctly: 'I wrote this by hand.'

He is not talking about unemployment but career transition

The hot search misinterpreted 'career transition' as a euphemism for unemployment, which is inaccurate.

His point is that top engineers can still remain competitive with their vision, engineering judgment, and intelligence, but the paradigm for producing operators has fundamentally changed. He has coined a term for this new role: the 'mech pilot' of agents. Instead of writing low-level instructions line by line, individuals step back and, based on their understanding of model requirements, cognition of hardware physical characteristics, and grasp of the overall system, give instructions to agents, set boundaries, and evaluate plans.

The job is secure. But he remarked lightly: 'I have more gears in my hands, but fewer beats in my heart.'

I read this sentence twice. It is not a complaint about treatment; it is about the broken connection between a person and their work. Previously, he would spend an afternoon profiling bit by bit, watching the throughput rise grid by grid, and the sense of accomplishment when outperforming the official operators from chip giants was, in his own words, like a speedrunner breaking a world record. Now, he describes the requirements, and the machine executes.

He wrote about this part very restrainedly, without sentimentality. But it is indeed the only part of the article with warmth.

The Other Half of the Chaotic World

Up to this point, it is still a record of an elite's predicament. The entire internet has only seen this layer.

But his article takes a turn in the latter half, which is why it deserves careful reading.

He poses a question that students face daily: the most common temptation in university and research training now is to take shortcuts with AI. However, code organization, system construction, forward-looking design, and abstraction skills were originally honed through countless collisions with walls and debugging sessions.

So, what will happen to these abilities? He provides two contrasts: will they slowly marginalize like 'writing x86 assembly,' or will they always be the cornerstone like 'understanding the entire computer system'?

If it is the latter, then AI brings not technological prosperity but systemic risks.

Because there is an asymmetry here: AI is a lever for both the strong and the weak, but it amplifies different things.

For someone like Liu Shengyu, AI amplifies his output while compressing his scarcity; he is being caught up with.

For someone who originally could not grasp system boundaries, AI amplifies their output but does not amplify their judgment. Previously, the messy code structure, random naming, and unhandled boundaries were obvious to anyone with a discerning eye. Now, the variable naming is standardized, comments are complete, and the structure is neat, but the errors remain the same; the alarm has been dismantled. AI does not eliminate garbage; it eliminates the appearance of garbage.

These two levers are at work simultaneously. The result is that the scarcity of high-quality output in the system is declining, while the absolute amount of low-quality output is surging, and the two are becoming increasingly difficult to distinguish externally.

This is what he means by 'even more chaotic.'

This judgment is not just his intuition. The article mentions an experiment conducted by NVIDIA and Cursor where a multi-agent system automatically optimized 235 CUDA operators for the Blackwell B200 over three weeks, achieving a geometric average performance improvement of 38%, with 19% of tasks speeding up more than twofold. When OpenAI was advancing its custom chips, it directly called upon agent tools to participate in operator generation and compilation stack construction, compressing tasks that originally required dozens of senior low-level experts and months of time into an extremely short period.

A reader named 'Hulu' in the comment section put it more directly. He said he is using Codex to develop Attention operators for domestic chips, and except for the inability to finalize a plan for him, everything else can be done quickly and well.

Note the only exception in that sentence: the inability to finalize a plan for him.

So, this is why he asks that question

At the end of the article, he poses a question: 'In future society, will power be more important than technology or intelligence?'

This question logically follows the previous reasoning.

In the past, ability itself was the sieve. If you couldn't write that operator, you couldn't get in; whether your words carried weight depended on whether you could accomplish what others couldn't. Technology and intelligence were the qualifications for entry and the source of discourse power (influence).

When ability becomes cheap and everyone can produce something that looks qualified, this sieve fails. So, what becomes the new sieve?

It is who has the authority to finalize a plan. Who decides what gets released, who bears the consequences if something goes wrong, and who has the power in the system to say, 'This won't work; do it again.'

This is not about ability; it is about position. And the more troublesome issue is that the vacancy in this position is not unfilled; it is that there is no system in place to cultivate people for this direction. Computer education tests whether you can write code. Interviews test whether you can implement an algorithm in 45 minutes. Performance evaluations count how many deliverables you have. Promotions look at what you have accomplished. From education to employment to incentives, the entire chain is designed around the already devalued metric of 'production capacity.'

His answer is to stay at DeepSeek, reasoning that he believes the most cutting-edge intelligence should be provided to everyone in an open and affordable manner. He even says that this might pull the world back a bit from 'the brink of 2077.' You might think this is overly idealistic, but at least it aligns with the previous reasoning. He is worried about the concentration of power, so he chooses to stand on the side that disperses ability.

A person who is being replaced by what he himself created is ultimately concerned about this. The hot search labels him as having 'AI anxiety.'

So, where should we direct our thoughts?

The question 'Will AI replace me?' has been asked for two years, yielding nothing useful. The answer is approaching yes; the only difference is the order.

Liu Shengyu's article provides a better question: under the lever of AI, am I among those being caught up with or those being amplified?

Those being caught up with lose their scarcity but retain their judgment; they still know what is good, but good things are no longer exclusive to them. Such individuals can transition and still stay in the game.

Those being amplified are more dangerous because their output increases, but feedback slows down. The things they produce look increasingly presentable, making it harder for them to realize where the problems lie. By the time issues are exposed, they are often already embedded in the system.

The dividing line between these two groups does not lie in whether they can use AI but in whether they possess the ability to 'finalize a plan,' knowing what should be done, what should not be done, and what will cause problems if done.

And this ability is precisely honed through collisions with walls and debugging. It is precisely the part that AI most easily helps you skip.

This is what Liu Shengyu is truly warning about. He is not saying he will lose his job. He is saying that when everyone can be highly productive, the ability to say, 'No, this cannot be released,' will become the scarcest and most critical thing. And we are collectively skipping the step of practicing it.

He buries his talent in yesterday. Meanwhile, many more people never had the chance to possess that talent in the first place yet are already highly productive.