Shiwen Dialogue Episode 17 | A DeepSeek Engineer's Confession: The Man Who Accelerated AI with His Own Hands Gave Himself a Six-Month Suspended Sentence

Opening: A 3,000-Character Essay, and a Man Who Wrote His Own Eulogy

In the early hours of September 14, Liu Shengyu, a machine learning systems engineer at DeepSeek, published a long essay of about 3,000 characters under a single-sentence title: “I Had to Bury My Talent in Yesterday.” He is the person in charge of DeepSeek V4.1’s core Attention operators—to put it plainly, part of how fast this model runs was written by his own hands. Within less than two days, starting September 15, the essay went viral in China and abroad. Zhihu saw a discussion question posted—“How do you view Liu Shengyu’s ‘I Had to Bury My Talent in Yesterday’?”—and tech media outlets ran follow-up after follow-up.

A man who accelerated AI with his own hands wrote his own eulogy. It sounds like clickbait, but the core judgment in his essay is stated calmly: within six months to a year, AI’s ability to write operators will catch up with, and even surpass, his own.

Shiwen: A top engineer publicly says his talent is about to become obsolete—what’s your first reaction?

Yongliang: The first thing I did was verify one thing—he isn’t venting after being laid off, and this isn’t a publicity stunt. He looked at what he does every day and reasoned his way, step by step, to his own endpoint. That chain of reasoning is worth hearing more than ten thousand takes on “will AI replace programmers.”

Shiwen: Then let’s talk it through: what exactly this essay says, why his words deserve more attention than the average person’s AI anxiety, how many layers are left in the moat of talent, and how to read that most controversial analogy.

Q1: First, the backstory—what kind of article can blow up like this?

Yongliang: Three factors stacked on top of each other: the speaker is solid enough, the statement is brutal enough, and the timing is perfect enough.

First, the person. Liu Shengyu is the lead for DeepSeek V4.1’s core Attention operators. Operators are the lifeline of a model’s inference speed, and only a handful of people in the industry can work in this position. Such people normally give no interviews and make no noise; when they finally speak, it’s 3,000 characters of eulogy for themselves—the contrast alone is news. Next, the statement. He didn’t complain about his company or the industry; the whole piece is an engineer’s farewell to his own craft—as media paraphrased it, he said, “I may not end up unemployed, but I have to change careers.” Farewells aren’t rare; what’s rare is one this lucid, complete with a timeline: six months to a year. Finally, the timing. AI writing code has been burning for over a year now, but the flames have only ever scorched application-layer programmers. This time, the fire reached the people who personally speed up the models—the flame made a full circle and burned the ones who fed it.

Q2: What exactly did he say? What do the six-month suspended sentence, the career change, and the “mecha pilot” mean?

Yongliang: He passed a six-month suspended sentence on this craft, then lined up his next stop: from the person writing operators to the person piloting Agents.

The core of the essay is a timetable: he believes that within six months to a year, AI’s ability to write operators will catch up with and even surpass his own. Writing operators used to be the craft of the very top sliver of chip engineers and model engineers—making matrix multiplication run faster within limited VRAM and power budgets, guided by intuition accumulated over years of experience. When the person doing this job himself judges that AI will reach his level within half a year, that signal carries a weight unlike any industry report out there: this is a diagnosis written by the terminally ill patient himself. According to media paraphrases, his exact wording was “I may not end up unemployed, but I have to change careers,” with the direction moving from designing, writing, and optimizing operators to becoming an Agent’s “mecha pilot”—that phrase, too, comes from the paraphrases. The meaning isn’t hard to grasp: value going forward lies not in building parts for the mecha, but in piloting it. On the evening of September 15, he also responded to the public attention on Zhihu—effectively adding another log to the fire with his own hands.

Q3: Why is a top engineer’s anxiety more worth listening to than the average person’s cries of AI anxiety?

Yongliang: Ordinary people’s anxiety is the fear of being hit by a car; his anxiety is discovering he’s the one building the engine—the closer you are to the endpoint, the more valuable your warning.

I’ve worked in software for seventeen years and have seen plenty of people crying wolf. But Liu Shengyu’s case is different, because he is one of the people who personally accelerated AI. AI writing operators is itself the doing of his cohort—the faster the model, the stronger AI gets at writing code and operators. It’s as if they were fitting parts into an engine while watching the speedometer climb. So he isn’t guessing at the endpoint; he’s reading the dashboard. What makes this perspective scarce: the vast majority of people discussing AI replacement base their judgment on gut feeling; his is based on every single line of his daily work. What deserves even more attention is how he handles the anxiety—no wailing, no jokes about switching careers to deliver food, but instead immediately starting to look for his next position. Seventeen years of leading teams taught me that the gap between people often lies not in who sees farther, but in who sets out first after seeing.

Q4: After seventeen years as a technology director, looking back—how many layers does the moat of talent have left?

Yongliang: My answer is three layers. This time he personally verified how fast the first two are falling, but nobody has lost the third yet.

The first layer is the tool layer: how to write code, how to tune operators, how to use frameworks. This layer is being eaten away at a visibly rapid pace—Liu Shengyu’s “six-month suspended sentence” was passed on exactly this layer. The second is the problem-definition layer: what the customer actually wants, where the system’s real bottleneck is, whether the thing is worth doing at all. AI can solve problems, but who poses the problem and why it needs solving is still decided by humans. The third is the responsibility layer: who is accountable when an incident happens, who signs off on medical data, who makes the final call on architecture. I’ve run technology for a hospital group—several hundred JVM services where registration, billing, and lab testing couldn’t stop for a second. When the system went down at 2 a.m., AI wasn’t going to answer the phone for me. So I understand Liu Shengyu’s choice to become a “mecha pilot”: it’s not fleeing technology, it’s moving from the first layer—the one being eaten—toward the second and third. The tool layer will be washed over again and again by wave after wave of technological change; the problem-definition layer and the responsibility layer are, for now, still in human hands. But note—only for now. He didn’t stop at that layer; he is moving toward a position closer to the steering wheel, and that very movement is itself the answer.

Q5: The most controversial analogy—“no less serious than letting Hitler get the atomic bomb before the Allies”—how should it be read?

Yongliang: First get the sourcing straight, then talk interpretation. My reading: this isn’t a political statement—it’s an engineer’s fear of “any single company leading forever.” You can disagree with him, but it’s worth understanding him.

First, let’s be precise about sourcing: the remark comes from his Zhihu response on the evening of September 15, as reported by Guancha.cn and other outlets; I’m citing it here only as paraphrased—he said, “I don’t trust Anthropic or OpenAI to do this, and I especially don’t want Anthropic to hold the most advanced AI or AGI—to put it in exaggerated terms, its severity would be no less than letting Hitler master atomic-bomb technology before the Allies.” Another line from the same response was paraphrased by media as: “Guess what—if Anthropic forever holds the world’s most advanced AI, will future society become communist, or 2077?” I won’t sensationalize this analogy or take sides for him, but two points are worth unpacking. First, this is closer to an engineer’s language of conviction than political language. Engineers believe that the greater the power, the greater the need for checks and balances; the object of his fear is “any one company permanently holding the strongest AI,” not any particular company itself—if a different company occupied that seat, his fear would most likely remain the same. Second, he aimed his criticism at both Anthropic and OpenAI, which shows that what he cares about isn’t a company narrative but the fact of “an unchecked, single-point lead.” This concern is debatable, but the precondition for debating it is admitting this: a person who personally pushed AI to this point has the standing to fear it resting in any one person’s hands. He also raised the issue of education—worrying that students using AI directly to complete homework and hands-on coursework would, over the long run, cause engineering ability to atrophy (per cryptocity’s paraphrase). That remark is really the same message as the atomic-bomb analogy: what unsettles him more than technological leadership is the disappearance, from the general population, of the ability to master technology.

Epilogue

Shiwen: Finally, sum up this episode in one sentence?

Yongliang: The six-month suspended sentence he passed on himself is really a wake-up call for everyone who makes a living by their craft—the length of the sentence may be off, but the suspended sentence itself is happening to each and every one of us.

Shiwen: That line goes out to everyone. See you next episode.

---

[Technical Deep Dive] What Is an Operator? Why Does “AI Writing Operators” Hit So Hard?

This episode’s throughline is “the man who accelerated AI with his own hands wrote his own eulogy.” For readers who want to go one level deeper, here’s a breakdown: what exactly an operator is, and why its fall carries such a weighty signal.

Layer one: operators are the lifeline of model speed. When a large model runs, it is, at its core, massive matrix computation. Operators are the concrete implementations of these computations—the same attention computation, written differently, can run several times faster or slower on the same chip. Half of the model experience is decided by the algorithms; the other half by the craft of operator engineers.

Layer two: it’s the craft with the highest density of experience. Writing operators requires understanding the temperament of the chip (VRAM bandwidth, caches, parallel units) as well as the structure of the model, and being deep enough on both ends. This kind of experience has historically been difficult to shortcut, making it one of the most classic “talent moat” positions.

Layer three: that’s why “caught up in six months” stings. If even the craft with the highest density of experience has only six months of shelf life left, then other experience-based technical roles will see shelf lives that get only shorter, never longer. This isn’t alarmism; it’s a valuation given by the people who understand the industry best—anyone can write a eulogy, but one you write for yourself is the one that counts.

Back to this episode’s throughline: talent in the tool layer is depreciating at an accelerating rate, while depreciation in the problem-definition and responsibility layers is far slower. Which layer to move toward is the real question this essay leaves for everyone.

---

Sources for This Episode

• Baidu Baike entry for “Liu Shengyu”; Tencent News report, 2026-09-16 (identity and role corroborated by both sources)
• Tencent News / Toutiao / TechWeb / WenxueCity: publication of the long essay (early hours of 2026-09-14) and its viral spread (from 2026-09-15)
• EET-China / TechWeb: paraphrased quotes “I may not end up unemployed, but I have to change careers” and “mecha pilot”
• Guancha.cn 2026-09-16, Tencent News: quotes from the Zhihu response (including the Hitler analogy and the communism/2077 remark—all paraphrased)
• x.com @J_Lingshan0909: paraphrase of the Zhihu response from the evening of 9/15
• cryptocity: concerns about education (paraphrased)
• Zhihu: discussion question “How do you view Liu Shengyu’s ‘I Had to Bury My Talent in Yesterday’?”