I Helped Build the AI That Will Replace Me

I have to bury my talent in yesterday

At 1 a.m. on September 14, Liu Shengyu typed the last line on his computer and hit publish.

Three thousand words. The title was like an epitaph.

He stood up, went to the kitchen, poured a glass of water. The cup was from the DeepSeek office, white ceramic, blue logo, a tea stain ring at the bottom he couldn't wash off. He carried it back to his desk. The screen was still on. The word count in the lower right corner: 3,174. He glanced at it, didn't revise, closed the editor, opened the terminal, and went back to running the day's operator performance tests.

Hours later, the article exploded in China's tech circle.

The hashtag "I have to bury my talent in yesterday" hit trending. It was translated into multiple languages and spread across overseas social media. People screenshotted it, wrote long reviews, argued in comments about whether it was self-pity. Liu Shengyu didn't respond much. He only added a note after the article, saying his intent was not to express anxiety about unemployment, but to say goodbye to the time when he wrote operators by hand.

Then he went back to writing operators.

1. The man holding up the computing foundation

Before understanding why this article resonated so widely, it's necessary to know what Liu Shengyu actually does.

Peking University Turing Class, class of 2021. Captain of PKU's Weiming Supercomputing Team. Represented the school at SC23. Got PhD offers from UC Berkeley and CMU, gave them up for financial reasons. Joined DeepSeek in April 2025.

A few days before publishing, he had just delivered the main Attention operator for DeepSeek V4.1. MQA attention with head dim = 512. A core component that directly determines model inference efficiency. Cui Tianyi, head of the DeepSeek harness team, called him "Operator Sage" in the comments.

Operator engineer. At the lowest level of GPU architecture. PTX assembly. SASS machine code. Hardware pipeline stalls. Register allocation. Shared memory scheduling.

Outsiders don't understand this. Insiders know how hard it is. It doesn't rely on memorization, on grinding problems, on calling libraries. It relies on feel. On staring at a latency curve in a profiling tool for three days until you suddenly know which bank conflict is the problem. On writing hundreds of lines of PTX, getting rejected by the hardware dozens of times, and on the last run, having it pass, with performance beating the vendor's official version by two percentage points. On intuition. On experience. On things that can't be written into documentation.

This field has long been considered one of the fortresses that human programmers are hardest to replace.

Liu Shengyu doesn't see it that way.

2. The weaver and the machine

He used a simple but precise metaphor in the article. A skilled weaver. Masterful hands. But everyone else is using machines. The joy of hand-weaving will eventually be crushed by the roar of the machine.

The metaphor stings because it doesn't paint AI as a distant threat. It paints it as a machine already at the door.

A year ago, AI was an assistant. Check docs. Read code. Find bugs. A year later, it can independently read CUDA, PTX, and SASS. It can use professional tools to analyze the stall time of every instruction. It can independently optimize operators.

Liu Shengyu's judgment: in another six months to a year, AI's ability to write operators will likely be as good as his. Or better.

He knows this because he himself is accelerating it.

"The better I write operators, the faster our new models train and infer, the faster model capabilities improve, and the sooner I will be replaced."

That line is in the article. He didn't add an exclamation mark. He didn't add "how ironic." He left it there. Like a stone.

Human beings have never hesitated when it comes to destroying themselves. He still tries his best to optimize operators for two reasons.

First, writing operators is like playing a game. It feels good. "When I see my operator performance far exceed the vendor's official version, I feel immense pride." He wrote that directly. You can imagine him at his desk, staring at the benchmark results, the corner of his mouth twitching upward. Then he screenshots it, sends it to a small group chat, with one line: "Not bad."

Second, colder. Even if he slacked off, deliberately sabotaged things, delayed model training, other companies' models would still develop as usual. "And in the end they will kill me just the same."

So he keeps writing. Keeps optimizing. Keeps pushing performance forward. Keeps accelerating the thing that will eventually replace him.

3. Why it went viral overseas

After the article came out, the reaction overseas was even bigger than at home.

People translated it. Shared it. Argued for three hundred comments on Hacker News. What they argued about wasn't "Will AI replace programmers?" That topic is too old. They argued about: Why is this person so honest?

For the past two years, the AI narrative has had only two voices. One from founders: disrupt everything, change the world, AGI next year. One from doomers: humanity is doomed, regulation must catch up, or we all die. Liu Shengyu stands on neither side. He didn't shout "AI is awesome." He didn't shout "AI is terrifying." He said: the future I helped accelerate is taking away the work I loved most.

A low-level engineer at a Chinese AI lab. Someone writing assembly at the lowest level of GPUs. Someone dealing with registers, shared memory, pipeline stalls every day. Suddenly he writes three thousand words about his craft disappearing.

Overseas readers hadn't seen this. They'd seen too many slide decks, too many launch events, too many "we believe AI will make the world better." They hadn't seen an engineer say: I know what I'm doing. I know what it will bring. I know what I will lose. But I did it anyway. Because if I didn't, someone else would. Because while I was doing it, I was happy.

There's also a passage about open source.

Liu Shengyu wrote: the most advanced intelligence should be supplied to everyone in an open, cheap way. He doesn't trust Anthropic or OpenAI to do that. "This is also why I chose, and insist on staying at, DeepSeek."

That passage was pulled out separately in overseas discussions. The focus had changed. Before, when readers looked at Chinese models, they only looked at cost-performance. Cheap, fast, usable. Now they're starting to look at the people. Their engineering ideals. Their value orientations.

A person who writes PTX, talking about open source. Talking about how intelligence should be distributed. Talking about who he doesn't trust. That, in itself, is more interesting than benchmark scores.

4. Unemployment or career change?

In his follow-up note, Liu Shengyu wrote: I am not trying to express anxiety about unemployment.

He said he won't be unemployed. But he will have to change careers. From craftsman to Agent mech pilot.

The rice bowl is still there. But "this may mean I will never again get to do the work I once loved."

He imagined the future. Writing operators by hand, even programming itself, will become entertainment. Like the javelin. No one uses a javelin to hunt anymore. But people still practice it. For competition. To see how far it can fly.

The javelin is still a javelin. The craft is still a craft. It's no longer needed for production.

This is deeper than unemployment. Unemployment is having no food. Career change is still having food, but watching your best dish get crossed off the menu. You can still go into the kitchen. You can still hold a spatula. But you know that the thing that kept you from wanting to leave at 3 a.m. is gone.

Liu Shengyu didn't write this article to cry. He wrote it to say goodbye. To that time of writing operators by hand. To that version of himself sitting at his desk, staring at a profile curve for three days, finally beating the vendor's official version.

5. The javelin no longer hunts

Technological history has always done this.

Assembly was replaced by C. C was replaced by high-level languages. Manual optimization was replaced by compilers. Every time the abstraction layer jumps up a level, a group of craftsmen lose their old positions. From a macro perspective, this is efficiency. Progress. Inevitability. From an individual perspective, it is a rupture in a generation's identity.

It's just that this time it's fast.

So fast that a person has just become an "Operator Sage" and already has to prepare to be a "mech pilot." So fast that "lifelong learning" is no longer a slogan but a survival necessity. So fast that after Liu Shengyu finished those three thousand words, the next day he still had to go back and run tests. Because the V4.1 operator was delivered, and V5 is still on the way.

He didn't stop.

At the end of the article, he didn't elevate. Didn't shout slogans. Didn't write "I believe humanity will eventually find new meaning." He put one sentence in the title:

I have to bury my talent in yesterday.

That sentence is not a question. Not an exclamation. It is a statement. Like a command he typed into the terminal. Enter. Execute. Continue.

6. After September 14

After September 14, Liu Shengyu's desk didn't change much.

Still the same computer. Still the white ceramic cup with the blue logo. The tea stain still at the bottom. He still runs tests every day, looks at profiles, adjusts registers, writes PTX. AI is still improving. Models are still iterating. He is still optimizing operators.

Only one thing changed.

He sent those three thousand words out. Someone saw them. Someone discussed them. Someone translated them. Someone remembered them.

What he buried was not his talent. It was the container his talent once clung to. The era of writing operators by hand is passing. He knows it. He helped it happen. He is a little sad. But he didn't stop.

He is still writing. Still running. Still pushing performance forward.

Like a javelin thrower who knows the javelin no longer hunts, but still stands behind the throwing line, runs up, turns, releases, and watches it fly.

Then walks over, picks it up, walks back.

Again.