"I certainly hope I won’t have to be caught up in a revolution—but if a revolution is unavoidable, I want the person who revolts against my life to be me.”
This quote is from a personal essay by DeepSeek operator engineer Liu Shengyu, titled “I Have No Choice but to Bury My Talent in Yesterday”. In just the past few days, it has swept the internet, garnering 100,000+ views.
First, a bit of the author’s background: a 2021 cohort student from Peking University’s Turing Class, former captain of the Unnamed Supercomputing Team. Last April, he joined DeepSeek to work on MLSys and operator optimization.
You’ve probably noticed how fast DeepSeek V4.1 Flash inference has been recently. A large part of the credit goes to the main Attention operator he delivered—an MQA attention with a head dimension of 512.
In the piece, he offers an on-the-ground judgment from an engineer: a year ago, AI could only help him look up documents and find bugs; a year later, AI can independently read CUDA, PTX, and even SASS machine code, analyze the stall time of each instruction, and then optimize the operator on its own. In another six months to a year, an operator written by AI will very likely be as good as his—or even better.
Translated: in the underlying optimization loop of model inference, AI is about to figure out how to fix itself.
What’s interesting is that the person worried about being replaced by AI first faces the criticism: “Is this article written by AI?” That mismatch, in itself, is the best footnote to the times.
Even more thought-provoking is the part where he goes head-to-head with the “downgrade consensus.”
On September 12, Dario Amodei of Anthropic published a long post urging the whole industry to slow down, arguing that recursive self-improvement (RSI) is accelerating AI evolution. Musk reshared it saying “he’s right”. Altman also voiced agreement, and even Hassabis from DeepMind joined the camp. A few of the fiercest rivals in daily competition suddenly reached a rare consensus on “go slower.”
So what happened next? After calling for slowdown, not a single one actually took their foot off the accelerator.
Amodei wrote that “the frontier models must be rate-limited” while pushing for Anthropic’s IPO, aiming to complete it as early as October, with a valuation potentially exceeding $2 trillion—which would become the biggest IPO in human history. OpenAI says it isn’t in a rush, yet is seeking over $1.2 trillion in financing to train the next generation of models. The four major cloud providers’ combined 2026 capital expenditures will total over $600 billion, with the vast majority going into AI data centers.
Liu Shengyu punctures the window with one sentence: slowing down protects the frontrunners, but kills the catch-up players. In an arms race where everyone is charging forward, the one who loosens their grip first loses the most—this isn’t a moral issue, it’s the prisoner’s dilemma.
This also explains why he and DeepSeek chose to keep accelerating. In his vision, if the strongest AI is forever held in the hands of a small number of closed-source companies, society will slide into “Cyberpunk 2077”: megacorp-style powers will maintain their own armies and define how the city runs, while ordinary people live much worse lives using inferior cyberware than people do today. Open source and broad access are the only rope to pull the world back from that side.
The market’s response has always been more honest than anyone else’s. On September 15, the Philadelphia Semiconductor Index fell 5.86%, posting the largest single-day drop in two months. NVIDIA wiped out about $176.6 billion in market value in one day; AMD, Intel, and ASML fell more than 6%. Across the entire industrial chain, more than $500 billion was wiped out in a single day. Selling spread across Europe and Asia. But the money didn’t leave AI—cybersecurity stocks surged more than 10% that day—they just changed pockets.
In the same week, Trump at the All-In Summit, in front of Jensen Huang, labeled the “slowdown theory” a fraud, saying data centers are “the oil of the next 20 to 25 years.” Jensen Huang echoed it: “We won’t let that happen,” and said that AI destroying humanity “has no scientific basis.”
In the end, compute power is power. Whoever controls the strongest compute can train the strongest models, define standards, and influence regulation. For countries, companies, or individuals alike, slowing down is the same as voluntarily handing over competitive strength—counting on any side to be the first to loosen their grip is almost never realistic.
Looking back at Liu Shengyu’s choice: he could have opted out, let the wheels spin, and slowed the models down so that that day would come later. But he didn’t. He will keep making the operator as good as possible—and revolt against his own destiny.
He said he wrote this piece not to express fear of unemployment anxiety, but to bid farewell to an era: “That quiet joy of sitting at the desk, writing an operator calmly for an afternoon—it may become the final song of this summer.”
The craft remains. But that calm state of mind will ultimately be crushed by the roar of machines.
This quote is from a personal essay by DeepSeek operator engineer Liu Shengyu, titled “I Have No Choice but to Bury My Talent in Yesterday”. In just the past few days, it has swept the internet, garnering 100,000+ views.
First, a bit of the author’s background: a 2021 cohort student from Peking University’s Turing Class, former captain of the Unnamed Supercomputing Team. Last April, he joined DeepSeek to work on MLSys and operator optimization.
You’ve probably noticed how fast DeepSeek V4.1 Flash inference has been recently. A large part of the credit goes to the main Attention operator he delivered—an MQA attention with a head dimension of 512.
In the piece, he offers an on-the-ground judgment from an engineer: a year ago, AI could only help him look up documents and find bugs; a year later, AI can independently read CUDA, PTX, and even SASS machine code, analyze the stall time of each instruction, and then optimize the operator on its own. In another six months to a year, an operator written by AI will very likely be as good as his—or even better.
Translated: in the underlying optimization loop of model inference, AI is about to figure out how to fix itself.
What’s interesting is that the person worried about being replaced by AI first faces the criticism: “Is this article written by AI?” That mismatch, in itself, is the best footnote to the times.
Even more thought-provoking is the part where he goes head-to-head with the “downgrade consensus.”
On September 12, Dario Amodei of Anthropic published a long post urging the whole industry to slow down, arguing that recursive self-improvement (RSI) is accelerating AI evolution. Musk reshared it saying “he’s right”. Altman also voiced agreement, and even Hassabis from DeepMind joined the camp. A few of the fiercest rivals in daily competition suddenly reached a rare consensus on “go slower.”
So what happened next? After calling for slowdown, not a single one actually took their foot off the accelerator.
Amodei wrote that “the frontier models must be rate-limited” while pushing for Anthropic’s IPO, aiming to complete it as early as October, with a valuation potentially exceeding $2 trillion—which would become the biggest IPO in human history. OpenAI says it isn’t in a rush, yet is seeking over $1.2 trillion in financing to train the next generation of models. The four major cloud providers’ combined 2026 capital expenditures will total over $600 billion, with the vast majority going into AI data centers.
Liu Shengyu punctures the window with one sentence: slowing down protects the frontrunners, but kills the catch-up players. In an arms race where everyone is charging forward, the one who loosens their grip first loses the most—this isn’t a moral issue, it’s the prisoner’s dilemma.
This also explains why he and DeepSeek chose to keep accelerating. In his vision, if the strongest AI is forever held in the hands of a small number of closed-source companies, society will slide into “Cyberpunk 2077”: megacorp-style powers will maintain their own armies and define how the city runs, while ordinary people live much worse lives using inferior cyberware than people do today. Open source and broad access are the only rope to pull the world back from that side.
The market’s response has always been more honest than anyone else’s. On September 15, the Philadelphia Semiconductor Index fell 5.86%, posting the largest single-day drop in two months. NVIDIA wiped out about $176.6 billion in market value in one day; AMD, Intel, and ASML fell more than 6%. Across the entire industrial chain, more than $500 billion was wiped out in a single day. Selling spread across Europe and Asia. But the money didn’t leave AI—cybersecurity stocks surged more than 10% that day—they just changed pockets.
In the same week, Trump at the All-In Summit, in front of Jensen Huang, labeled the “slowdown theory” a fraud, saying data centers are “the oil of the next 20 to 25 years.” Jensen Huang echoed it: “We won’t let that happen,” and said that AI destroying humanity “has no scientific basis.”
In the end, compute power is power. Whoever controls the strongest compute can train the strongest models, define standards, and influence regulation. For countries, companies, or individuals alike, slowing down is the same as voluntarily handing over competitive strength—counting on any side to be the first to loosen their grip is almost never realistic.
Looking back at Liu Shengyu’s choice: he could have opted out, let the wheels spin, and slowed the models down so that that day would come later. But he didn’t. He will keep making the operator as good as possible—and revolt against his own destiny.
He said he wrote this piece not to express fear of unemployment anxiety, but to bid farewell to an era: “That quiet joy of sitting at the desk, writing an operator calmly for an afternoon—it may become the final song of this summer.”
The craft remains. But that calm state of mind will ultimately be crushed by the roar of machines.