A DeepSeek engineer's viral post on talent, power, and humanity is a rare snapshot of work at China's frontier AI labs

DeepSeek developer Liu Shengyu wrote a viral blog station astir the anticipation that AI could soon surpass him astatine the activity he loves. CN-STR / AFP via Getty Images

A DeepSeek technologist who helped constitute a cardinal constituent of the company's latest exemplary has fixed his readers a uncommon penetration into activity astatine a Chinese frontier laboratory — and really China's workers are contending pinch AI's risks.

In a blog post published successful Chinese on his nationalist WeChat relationship titled "I Have No Choice but to Bury My Talent successful Yesterday," Liu Shengyu, an AI developer astatine DeepSeek, reflected connected the unusual reality of helping build AI that could yet make his ain specialized skills obsolete.

Liu confirmed to Business Insider that he wrote the post, but declined to comment.

He wrote that, successful astir a year, AI had gone from a instrumentality that could thief him hunt archiving and spot bugs to 1 tin of reference GPU codification and optimizing the package operators.

As overmuch arsenic he is "proud of the occurrence of DeepSeek v4.1," Liu said he is "well aware" that the operators written by AI "will astir apt beryllium arsenic bully arsenic mine, aliases moreover surpass mine."

And successful 1 peculiarly poignant portion of the essay, Liu considers what it intends to thief build exertion that could switch him.

Human beings person ne'er hesitated erstwhile it comes to engineering their ain destruction.Why, then, do I still effort my hardest to optimize kernels, knowing afloat good that the amended I constitute them, the faster our caller models tin beryllium trained and run, the faster their capabilities will advance, and the sooner I'll beryllium replaced?Partly, it's because penning kernels feels for illustration playing a video crippled to me. I get an tremendous thrill from it. The excitement I consciousness erstwhile I invent a caller technique, aliases spot 1 of my kernels execute better, is nary little aggravated than what a runner feels erstwhile they break their ain record. And erstwhile my kernel dramatically outperforms the hardware vendor's charismatic version, I consciousness an immense consciousness of pride.But there's another, much important reason: Even if I decided to "lie flat," aliases deliberately threw a wrench successful the useful to slow down exemplary training, rival companies would support advancing their models, and I'd still extremity up being replaced each the same.Of people I'd alternatively not beryllium overthrown by the revolution. But if gyration is inevitable, I'd alternatively beryllium the 1 to "revolutionize myself."When everyone is this wished to bring astir their ain destruction, I person small prime but to subordinate this sadistic arms race.

A caller business revolution

Liu said he does not deliberation he will beryllium retired of work, but is worried that AI could return distant the hands-on portion of the occupation he loves, turning him from personification who writes operators himself into personification who chiefly steers AI agents to do it.

He likened the displacement to a talented knitter who has to springiness up making sweaters by manus and usage a faster instrumentality because everyone other successful the waste and acquisition has 1 too.

"But the pleasance of listening to the rainfall by the window, threading a needle, and slow passing the clip was yet crushed by the roar of the machines," Liu wrote.

Liu, successful his post, wondered if his readers felt the aforesaid way: having to prosecute pinch AI to protect their livelihoods, while knowing that things will alteration and return distant what they for illustration astir about their work.

"I thin to support my logical and affectional sides rather separate. When a problem calls for reason, I tin beryllium very rational. But location are times erstwhile my sentimental broadside comes through," Liu wrote.

"I retrieve sobbing erstwhile I moved retired of an flat I had rented for conscionable a year, incapable to carnivore parting pinch the memories I had made there," he added. "Saying goodbye coming to the era of hand-written kernels and optimization driven by quality ingenuity is undoubtedly acold much painful."

Anxiety complete who should power frontier AI

Liu besides warned against frontier AI being controlled by a fistful of companies, which he wrote could widen inequality by giving those pinch entree to the astir tin models an moreover greater advantage.

"Therefore, I still judge that cutting-edge intelligence should beryllium disposable to everyone successful an unfastened and affordable way," Liu wrote.

"I don't spot Anthropic aliases OpenAI to do this, and I particularly don't want Anthropic to maestro the astir precocious artificial intelligence aliases AGI; to exaggerate, the seriousness of that would beryllium nary little than Hitler acquiring atomic explosive exertion earlier the Allies," he wrote.

Liu's station comes arsenic AI leaders sound the siren astir the technology's trajectory.

Last week, Jacob Coxon, a erstwhile interrogator astatine Anthropic and OpenAI, announced he had discontinue Anthropic complete AI information concerns. In a station connected X, he wrote that AI companies were "racing consecutive to self-improving superintelligence" and "gambling pinch our lives."

Coxon's words group disconnected a wider debate, pinch leaders including Dario Amodei, Sam Altman, and Elon Musk, weighing successful connected whether the title to build much powerful AI is moving excessively quickly.

The warnings travel arsenic China makes its ain push to go a awesome unit successful frontier AI. Like DeepSeek earlier it, Moonshot AI's Kimi K3 has renewed attraction connected really quickly Chinese labs are advancing.