A DeepSeek Engineer’s Warning About AI, Obsolescence, and Power

Shengyu Liu, the DeepSeek machine-learning systems engineer known online as “interestingLSY,” framed his essay I Have No Choice but to Bury My Talent in Yesterday as more than a prediction about coding tools. It is a reflection on what happens when an AI builder sees the technology approaching the specialized craft that gives his own work meaning—and a political argument against concentrating frontier AI capability in a small number of closed companies. Liu’s public profile identifies his work as machine-learning systems and kernel design and optimization at DeepSeek. 7

From assistant to potential peer

Liu’s central fear is not simply that AI will automate routine programming. He described a rapid progression: first, systems helped engineers find documentation and debug problems; then they became capable of examining GPU-level code, understanding an operator’s implementation and performance constraints, and suggesting or carrying out optimizations. 1 11

For a kernel optimization engineer, that shift matters because the valuable work lies in specialized, painstaking judgment about low-level performance. Liu’s concern was that increasingly capable systems could eventually perform the part of engineering he values most. The result, in his view, may not be immediate unemployment. An engineer could remain employed while moving from hands-on creator to supervisor of automated agents—a loss captured by the essay’s title. 1 3

That distinction is important. The essay is a personal account of perceived technological change, not proof that AI can already autonomously replace kernel engineers across real production environments. Liu is expressing a frontier engineer’s expectation about where capabilities are headed.

Why he objects to concentrated AGI power

Liu paired his concern about automation with an argument about ownership and access. He said he did not want Anthropic to control the world’s most advanced AI or AGI, using an intentionally extreme analogy: the stakes, he wrote, were comparable to Hitler obtaining atomic-bomb technology before the Allies. 2 9

The analogy is Liu’s rhetoric, not an evidence-based comparison between Anthropic and Nazi Germany. Its underlying claim is clearer: if a single private organization gains a decisive lead in highly capable AI, it could hold an extraordinary and insufficiently accountable concentration of economic and geopolitical power.

His preferred answer is frontier AI that is open and affordable enough to be broadly accessible. Liu presented DeepSeek’s open-source orientation as a counterweight to closed Western frontier labs, and as part of why he chose to work there. 1 3 This is a normative case for openness, not a settled safety conclusion. Wider access may diffuse control, but it can also make powerful capabilities easier for malicious or irresponsible actors to obtain.

Distrust of Anthropic and OpenAI

Liu’s distrust of Anthropic and OpenAI is political as well as commercial. He questions whether companies developing proprietary frontier systems—and advocating policies that can limit rivals’ access to advanced computing—should be trusted as neutral stewards of AGI.

That is Liu’s judgment. The available reporting does not establish that either company is acting as an unaccountable controller of AGI, nor that Liu’s preferred open-access model is safer. The dispute is fundamentally about which risks deserve priority: concentration of power, rapid proliferation of capabilities, competitive imbalance, or catastrophic misuse.

The argument sits inside a U.S.–China AI conflict

Liu’s essay surfaced amid a sharper public confrontation between U.S. and Chinese AI interests.

Anthropic has alleged that DeepSeek, Moonshot, and MiniMax conducted large-scale unauthorized distillation campaigns using Claude outputs to improve their own models. Anthropic said the activity involved more than 16 million Claude exchanges through roughly 24,000 fraudulent accounts. These are Anthropic’s allegations; they should not be treated as adjudicated findings in this context.

At the same time, Anthropic CEO Dario Amodei has called for the industry to slow the rate at which it improves frontier-model capabilities, arguing that more time is needed for safety work, oversight, and risk management. He has also urged continued restrictions on exports of the most advanced AI chips and chipmaking equipment to China, saying a Chinese AI lead would pose serious risks to the United States and the world.

Chinese officials and state-backed media have rejected that framing. China’s Foreign Ministry criticized Amodei’s call to curb China’s AI capabilities, while the Global Times characterized the slowdown proposal as a “Cold War playbook” aimed at preserving U.S. technological advantage.

A second critique: the race itself

The debate does not split neatly into “Western safety advocates” versus “Chinese accelerationists.” Former Anthropic and OpenAI researcher Jacob Coxon resigned from Anthropic and left the AI industry in September 2026, saying that both companies were racing toward self-improving superintelligence without acting responsibly. He described that race as “gambling with our lives.” 17

Coxon’s critique differs from Liu’s: Coxon focuses on the dangers of building increasingly capable AI too quickly, while Liu emphasizes the danger of letting a closed U.S. lab dominate it. But both reject the idea that the current competitive race is automatically legitimate or benign.

The unresolved question

Liu’s essay turns a technical career anxiety into a larger governance problem. If AI can absorb highly specialized human skills, who sets the rules for its deployment? If a few firms control the most capable systems, how should their power be constrained? And if those capabilities are opened widely, how can society manage misuse and proliferation?

There is no consensus answer. Liu argues that openness and affordability are the necessary counterweight to private concentration. Anthropic argues for paced development, stronger safety controls, and restrictions intended to preserve a U.S. lead. Chinese critics see those restrictions as containment.

What Liu adds is the perspective of someone building the systems at the center of that conflict: the possibility that AI’s first major disruption may be not only to ordinary work, but to the expert craft of the people training it to advance.