A DeepSeek Researcher's Late-Night Confession: The Dilemma of a Tech Idealist Burying the Work He Loves

A Brilliant Engineer's Late-Night Confession

An article posted late at night by DeepSeek researcher Liu Shengyu has been making waves in tech communities abroad. He contributed to writing the main Attention kernel for DeepSeek V4.1 — one of the most demanding and technically sophisticated tasks in low-level performance optimization for large language models. Yet this engineer, widely regarded as exceptionally talented, chose to confront a rather stark reality in his post: AI is very likely about to take his job.

His words carry a tone of near-tragic calm: "Of course I'd rather not be the one getting disrupted — but if disruption is inevitable, I want to be the one disrupting myself." This line sets the emotional tenor of the entire piece and gives voice to the internal conflict that countless AI practitioners quietly carry.

The Accelerating Arc from Assisted to Surpassed

Liu draws on his own experience to illustrate how rapidly AI capabilities have evolved. A year ago, AI played a supporting role in his workflow — helping him look up documentation and track down bugs. Today, AI can read low-level assembly code like CUDA PTX and SASS and independently optimize kernels.

His assessment is calm but striking: within six months to a year, AI will very likely surpass his own ability to write kernels. Coming from someone who has pushed kernel optimization to its limits, this isn't alarmism — it's a professional reading of the technology curve.

Facing this prospect, he nonetheless chooses to keep doing the best kernel work he can. His reasoning is practical: if he doesn't do it, someone else will. Since this revolution can't be avoided, he'd rather be the one leading the charge against himself. There's a certain dignity in that posture — proactive transformation beats waiting passively to be made obsolete.

CUDA PTX (Parallel Thread Execution) and SASS (Shader Assembly) are two low-level instruction sets for NVIDIA GPUs. PTX is an intermediate assembly language that compilers generate from CUDA C++ code, designed for portability across hardware architectures. SASS is the native machine instruction set that actually runs on specific GPU hardware, directly controlling register allocation and pipeline scheduling. Being able to read and optimize code at both these levels requires deep understanding of GPU physical execution models — including warp scheduling, shared memory bank conflicts, and register pressure — the fine-grained microarchitectural details that live at the very bottom of the stack. This is the highest-barrier layer in large model performance optimization, typically mastered by only a handful of specialists. AI now operating independently at this level marks a significant milestone in its capability leap.

Not Unemployed — But Possibly Robbed of What You Love

The most affecting part of the article isn't anxiety about job loss — it's a subtler kind of grief. Liu believes he's very unlikely to lose his job, but he may lose what he once loved about it.

He describes a shift in professional identity: from someone who personally writes kernels to someone who becomes a "mech pilot commanding agents." The work remains, but the joy of carving out code line by line, of communicating directly with hardware at the lowest level — that may be gone for good.

"I am forced to bury my talent in yesterday." This line captures a predicament unique to technical people in the AI era: when the tools become powerful enough, the human role degrades from creator to coordinator, and the sense of achievement that comes from hard-won expertise gets diluted. It's not that your abilities are being denied — it's that the thing you loved is being hollowed out. In some ways, that's harder to accept than unemployment.

The "operator" (also called a kernel) mentioned in the article is the fundamental computational unit in deep learning frameworks that executes specific mathematical operations — matrix multiplication, attention mechanisms, normalization, and so on. The Attention kernel is particularly critical: it is the core component of the Transformer architecture, governing how the model captures long-range dependencies across a sequence. Its computational efficiency directly determines training speed and inference latency for the entire model. Writing and optimizing a high-performance Attention kernel from scratch demands fluency in algorithmic mathematics, parallel computing principles, and GPU microarchitecture simultaneously. It sits at the deepest intersection of hardware and software, requiring the most creative investment of any part of the stack. Many engineers describe this work as "actually talking to the machine." Losing that role and becoming an AI-orchestrating "mech pilot" is a fall from creative heights that is entirely understandable.

Two Futures: Liberation or Cyberpunk 2077

In the article's final section, Liu zooms out from the personal to the societal. He sees AI pushing the world toward one of two extremes.

At one end lies the total liberation of human productivity, with the benefits of technology reaching everyone. At the other end lies something resembling Cyberpunk 2077: a handful of tech companies controlling the most powerful AI and resources, while ordinary people are left with weaker models, and the technology gap hardens into a new class divide.

He even names Anthropic directly, raising a sharp question: if the most advanced AI always remains in the hands of a few companies, where does the world end up? This cuts to the core equity question of the AI age — could intelligence itself become a luxury good as technology concentrates?

The Anthropic that Liu names is one of the world's leading AI safety research companies, founded by former core members of the OpenAI team and developer of the Claude series of large models. Unlike OpenAI and Google, Anthropic's models are not currently open-source; its most advanced versions are accessible only via paid API access, making it a prime example of frontier AI capability concentrated within a small number of commercial institutions. Liu's choice to name Anthropic is not incidental — it points precisely at the inherent tension in the AI industry between "safety alignment" rhetoric and monopolistic tendencies. Whether closed-source development under the banner of safety is itself contributing to an asymmetric distribution of technological power is a question well worth examining.

The Reason to Stay: Making the Future Belong to More People

Given this framing, Liu's choice becomes clear: stay at DeepSeek, work on AI that is powerful, fast, and broadly accessible — and open-source it.

He believes open source might pull the world back from "the 2077 end of the spectrum." This is not just a career decision; it is a statement of values — using technological democratization to counter technological monopoly, openness to counter closure.

"Since the future is coming regardless, do everything you can to make it belong to more people." This closing line transforms personal loss into a larger sense of purpose. When someone foresees being surpassed by the very technology they helped build, yet still chooses to accelerate that process — and strives to ensure more people share in what it produces — that reflects a form of technical idealism that transcends personal gain and loss.

Conclusion: The Collective Voice of AI Practitioners

The reason Liu Shengyu's confession resonated so strongly abroad is that it articulates what many AI practitioners won't say out loud. The benefits and costs of technological progress often fall on different people — and those standing at the very front of the wave are the first to feel its pull.

The value of this article lies not in providing answers, but in honestly raising the questions: when AI becomes powerful enough, where do we put the things we love? Should technology belong to a few or to everyone? There are no standard answers — but these are questions every person living through this moment deserves to sit with seriously.