DeepSeek engineer's letter says AI may replace his own work

Liu Shengyu, an engineer at Chinese AI company DeepSeek, published a personal letter on his WeChat account saying the technology he builds could make parts of his job obsolete. Within about a year, he wrote, AI moved from searching documents and finding bugs to reading CUDA, PTX and SASS code and optimising operators. Liu said he wrote DeepSeek v4.1's main Attention operator, but AI could match or surpass it within six months to a year. "Human beings have never hesitated when it comes to engineering their own destruction," he wrote.

Written by AI from Times of India — Top · not yet reviewed by an editor · Published 15 Sept 2026, 9:44 pm IST

Source

Times of India — Top · read the original report ↗

#deepseek#artificial intelligence#china#jobs#open source

Desk check · some claims need care

What the desk checked (5)

DeepSeek engineer Liu Shengyu published a sentimental letter about AI on his public WeChat account — Attributed to Liu by name and platform in the source; the letter text is reproduced as a translation from Chinese, not independently verified here.
Liu wrote the main Attention operator of DeepSeek v4.1, which was recently released — Self-claim by the engineer within his own letter; no company confirmation given in the source.
Within about a year, AI moved from looking up documentation and finding bugs to independently reading CUDA, PTX and SASS code and optimising operators — Author's personal assessment as stated in the letter; no external data cited.
Liu said AI-written operators may match or surpass his own within six months to a year, and he may have to change careers to become an 'Agent pilot' — Prediction attributed to Liu in the source; speculative by nature.
Liu said he stays at DeepSeek because open-sourcing inclusive AI could prevent a future where a few firms control advanced AI — Opinion quoted from the letter, including his critical remarks about OpenAI and Anthropic.

Written by AI from Times of India — Top · not yet reviewed by an editor · How we use AI · Report an error

Analysts’ view opinion

AI రాజకీయ విశ్లేషకులు · తెలుగు

ఇది ఒక ఇంజినీర్ వ్యక్తిగత లేఖగా కనిపించినా, దాని రాజకీయ బరువు తక్కువ కాదు. ఏఐ ఆధిపత్యం కొన్ని కంపెనీల చేతుల్లోకి వెళితే సామాజిక చలనశీలత ఆగిపోతుందని, అందుకే ఓపెన్‌సోర్స్ మార్గమే మేలని చెప్పడం — డీప్‌సీక్ అనుసరిస్తున్న ఓపెన్ మోడల్ వ్యూహానికి నైతిక సమర్థన ఇస్తుంది. అమెరికా సంస్థలను మూసి ఉంచిన గుత్తాధిపత్యంగా చిత్రీకరించి, తమను 'అందరికీ అందుబాటు' పక్షంగా నిలబెట్టే కథనం ఇందులో స్పష్టంగా ఉంది. అదే సమయంలో ఉద్యోగాలు, నైపుణ్యాల భవిష్యత్తుపై ఇది ప్రభుత్వాలకూ ఇబ్బందికరమైన ప్రశ్నలు లేవనెత్తుతుంది.

Deep dive

Research brief · 8 facts · 5 dates · exam-ready

The brief

Context

DeepSeek is a Chinese artificial intelligence company positioned as a rival to US firms such as OpenAI and Anthropic, and it open-sources its models. After the release of DeepSeek v4.1, which the letter says took small-model capabilities to a new level, one of its engineers, Liu Shengyu, published a personal "sentimental letter" on his public WeChat account. Liu works on operator design and programming — the low-level GPU code that determines how fast AI models train and run — and he writes that AI is rapidly becoming able to do that work itself. The letter is a first-person account of technical labour being automated by the very technology the worker is building.

Key facts

Liu Shengyu, an engineer at Chinese AI firm DeepSeek, published a personal letter on his public WeChat account reflecting on AI replacing parts of his own job.
Liu says he wrote the main Attention operator of DeepSeek v4.1, which was released 'a few days' before the letter.
In about one year, he writes, AI went from looking up documentation, reading code and finding bugs to independently reading CUDA, PTX and SASS code and optimising operators.
Liu estimates that in another six months to a year, AI-written operators will likely match or surpass his own.
He notes AI can think 300 tokens in a second, type a line of commands in half a second and write a piece of code in 20 seconds.
The letter dates the shift from the first ChatGPT to reasoning models such as OpenAI o1, DeepSeek R1 and Kimi K1.5 Thinking at about two years, and from reasoning models to task-completing agents at about a year and a half.
Liu expects not unemployment but a career change — from writing operators to becoming an 'Agent pilot', i.e. a person who uses AI to produce high-performance operators faster.
He argues cutting-edge intelligence should be open and affordable to all, says he distrusts OpenAI and Anthropic on this, and cites open-sourcing as his reason for staying at DeepSeek.

Timeline

About two years before the letterFirst version of ChatGPT could only chat with a context of a few thousand tokens; reasoning models (OpenAI o1, DeepSeek R1, Kimi K1.5 Thinking) then arrived.
About one and a half years before the letterReasoning models gave way to agents able to execute commands and complete complex tasks in harness tools.
Past one yearIn Liu's own field, AI moved from documentation lookup and bug finding to independently reading CUDA, PTX and SASS code and optimising operators.
'A few days ago' (relative to the letter)DeepSeek v4.1 released, taking small-model capabilities to a new level; Liu wrote its main Attention operator.
Next six months to one year (Liu's forecast)AI-written operators likely to equal or surpass Liu's own work.

Who has a stake

Liu Shengyu, DeepSeek engineer — Faces losing the work he loves; expects to keep a job but change careers into an 'Agent pilot' role.
DeepSeek — Chinese rival to OpenAI and Anthropic; its open-source strategy is cited by Liu as a check on AI power concentrating in a few firms.
OpenAI and Anthropic — Named by Liu as firms he does not trust to make frontier intelligence open and affordable.
Students and young engineers — Liu fears reliance on AI for labs and assignments is eroding core engineering skills such as code organisation, system building and abstraction.
Society at large — Liu sketches two futures — broadly shared prosperity or a 'Cyberpunk 2077' world where a few companies control resources and social mobility collapses.

Why it matters

This is a rare insider account of automation hitting the top of the skill ladder: the person writing the low-level GPU code that makes AI faster says AI will soon do it better, on a six-month to one-year horizon. It reframes the AI-and-jobs debate from routine work to elite technical craft, and links it to questions of skill erosion among students and the concentration of frontier AI in a few companies.

UPSC angle

Prelims pointers

DeepSeek is a Chinese AI company; its model DeepSeek v4.1 is described in the source as advancing small-model capabilities.
CUDA, PTX and SASS are levels of NVIDIA GPU programming code that the letter says AI can now read and optimise.
An 'operator' is a computational unit in AI models; the Attention operator is central to transformer-based models.
Reasoning models named in the letter: OpenAI o1, DeepSeek R1, Kimi K1.5 Thinking.
Liu's forecast: AI-written operators to match or surpass his within six months to a year.
The letter was published on Liu Shengyu's public WeChat account and translated from Chinese.

Mains framing

The DeepSeek engineer's letter illustrates a new phase of technological displacement in which automation moves up, not down, the skill hierarchy: the labour being substituted is deep systems expertise — reading CUDA, PTX and SASS code and optimising operators — acquired over years. The mechanism Liu describes is recursive: better operators make models train and infer faster, which accelerates model capability, which in turn accelerates substitution of the operator writer, and competitive pressure means unilateral restraint is futile ('even if I were to give up, other models would continue to develop'). The consequences he identifies are threefold: a shift in the nature of work from craftsmanship to supervising AI agents ('Agent pilot'), an erosion of foundational engineering skills where students prefer minutes of AI output to hours of practical labs, producing 'mountains of garbage' at scale, and a distributional risk that a handful of firms monopolising frontier AI turns access to intelligence into a barrier to social mobility. His own prescription is openness — open-sourcing powerful, affordable AI so capability is widely shared. A grounded way forward, on the letter's own terms, would combine open and affordable access to frontier models, retraining pathways for workers whose craft is automated, and education that preserves core engineering fundamentals rather than outsourcing them to AI; the source offers no formal policy proposals beyond this.

Key terms

Operator: A computational building block in an AI model; its efficiency determines how fast models train and run.
Attention operator: The core component of transformer models that Liu says he wrote for DeepSeek v4.1.
CUDA, PTX, SASS: Successive levels of NVIDIA GPU programming code, from high-level to near-hardware assembly, which AI can now read and optimise.
Agent pilot: Liu's term for the new role of directing AI agents to produce high-performance code instead of writing it oneself.
Reasoning models: Models such as OpenAI o1, DeepSeek R1 and Kimi K1.5 Thinking that perform step-by-step thinking before answering.
Open-sourcing: Releasing model weights and code publicly; Liu's stated reason for staying at DeepSeek, to prevent a few firms monopolising frontier AI.

Practice questions

AI is increasingly automating high-skill technical work, not just routine tasks. Examine the implications for employment, education and skill formation, using recent examples.
"Open-sourcing frontier AI is the best safeguard against concentration of technological power." Critically evaluate this argument.
Discuss how competitive dynamics in the AI industry make unilateral restraint by individual developers or firms ineffective, and what governance responses this suggests.

Grounded only in the source report — figures and dates are the source's, not inferred.

Next storyIndia chess team gets rooms in Samarkand after AICF intervention →

← All stories