← All links
Theme
Using automatic theme
Back to top

Article notes

I Have to Bury My Talent in Yesterday

By intlsy

Read the original

Listen

Audio version

A direct reading of the notes, with clickable timestamps throughout the article.

Total length: 8:44
8:44 remaining

In one sentence

Technological progress can preserve a person’s livelihood while destroying the particular craft, identity, and pleasure that made work meaningful. Knowing this, the author still advances AI because opting out would not stop replacement—and because he would rather help shape the force that replaces him than be passively overtaken by it.

Overview

Written after the release of DeepSeek v4.1, the essay begins with the astonishing speed of AI progress, from early conversational ChatGPT to reasoning models and tool-using agents. The author then applies that trajectory to kernel engineering, his own specialty. Within roughly a year, AI has moved from documentation and debugging assistance toward reading CUDA, PTX, and SASS, analyzing instruction stalls, and independently optimizing kernels.

The author takes pride in having written DeepSeek v4.1’s main attention kernel, but expects AI-generated kernels to match or surpass his work within six months to a year. He emphasizes a basic asymmetry: AI can scale speed, reasoning effort, tool use, and parallelism; an individual human cannot.

His continued effort is not simple denial. Kernel optimization gives him intense, game-like satisfaction, but deliberately slowing down would only leave competitors to advance and replace him anyway. Thus he participates in the race, hoping—through a pun on “revolution”—that if he must be displaced, he will help build the system that does it.

The personal conclusion is restrained but painful: he probably will not become unemployed, yet he may have to change professions and lose access to the work he genuinely loved. The title names that loss: his talent will remain real, but its economic and practical relevance may belong to yesterday.

The essay then widens from one engineer’s vocation to education, engineering quality, and political power. The author worries that students will rationally skip the practice through which judgment is formed, and that weak engineers equipped with AI may create bad systems faster. He ends by contrasting broadly shared abundance with a cyberpunk future in which a few companies control frontier intelligence. His commitment to open, inexpensive access—and his decision to remain at DeepSeek—follows from that political concern.

Core ideas

Capability progress can outrun occupational adaptation

The author treats AI progress as compounding across several dimensions at once: model depth, reasoning intensity, tool calls, environmental interaction, and parallelism. Human expertise improves more slowly and cannot scale in the same way.

Replacement may mean loss of vocation, not loss of income

The most important distinction is between employment and meaningful work. A person may remain economically useful after machines surpass their specialty, while losing the activity that supplied pleasure, pride, identity, and daily rhythm.

The innovator is trapped in an arms race

The author’s work accelerates the systems that threaten his own comparative advantage. Refusing to contribute would not prevent the transition because rival organizations would continue developing the same capabilities.

Craft can be intrinsically rewarding even when economically temporary

Kernel writing is compared to a speedrunning game: inventing techniques, improving performance, and surpassing official implementations provide immediate satisfaction independent of long-term career security.

Self-displacement offers agency, not safety

Helping create the technology that supersedes him does not eliminate the loss, but gives him some control over its direction, timing, and beneficiaries. The essay frames this as preferable to being passively displaced by someone else.

Access to intelligence is a question of power

The author does not treat model capability as politically neutral. If frontier intelligence remains concentrated inside a few companies, it may harden class divisions by making access to the strongest AI a prerequisite for mobility. Open-source, affordable intelligence is presented as a counterweight to that outcome.

The argument is personal rather than a universal forecast

The six-month-to-one-year estimate concerns the author’s specialized domain and his own judgment. It is a vivid insider warning, not proof that every occupation will follow the same timetable.

Practical takeaways

Caveats and counterpoints

Questions worth revisiting

Return to this when…

Return to this essay when thinking about AI and work beyond the slogan of “job loss.” Its distinctive contribution is the idea that automation may take away a vocation before it takes away a paycheck—and that the people building the replacement may do so knowingly, competitively, and with genuine love for the craft they are making obsolete.

References

  1. Original Chinese article