Fei-Fei Li Predicts Only Two Types of Workers Will Remain in Ten Years

Intelligent cost never reaches zero

Industry claims that AI’s cost is approaching zero repeat a pattern from earlier eras (e.g., bandwidth cost). Each layer—bandwidth, compute, transmission—still incurs real expenses. Model prices are falling, but they have never been free; in many scenarios a powerful model costs as much as or more than hiring an intern.

Two common misinterpretations

AI as a layoff lever – Some managers assume that AI lets product managers code themselves ("vibe coding") and eliminates engineers. The interview contrasts this view with the observation that designers and engineers shift from executors to problem‑solvers; AI raises the skill ceiling rather than replacing roles.

Equating “can do” with “can do well” – People often treat the ability to perform a task with AI as evidence that the task is solved, ignoring the quality gap between merely possible and truly excellent outcomes.

Future workforce: two survivor categories

David Rogier proposes a bell‑curve effect: the middle tier of workers will be squeezed out, leaving only:

Top‑level specialists (e.g., world‑class copywriters, researchers) whose expertise cannot be replicated by current AI.

Generalists who build their own AI‑powered tools (e.g., Rogier’s self‑built CEO tool stack, a todo‑list app that forces decisions within 1.5 days). These individuals can deliver complete products in a weekend instead of months.

Both groups are driven by proactive initiative: experts deepen their unique domain; generalists create and define workflows.

What AI still cannot replicate

Li Fei‑Fei’s research on spatial intelligence illustrates the gap with a basketball‑shooting scenario that simultaneously requires language reasoning, spatial awareness, and physical coordination. Language intelligence has advanced rapidly (large language models excel at analysis, coding, reasoning), but perception, spatial, and bodily intelligence—evolved over >500 million years—remain far beyond AI. She estimates bridging this gap will take decades, not years.

Two flawed AI‑driven approaches

Approach 1: Treat AI as a replacement for staff – Example: product managers now prototype alone using AI, compressing cycles, leading some managers to think they can hire fewer engineers. The interview counters that AI lifts designers and engineers to tackle harder problems rather than eliminating them.

Approach 2: Conflate “AI can do a task” with “AI can do it well” – Li Fei‑Fei criticizes “vibe coding” dashboards that look good for an hour but lack backend integration, causing them to fail once data connections are needed. The gap between feasibility and quality is emphasized.

Four‑step breakdown of spatial intelligence

Spatial intelligence comprises four components: understanding, reasoning, generation, and interaction. The basketball example shows language reasoning (recognizing a made shot), spatial reasoning (seeing the court and player positions), and physical execution (adjusting the body). These modalities operate concurrently, not sequentially. Evolutionary time scales differ: spatial and bodily intelligence required >5 hundred million years, whereas language intelligence evolved much more recently.

Practical first step

Li Fei‑Fei suggests observing a person under 25 who already uses AI daily. By letting such a person demonstrate their workflow—without requiring technical knowledge—one can demystify AI and begin integrating it responsibly.

Conclusion

The premise that AI will automate intellectual labor like the industrial revolution automated physical labor is rejected. The future depends on whether individuals passively accept AI or actively redesign their relationship with it, resulting in the two survivor categories described above.

Code example

本文
约3000字
，建议阅读
7
分钟
AI正在重新划分职场的结构，10年之后，可能只剩下两类工作者。