Everyone is anxious about AI taking jobs. “Godmother of AI” Fei-Fei Li’s latest interview cuts straight to the truth of our era: in ten years, the workplace will only retain two types of workers. The mediocre middle, neither outstanding nor terrible, will see their living space continuously shrink. The industry is buzzing with the claim that “the cost of intelligence is approaching zero,” but she calls this a highly misleading, one-sided assertion. Human intelligence—our unique perception, empathy, spatial reasoning, and physical coordination, refined through billions of years of evolution—is far beyond what large models can replicate.
Most companies fall into three automation mindset traps: treating AI as a layoff replacement tool, piling up tools and calling it digital transformation, and issuing crude company-wide AI deployment orders that trigger employee panic. True AI transformation never eliminates human labor; it elevates human value. The article throws out a sharp “barbell effect”: on one end are the top 1% of deep-domain experts who use AI to strip away repetitive tasks and focus on irreplaceable, in-depth judgment; on the other are polymath generalists who actively restructure workflows, autonomously harness tools, and build their own dedicated business systems.
Ordinary people need not fear technological iteration. The core breakthrough lies in abandoning a passive execution mindset. Below, enjoy:
“Godmother of AI” Fei-Fei Li and MasterClass founder & CEO David Rogier had a podcast interview titled “The Godmother of AI: In 10 Years, Only Two Types of Workers Will Remain.”
The host asked Fei-Fei Li what she thought about a widely circulated claim: the Industrial Revolution automated physical labor, and now AI will automate intellectual labor—what should we do?
Fei-Fei Li countered that the Industrial Revolution did not automate labor. It made labor more efficient, expanded the scale of labor, and indeed changed the labor market. But it did not automate labor. And we cannot imply that labor lacks intelligence—that assumption is terribly wrong.
It changed the form of labor, but the judgment people pour into their work—the intuition accumulated by craftsmen over a lifetime, the cognitive judgments embedded in physical labor—has never truly been automated.
The same misunderstanding is being repeated with AI.
01. “The Cost of Intelligence Is Approaching Zero” Is a Misunderstanding
A phrase is popular in the AI industry lately: the cost of intelligence is approaching zero.
Fei-Fei Li directly responded: physical labor, cognitive labor, emotional labor—human activities and human intelligence are deeply intertwined.
Human intelligence remains an unsolved mystery to nature. We don’t really understand the depth and nuances of human intelligence. So anyone out there claiming “the cost of intelligence is approaching zero” is making an irresponsible statement.
Immediately, she gave a second reason.
Even considering just linguistic intelligence, large language models are indeed powerful. “They have shown their might in assisting business intelligence, software engineering, deductive logical reasoning, and even deeper tasks.”
But beyond the linguistic intelligence we’re more familiar with, we also have perceptual intelligence, spatial intelligence, bodily intelligence, and emotional intelligence. We haven’t even figured out where creativity comes from. Each person’s creativity arises from different parts of their brain and from different parts of their entire life experience.
For example:
When a teacher judges why a student isn’t learning, it’s not just text analysis; it’s observing expressions, tone, moments of hesitation.
When a team leader decides whether to say that line in front of a key client, no algorithm can make that judgment for them.
If “the cost of intelligence approaching zero” is taken as a premise for managerial decisions, what it misses is precisely the most expensive part of a person.
02. Three Practices That Expose Automation Thinking
Fei-Fei Li repeatedly reaffirmed the same stance in the interview: I genuinely believe it is a technology—that is, it is merely an extremely powerful tool. But it is a tool that humans can use to make things better. At the same time, we must be very vigilant about how we use this tool.
She added a more important line: “We teach children how to use fire, knives, and then the internet. Now, as a species, as a society, we must learn how to use AI.”
Whether a tool is used to replace people or uplift them does not depend on the technology, but on the people deploying it.
Three practices in the interview correspond to three inertia traps of automation thinking.
First, treating AI as a headcount replacer.
Fei-Fei Li gave the example of product managers.
The standard product manager a decade ago “was more like a conductor. They didn’t need to write code; they usually weren’t software engineers.”
For a prototype, you looked for a designer. For development, you waited for engineers. You got the prototype, sent it to users, waited for feedback, then integrated it. “That product management lifecycle could take months in a typical company.”
Now?
Many product managers now write code themselves. They don’t need to wait for a team to build a prototype; they can use AI to help design some very simple things and do “vibe coding.” That immediately shortens the cycle.
But that doesn’t mean we should ditch designers and software engineers—it just saves time so they can work on the more complex parts of the job.
AI has not replaced anyone. It has pushed everyone up a step. The product manager went from “conductor” to “doer.” Designers and engineers went from “executor” to “specialist tackling the hardest problems.”
A manager stuck in automation thinking, upon seeing this example, would likely react with “then we can hire two fewer engineers in the future.” Same fact, same tool, radically different conclusion. The difference isn’t in the technology, but in how the technology is understood.
Second, equating “the tools are live” with “we’ve done it right.”
Buying tools, conducting training, teaching employees to write prompts, and once taught, considering the task done.
Fei-Fei Li said something about education in the interview: The goal of education is not whether exams are closed-book or open-book, nor standardized test scores.
The goal of education is to cultivate people, enabling each person to become a meaningful contributor to their community and society, and to live a meaningful life. AI should not strip away any of these fundamental goals. But AI should help achieve these goals better and more effectively.
Replace “education” with “management,” and “student” with “team,” and every sentence holds true. The goal of introducing AI isn’t “getting the tools installed,” but what exactly you are redesigning with AI.
Rogier is not from a technical background, but he mentioned something he is doing:
“I found that most of the apps I use are ones I built myself, using Claude Code or Cursor.”
My CEO tool stack is entirely composed of apps I made myself. Even my productivity app and my to-do list app are ones I built. The cost of making an app has shrunk from months to a single weekend.
He’s not showing off tech; he’s demonstrating: in the AI era, excellent people aren’t “those who are better at executing tasks,” but “those who are better at designing work systems.” No matter how many tools you buy, if the team lacks a design mindset, the tools will just become digitized old processes.
Third, assuming that “roll out AI company-wide” is a technical order.
Send a notice saying “the company will roll out AI across the board,” and employees hear “layoffs are coming.” Sit down and tell them “let’s see what you can do with AI that you couldn’t do before,” and employees hear “you can become stronger.”
This is an interesting phenomenon: employees hesitate at first not because they don’t know how to use the tool, but because they don’t know what the manager really intends to do with AI—replace them, or uplift them.
Same tool, same budget, same person. Different preconditions lead to completely different outcomes.
In the latter part of the interview, when the host asked her what is the simplest way for someone who has no idea where to start to get started with AI, Fei-Fei Li gave this answer:
“Find a young person. Your kid, a niece or nephew—basically anyone under 25—most of them are already using AI.”
With pure curiosity, ask them to show you how they use it, what they’re doing with AI in their daily lives. Once you really understand what it is, that world won’t be so scary anymore.
03. In the Next 10 Years, Only 2 Types of Workers Will Remain in the Workplace
After escaping the trap of automation thinking, Rogier’s description of the workplace structure takes on a completely different meaning at the two ends of the barbell.
Rogier said in the interview: My hypothesis is that you will see a barbell effect emerge. There is a cohort of people becoming true experts.
A so-so copywriter—now anyone can do a decent job with a large language model. But if I’m the best copywriter in the world, or in the top 1%, then you can’t easily beat me.
And the other role we’re seeing is the high-agency generalist. They can do many different things and have strong skills in judgment and initiative.
The two ends of the barbell: one end is the top 1% expert, the other the high-agency generalist who can juggle multiple things. The “good enough” people in the middle are seeing their space compressed.
Fei-Fei Li agreed with this judgment and added a layer of analysis:
Whether you are on the expert side or the generalist side, you need to have agency; you should be able to use tools in a unique, creative, deep way.
The top expert on the left end is the person who maximizes augmentation. AI helps them filter out 90% of repetitive work, allowing them to focus on the 10% that most requires human judgment. Their value isn’t compressed—it’s unleashed.
The generalist on the right end is the one who proactively initiates augmentation. They do it themselves, build their own tools, define their own workflows. They aren’t waiting for an augmented future—they are the starting point of augmentation.
The middle layer’s problem isn’t a skill issue—it’s a posture issue. AI has raised the bar of “good enough” execution to an extremely high level. Anyone who stays at the level of “able to execute” will be caught up with, no matter what they do.
But as long as they shift from “waiting to be told how to use AI” to “let me see what I can do with it myself,” the middle layer has a chance to push themselves to either end of the barbell.
Fei-Fei Li also specifically talked about this shift, saying: the word “entrepreneur” is, to a large extent, a synonym for “initiative.”
04. Why Augmentation Isn’t Wishful Thinking
Someone might ask: What if technology advances further, and human judgment, creativity, and emotional intelligence are all automated?
Fei-Fei Li devoted a lengthy section of the interview to the scientific version of the same issue. Her company and research focus is spatial intelligence.
Spatial intelligence boils down to four things: understanding, reasoning, generating, and interacting.
It encompasses multiple abilities that we humans demonstrate today in 3D (three-dimensional) environments.
First, we can understand what is happening;
Second, we can reason;
Third, we can generate;
Fourth, last but not least, interaction.
Fei-Fei Li gave the example of shooting a basketball:
Even the act of shooting a basketball itself is a highly complex moment of intelligence, and linguistic reasoning is involved. Because as an athlete, you are keenly aware of whether the shot went in or not, and what it means for the game, for that moment.
At the same time, seeing the entire court, seeing the positions of other players, and aiming at the basket—this is deeply spatial. Then adjusting your body, knowing how to execute that movement—this is deeply physical.
Three kinds of intelligence—linguistic, spatial, and physical—work simultaneously and synergize in the split second of a jump shot, not an assembly line of “first language, then space, then body.”
And the vast majority of things we do in life are actually a mixture of linguistic intelligence, spatial intelligence, and physical intelligence. They are highly complementary and work together synergistically.
Then Fei-Fei Li offered a judgment from the perspective of evolution: It took evolution over 500 million years to mature spatial intelligence, while linguistic intelligence took far less time. So this is a very deep, ancient, and fundamental cognitive ability, one that both animals and humans possess.
These judgments, taken together, point to the same thing. What AI can truly accelerate today are tasks at the language level: writing reports, looking up information, doing data analysis, writing code, generating images.
It gives people more time and energy to do things beyond language: judgment, creation, empathy, making decisions in ambiguous zones, staying calm under pressure, and focusing on the most important signal when various signals contradict each other.
Augmentation is not a wish or a value choice. At this stage of technological development, it is a scientific judgment: humans still have plenty of things that AI cannot catch up to.
Using the framework of augmentation, what you save is repetitive labor, and what you gain is liberated professional judgment.
Fei-Fei Li said something in the interview that can serve as a litmus test for all AI management decisions: We teach children how to use fire, how to use knives, and then how to use the Internet. Now, as a species, as a society, we must learn how to use AI.
The key word is not “learn,” it’s “we.” It’s not about letting employees learn it on their own, nor letting the IT department deploy it. It’s about managers and teams together, treating AI as something they need to collectively figure out, and using it to push everyone up one level.
