Last weekend, I gave a talk at a Silicon Valley Tsinghua alumni forum titled "Eight Startup and Investment Openings Under the Pressure of Foundation Models."
The room was packed with several hundred people. What impressed me most was not the size of the audience, but how closely people listened. As soon as the talk ended, more than twenty people pulled me into a side room and continued the discussion for a long time. The next afternoon, I had a similar conversation at Stanford with PhD students, postdocs, and potential founders. The reaction was just as strong.
What struck people was not a particular AI application or a clever list of sectors. It was a different way of looking at AI.
Most conversations about AI startups begin from technology: models, compute, agents, robotics, infrastructure, applications. Other conversations begin from the humanities: ethics, job loss, education, mental health. What I tried to do was connect the two. I wanted to ask: How is technology reshaping human beings? And once human beings are reshaped, what new needs, pains, and industries will appear?
This is what I call a techno-anthropological view.
Over the past year, I have become increasingly convinced of one thesis:
The Achilles' heels of foundation models are precisely where the next openings for AI startups, investing, and work will emerge.
The AI startup world is full of anxiety.
Investors are anxious because more and more startups look as if they could be swallowed by model companies. Founders are anxious because every time they think of a useful feature, it seems possible that a foundation-model company could ship the same capability very soon. White-collar workers are anxious because many office functions are being rewritten by AI. New college graduates are anxious because they are entering a labor market that is already changing underneath them.
In the past, a software startup could build a strong company around a workflow, a vertical SaaS product, or a knowledge-work tool. Today the question is much harsher: if a product mainly lives on a screen, in text, in documents, in workflows, in code, or in structured knowledge, how long can it remain outside the reach of foundation models?
In that sense, we are seeing a scorched-earth effect. Wherever foundation models go, many familiar opportunities disappear.
But as seed-stage investors, our job is not to stand there and lament the scorched earth. Our job is to ask: After this land has been burned through, where can new life still grow?
Seed investing is about seeing the next road before it becomes obvious. We have to discover new terrains before others do, and then position ourselves early.
So the question becomes:
What can foundation models not do?
What should foundation models not do?
And the more powerful foundation models become, what will human beings need even more?
These three questions are the starting point of the whole argument.
The first Achilles' heel of foundation models is the physical world.
Today's foundation models are much better at replacing cognitive labor than physical labor. The reason is simple: they are most at home in language, text, code, images, video, and other symbolic representations. The real physical world is not just a symbolic world.
During the talk, I used cooking as an example.
Imagine that we want to train a humanoid robot to learn from a master chef. On the surface, this seems straightforward. We can record the chef with cameras. We can collect motion data. We can let the robot imitate the movements. We can build a dataset from the whole process.
But anyone who has actually cooked knows that cooking is not just vision and motion.
What exactly is the wrist force behind a proper wok toss? How does the chef feel the heat on the face? What does the sizzling sound of the oil tell you? What does the change in smell mean? And what exactly is wok hei - the so-called breath of the wok - that experienced Chinese chefs talk about?
These things are difficult to describe fully in language. They are also difficult to capture with visual data alone. The physical world is multi-sensory, multi-variable, low-tolerance, and full of unexpected edge cases.
That is why Physical AI is the first opening.
This does not mean robotics will not develop. Quite the opposite: I believe Physical AI is a huge opportunity. But its development will not look like pure software. It will not be as linear, as fast, or as cheap. Physical environments contain endless details: lighting, temperature, humidity, material, surface, occlusion, foreign objects, misrecognition, abnormal behavior, user mistakes, equipment aging. If a system makes a mistake in a critical physical scene, the user may simply stop trusting it.
So the opportunity in Physical AI is not to shout about a universal robot. It is to go deep into vertical physical domains.
Industrial robotics, agricultural robotics, warehousing, manufacturing, repair, care work, dangerous-environment operations, and future manufacturing all contain real opportunities. In the longer run, we may even see new forms such as the "factory in a container": manufacturing no longer as a fixed building, but as a modular, unmanned, movable capability that can be placed wherever production is needed.
The physical world is one of the hardest places for foundation models to completely absorb in the near term. That makes it the first major opening for AI entrepreneurship and investment.
The second Achilles' heel is more fundamental.
The physical world is something AI still finds hard to do. Human interests are something AI should not be allowed to fully take over.
I see this at three levels.
First, AI cannot be the final judge of humanity's future. No matter how capable AI becomes, it should not become the ultimate decision-maker for the destiny of the human species. Technology can assist judgment, produce options, and improve efficiency. But human beings cannot outsource their future completely to machines.
Second, personal sovereignty cannot be handed over to machines or platforms. We are already being arranged by algorithms: what to watch, what to buy, what to believe, whom to interact with, when to be angry, why to feel anxious. Algorithms may be very smart. But if their objective function is not my interest - if it is platform engagement, ad conversion, or commercial profit - then I am being alienated by the system.
Third, human beings cannot lose meaning. If AI replaces more and more cognitive labor, and later more and more physical labor, what will humans do? Many people assume that not having to work is happiness. From an anthropological point of view, this is dangerous. Work is not only a way to make a living. It is also an infrastructure of meaning. When people lose meaning, mental health suffers.
So the second Achilles' heel of foundation models is human interests.
From here on, we are no longer talking only about AI capability. We are talking about the human position in the AI age. That is why I find it necessary to approach AI startups, investing, and employment through techno-anthropology.
For me, techno-anthropology is not an academic label. It is a method that has emerged from forty-five years of observation and participation.
I have lived through several worlds.
In 1980, at Tsinghua, I received a classic engineering education - in some sense, an education for becoming a very capable toolmaker. In 1991, I came to the United States for graduate study and experienced a deep cultural shock. In 1996, I entered Silicon Valley and experienced its technical freedom and techno-libertarian spirit from the inside. Around 2006, because I could not really understand social media and social networks through technology alone, I began a long period of reading media theory, Western postmodern philosophy, and also reconnecting with Eastern traditions such as Daoism and Buddhism. In 2025, I attended a ten-day Vipassana meditation retreat, which helped me understand the structure of the inner life more directly.
I have also lived with several insider identities: scholar, engineer, founder, investor, and practitioner of inner cultivation.
Because of these identities, I cannot look at AI investing only through a technical roadmap. Nor can I look at it only through humanistic anxiety. What I care about is this:
After technology changes human beings, what will society reorganize?
What will people lose?
What will people take back?
What will people need again?
From this perspective, the next seven openings all revolve around the same issue: human interests.
In the talk, I asked a simple question:
Why are there no white-collar workers in a wolf pack?
A wolf pack is also a species-level social system. It has cooperation and division of labor. But it does not have white-collar work. Why? Because wolves do not have writing, complex institutions, or symbolic coordination across time and space.
Human reality can be divided into three layers.
The first layer is natural reality: mountains, rivers, plants, animals, bodies, food, and the physical world.
The second layer is social coordination: how human beings organize, divide tasks, cooperate, and act together.
The third layer is symbolic-institutional reality: writing, law, contracts, accounting, procedures, reports, spreadsheets, organizational structures, and institutional systems.
Large-scale white-collar work appeared after industrial civilization. Factories became more complex. Companies became more complex. Law, finance, and management became more complex. A large class of people emerged who did not directly tighten bolts, farm land, or move goods. They processed documents, contracts, procedures, institutions, reports, code, and communications.
In other words, white-collar work is a product of the third layer of reality.
And large language models are especially good at that third layer. The documents, contracts, emails, reports, code, knowledge bases, meeting notes, and process manuals produced by white-collar workers over decades have become training material for AI. The first thing AI learned was white-collar work.
The impact, therefore, is not just task substitution. It is institutional compression.
Why does human society have so many procedures, approvals, reviews, contracts, and legal mechanisms? A big reason is that humans are unreliable. We make mistakes. We have biases. We deceive. We have private incentives. We exploit loopholes. So society has built complex systems to guard against human unreliability.
But when AI agents interact with other AI agents, if inputs and outputs are clear, permissions are defined, logs are traceable, and responsibility is auditable, many intermediate procedures designed to guard against humans can be compressed.
This is why I say white-collar work is not simply being replaced by AI. The civilizational organ on which white-collar work depends is being reconstructed.
Then what should human beings do?
I propose the concept of black-collar work.
Black-collar work is neither blue-collar nor white-collar. The best metaphor is a soccer referee.
On the field, the players are the ones kicking the ball. In many future work settings, AI will be the one executing the tasks. The black-collar worker does not kick the ball. But he runs across the field, watches for offside, calls fouls, gives yellow cards, gives red cards, and stops the game when necessary.
Black-collar work has three faces.
First, the black-collar worker is a boundary designer. What can AI do? What can it not do? Which permissions can be delegated to AI? Which red lines must remain under human control? These boundaries must be designed by humans.
Second, the black-collar worker is the gardener of an artificial nature. Future agent systems will increasingly resemble a self-growing artificial nature. New agents will be created. New processes will emerge. New capabilities will grow. The human role will be like a gardener: remove weeds and pests, prune excessive branches, discover valuable new species, replicate them, and maintain the health of the ecosystem.
Third, the black-collar worker is the representative of human interests. No matter how powerful AI becomes, someone must represent the interests of individuals, organizations, companies, and ultimately human beings. Someone must sign, take responsibility, and press the stop button at critical moments.
This will create a whole set of startup opportunities: AI governance, agent management, permission systems, audit systems, red-line mechanisms, responsibility signing, black-collar toolkits, and eventually black-collar professional training.
The next opening is the personal agent, or PA.
To understand PA, we must first understand a fundamental problem of the internet over the past thirty years: free content.
On the surface, we have enjoyed free content, free search, free social networking, and free video. But there is no free lunch. What we paid was something far more expensive: our attention, our emotions, and our nervous systems.
The platform-driven attention economy works like this: platform algorithms push content into human nervous systems, and then monetize that attention through advertising and transactions. The platform algorithm is not designed around my interests. It is designed around engagement, click-through, conversion, and revenue.
Many influencers function in a similar way. A small number of creators produce original and valuable work, and I respect them. But a large number of influencers are essentially human-shaped algorithms. They redistribute content and use emotional stimulation to extract attention.
Many generative AI tools today are only making this machine more efficient. They help platforms and creators generate articles faster, images faster, videos faster, and floods of content faster. But they do not answer the question that matters to me: Is this content actually good for me?
That is why the PA is not merely a smarter assistant. It is about taking back the entrance to my own mind.
A real PA must satisfy several conditions: I pay for it; I keep it; it understands me; it is loyal to me; it protects my privacy; it represents my cognition, my mental life, my health, and my well-being.
If an agent is not paid for and maintained by me, it cannot fully represent me. An "assistant" offered by a platform is still the platform's agent. It is not my agent.
The PA is my intermediary and my cognitive firewall. It stands between me and the platform. It filters, judges, and negotiates before anything reaches my eyes. My attention and emotions should not be directly exposed to platform algorithms.
In the future, a PA will represent the demand side - me as the individual consumer - and interact directly with the agents of suppliers and service providers. It will make contact, negotiate, and transact. In principle, much of this can happen without the platform as the central gatekeeper.
This will challenge every platform built on owning the traffic entrance. Today, many platforms derive their power from controlling discovery and distribution. Once the entrance shifts from platforms to personal agents, the power structure of the internet changes.
This is not a small feature. It is a struggle over the next entry point of the internet.
Once PAs exist, agentic media naturally follows.
Today's social media works like this: the platform recommends something, and I watch it. The algorithm reaches my eyes first, affects my emotions, and then monetizes my attention.
Agentic media reverses this order.
Content no longer rushes directly toward my eyes. It first passes through my PA. My PA decides, based on my real interests, long-term goals, current mental state, and cognitive needs, what deserves to enter my consciousness.
The essence of agentic media is this: content selection shifts from the platform to the personal agent; interaction happens first among agents before it is presented to humans; and media shifts from a recommendation mechanism to a negotiation mechanism.
This changes the objective function of media.
Social media rewards extremity, immediacy, emotional stimulation, and time spent. Agentic media may reward patience, long-term value, inner consistency, and genuine relevance.
In one sentence:
Social media competes for eyeballs. Agentic media protects consciousness.
From a techno-anthropological point of view, major technological shifts often affect society first through media. Printing, radio, television, the internet, and social media all did this. The agent era may also produce its first deep social change through media.
After agentic media comes agentic social.
Today's social networks solved the problem of connection, but they did not truly solve the problem of relationship.
We add many contacts, connect with many people on LinkedIn, and join many groups. But most of these connections are thin. A "friend" is often just a node. We do not really understand the person. There is little continued interaction, and even less long-term trust.
This is the problem of today's social networks:
Connections are excessive, but relationships are thin.
Agentic social should not mean that AI makes friends on our behalf. It should mean that agents help us restore relational context, maintain boundaries, calibrate different social circles, and turn connection back into something with memory, boundary, and warmth.
This is a shift from the social graph to relationship ecology.
It matters not only for individuals, but also for institutions.
For example, in a venture fund, our most important relationships fall into two groups. One group is LPs, the limited partners who invest in the fund. We need to build long-term trust so they continue to understand and support us. The other group is founders. Many founders connect with us on WeChat or LinkedIn, but if there is no continued interaction, the connection becomes a silent node. The real value is in sustained understanding, trust, and collaboration.
Agentic social will therefore create a new relationship operating system: personal social agents, intelligent matchmakers, small-community stewards, relationship memory, privacy authorization protocols, and tools for organizing meaningful offline gatherings.
Its objective function is not traffic. Its objective function is relational calibration.
If a PA is to truly represent me, it must truly understand me.
But to understand a person, it is not enough to understand what that person says. It must also understand emotion.
Today's foundation models are strong in language and reasoning. In that sense, their IQ is rising quickly. But their emotional intelligence is still thin. An AI that does not understand emotion cannot truly understand human beings.
The reason is simple: much of human emotion is difficult to name.
Language is only the shadow of emotion, not its full substance. Many feelings cannot be stated clearly. Many emotions cannot be named accurately. Facial expression, body movement, tone of voice, pitch, silence, breathing, heartbeat, physiological signals, smell, and chemical signals all participate in emotional communication.
That is why affective AI, or emotion-aware AI, is an important opening.
It can be applied to emotionally intelligent personal agents, customer service, sales, education, companionship, mental health, elder care, vehicle safety, and artistic creation.
But it is also a highly sensitive field. Emotional data is more private than ordinary behavioral data. If a system can detect my anxiety, loneliness, vulnerability, fear, and desire, it must be governed by stronger rules around consent, privacy, transparency, and anti-manipulation.
So affective AI is both an opportunity and a reason we will need more black-collar governance. The deeper AI enters human emotion, the more it needs boundaries and responsibility.
At this point, the question changes.
If AI replaces more and more cognitive labor, and later more and more physical labor, what will human beings do?
My prediction is that human labor will move from physical labor to cognitive labor, and then further into inner work.
After agricultural civilization, many farmers left the land and entered factories as workers. After industrial civilization developed, many forms of blue-collar work decreased relative to white-collar work. Today, if cognitive labor is also substantially automated by AI, human beings will not simply have nothing to do. A new form of work will appear.
I call it inner work, or heart-mind labor.
Inner work focuses on the internal capacities of human beings: attention, emotional resilience, self-awareness, meaning-making, relational ability, inner stability, compassion, and well-being.
Historically, many huge industries have emerged as compensatory mechanisms of civilization.
After physical labor declined, the fitness industry grew. A farmer two hundred years ago, after working in the field all day, would not come home and say, "I am going to the gym for a two-hour run." Modern office workers go to the gym precisely because their bodies lack physical labor and need compensation.
After industrial civilization and urban life developed, the entertainment industry expanded. People had more leisure, income, and mental exhaustion. Film, television, games, platforms, and content industries all grew out of this condition.
So when AI replaces more and more cognitive labor, what will human beings lack most?
I believe they will lack inner strength.
The inner-development economy - including wellness, mental fitness, meaning education, spiritual practice, and body-mind disciplines - will become a major industry of the AI age. It should not be dismissed as marginal mysticism. At its core, it is another civilizational compensation mechanism.
Many sectors will emerge: inner-strength training, psychological technology, functional nutrition, meaning education, and spaces designed for mental and emotional restoration.
Inner-strength training may become as common as going to the gym. The difference is that the training target is not muscle, but attention, emotional resilience, and self-awareness. Psychological technology may combine wearables, HRV, AI mental-health assistants, and long-term state tracking. Functional nutrition will increasingly focus on emotion, sleep, the gut-brain axis, and overall body-mind state. Meaning education will help people build values, direction, and inner order. Physical spaces may change as well: meditation corners at home, quiet capsules in airports, and new forms of retreat or wellness environments.
The fundamental reason is simple:
Human beings cannot live without work.
Without work, people lose meaning.
Without meaning, mental health suffers.
The stronger AI becomes, the more human beings will need to build their inner life.
One of the most meaningful reactions after the talk came from young parents, including mothers who brought their children to the event. They were especially interested in the final opening: education.
This is not surprising.
In the AI age, the real question parents face is: What should children learn?
If AI can write, code, solve problems, summarize, make slides, and conduct research, how should we educate the next generation?
My answer is that, in addition to knowledge education and skill education, we need to develop heart-mind education.
Schools have physical education because the body needs training. Why should there not also be heart-mind education? Do attention, emotion, self-awareness, inner stability, and a sense of meaning not also need training?
Many children today are smart, but not internally strong. They can take tests, use tools, and learn techniques, but they may not know how to face frustration, loneliness, anxiety, comparison, failure, or a loss of meaning.
Heart-mind education should become the next generation's cognitive-emotional infrastructure.
In elementary school, it can begin from the body: breath awareness, naming emotions, mindful eating, and learning to use the body as an anchor. In middle school, it can include emotional journaling, attention training, and awareness of inner dialogue. In high school, it can include values clarification, retreat experiences, and the exploration of meaning. At the university level, heart-mind education should become part of general education, and we may even explore what I call SQ - spiritual quotient.
By spiritual quotient, I do not mean a religious metric. I mean whether a person has inner strength, self-awareness, the ability to relate to others in a healthy way, compassion, responsibility, and a sense of meaning.
In the future, companies may not evaluate people only by IQ and EQ. They may also care about inner stability. A person with a strong inner life can live better, work better with others, and have a positive influence on teams, organizations, and society.
The most important education in the AI age is not to train children to become more machine-like. It is to help them become more fully human.
Let us return to the original question: where are the opportunities in AI startups, investing, and work?
If we look only from the technology side, we see models becoming stronger, application space becoming narrower, many startups being compressed, and some jobs disappearing.
But from a techno-anthropological perspective, we see not the disappearance of opportunity, but the migration of opportunity.
The stronger foundation models become, the more they will absorb the parts of civilization that are symbolic, procedural, and automatable. But precisely because AI becomes stronger, human beings will need to defend and rebuild what cannot be fully replaced by machines, manipulated by platforms, or governed by algorithms: the physical world, human interests, personal sovereignty, emotional relationships, inner growth, and meaning.
This is the common logic behind the eight openings.
Physical AI is the physical world that foundation models cannot fully enter.
Black-collar work is the human boundary-keeping role in an age of autonomous AI systems.
The PA is the tool through which individuals recover attention sovereignty and control over the entry point to their minds.
Agentic media is the media revolution from capturing eyeballs to protecting consciousness.
Agentic social is the social revolution from excessive connections to restored relationships.
Affective AI is the emotional layer that allows AI to understand humans more deeply.
The inner-development economy is the civilizational compensation mechanism after cognitive labor is automated.
Heart-mind education is the cognitive-emotional infrastructure of the next generation.
I ended the talk with three sentences:
Replacing people with machines is an efficiency revolution.
Reclaiming people from machine logic is a sovereignty revolution.
Returning human beings to themselves is a civilizational revolution.
For decades, we have used technology to increase productivity. We have replaced labor with machines and, now, many forms of cognitive work with AI. That is the efficiency revolution.
But in the process, human beings have also become increasingly machine-like. Our time, emotions, cognition, and relationships are being arranged by systems, driven by platforms, and manipulated by algorithms. The next task is to reclaim human beings from machine logic. That is the sovereignty revolution.
Ultimately, human beings must return to themselves: to the body, to emotion, to relationship, to the inner life, and to meaning. This is not a small sector or a passing trend. It is a civilizational revolution.
The next wave of opportunities in AI startups, investing, and work is not simply about finding the next tool, the next application, the next platform, or the next job title.
The deeper question is this:
When machines become better and better at doing what humans used to do, what will human beings need again?
That is where the true openings behind the Achilles' heels of foundation models begin.
[In Chinese now. English versions coming soon]
· White-Collar Work Is Becoming Redundant, but Black-Collar Work Has Just Begun - 白领正在变得多余，但“黑领”刚刚登场
· How Black-Collar Workers Will Be Made: What Children Should Learn - 黑领将怎样炼成（教育篇）：孩子应该学些什么
· Personal Agent: Taking Back Cognitive Sovereignty from the Free Internet - 个人智能体 PA：从免费互联网夺回认知主权
· Agentic Media: Why Platforms and Influencers May Lose Their Grip - 智能体媒体：会让平台和网红集体失灵
· How AI Will Disrupt and Rebuild Social Networks - AI怎样颠覆并再造社交网络
· In the Post-AI Age, Your Destiny Depends Not on Compute, but on Inner Strength - 后AI时代，决定你命运的，不是算力，而是心力
· Spiritual Quotient: In the AI Age, Humanity Must Upgrade the Life System - 灵商：AI时代，人类真正要升级的是生命系统
.png){kind=icon}
