Kai-Fu Lee: Originality Pride Is a Handcuff on What a Company Can Become

Lex Fridman opens with a poetic question. He emigrated from Russia to the United States at thirteen; Kai-Fu Lee emigrated to the U.S. at about the same age. Each people, Fridman suggests, has a soul that lasts across generations. What, he asks, defines the Chinese soul?

Lee answers from poverty, not poetry. Today's Chinese live with centuries of burden from a poor country, then a sudden shine of hope as China opened up and embraced a market economy over the past forty years. Tradition and hardship pull one way; the desire to be first to become successful and wealthy pulls the other. The result, he says, is a very strong hunger, a strong desire, and a strong work ethic.

Under that hunger sits an education built on excellence, dedication, and results. Chinese exams traditionally start by memorizing ten thousand characters, then history, philosophy, literature, and poetry. Lee calls it probably the strongest rote-learning mechanism ever created to guarantee memory and command of detail. The side effect is also clear: it suppresses the breakthrough innovation associated with the United States, while it enhances speed, execution, and getting results. Fridman hears an echo of Russian schooling, with its own piles of memorized poetry and its obsession with perfection. The two agree that perfection and creativity are not the same thing.

Fridman asks whether a Chinese AI engineer today differs from an American one, rooted in that culture. Lee says much is similar. AI is mastering known science, using known techniques, and trying new things. Picking among network architectures and tuning parameters is somewhat rote. Anyone who has shipped a product also knows the nightmare: more data makes the system better, and the data is unstructured, error-prone, and unclean.

The split is in the bet. The American strength, as Lee describes it, is trying things people have not done and using technology to solve most problems, including ugly data: make the algorithm error-tolerant and work around the mess. The Chinese path is to enumerate every possible setup, buy a lot of machines, and spend money and time cleaning. The engineer may write cleansing algorithms and work with thousands of people who label and correct. They ask for more data and spend less time inventing new algorithms that might paper over the rest.

Asked where the next decade's impact lies, Lee is blunt about applied work. If you are inside a company and have to deliver results, using known techniques and enhancing data is the more expedient, lower-risk path, and it tends to generate better and better results. That is why the Chinese approach has done quite well. Harder problems, including autonomous vehicles and medical diagnosis, may not yield to existing algorithms. Those are where breakthrough innovation, and the American edge, still matter.

Fridman's own intuition is that data can take us extremely far. Lee agrees up to a point. Huge amounts of data will probably solve trucks on highways, a high-value problem China is likely to lead. Full L5 autonomy, he thinks, will need technologies we do not yet know, the kind of work that wants academia and industrial research together, where the United States has an advantage.

Tesla, Fridman notes, is going full steam at L4 and L5 with data alone, including urban edge cases, not just highways. Lee thinks that will be very hard. One could characterize Tesla's approach as a Chinese-strength approach: gather all the data you can and hope it overcomes the problems. But many driving decisions are not solved by aggregating data and closing a feedback loop. They are closer to human thinking. Integrating that kind of intelligence, even the taboo phrase "expert systems," with machine learning has not yet been demonstrated.

Tesla also forgoes lidar, a valuable and reliable sensor, because the company thinks using it is foolish. The offsetting advantage is data nobody else has seen. Lee has watched Chinese companies accumulate vision and speech data unavailable in the West and deliver superior results. Those tasks, though, are relatively suitable for deep learning and do not need the analytical planning of human intelligence. A purely machine-learning path to a full conversational experience across arbitrary domains is, in his view, akin to L5. He is hesitant to say "Turing test," because the original definition was probably too easy; he means the spirit of it.

Lee spent major stretches of his career at Apple, Microsoft, and Google, so Fridman asks for the culture of Silicon Valley against China, and for the personality of each of those three companies. In aggregate, including Microsoft even though it is not in the Valley, he sees companies that dream big, hold visionary goals, and believe technology will conquer all, plus a self-confidence and self-entitlement that whatever they produce, the whole world should use and must use.

Steve Jobs's line that he does not do focus groups, he looks in the mirror and asks the person there what he wants, is for Lee the most exhilarating description of the Valley's essence. A great company should not merely ask users what they want. It should build something users will know they want when they see it but could never invent themselves. Nobody in the BlackBerry era would have written "I want multi-touch" on a survey; nobody in the FTP era asked for a browser. Once they saw those products, they wanted them.

Success also bred a belief that these are the only things worth doing, and that companies should not tread on one another's territory. Groupon, Yelp, OpenTable, and Grubhub each keep to a field; going after another company's business would not be the pride of innovating. The Chinese approach is the opposite: do whatever it takes to win. In internet markets the leader extracts almost all the value, and the "system" is not a narrow category but a sphere that keeps getting broader. That produces a practical, result-oriented, ultra-ambitious, winner-take-all gladiatorial mentality. If winning means copying a competitor without breaking the law, they copy. If it means forking a codebase into something ugly for a foreign market, they fork.

Apple, in Lee's telling, is about pleasing the user: design, brand, attention to detail, and respect for user needs, the rare tech company people will pay a premium for because of a halo. Microsoft is a platform company that builds giant products with strong moats, well architected at the bottom, work delegated to individuals, small parts summing into a whole. He calls it the most effective high-tech assembly line, a process competitors cannot easily repeat. Microsoft, Facebook, and Amazon are the three American firms he says have Chinese characteristics: they tenaciously go after adjacent markets and understand the value of platforms.

Google he sees as genuinely value-oriented, a company with a heart and soul that wants to do great things by connecting information, with strong technology genes and out-of-the-box ways of delivering value. Fridman notes the old slogan "don't be evil" and the darker public tint around Facebook. Lee's first move is a bucket of cold water. That mantra is very dangerous, because every employee's definition of evil is different, and it has led to difficult employee situations. Watching Alphabet put resources into healthcare and eradicating mosquitoes, work that is not the business of an internet company, is what actually shows heart and soul.

Trust is still hard. Many people aim their unhappiness at Facebook; some put Google in the same category and have called its practices predatory. A company is a body with two parts. The brain owes shareholders maximized profit; the heart wants to do good things that may run against the brain. On the fear that too few firms will control our data, Lee is not the loudest complainer. Lots of data does perpetuate strength and limit competition, but AI is much broader than the internet. Finance, retail, manufacturing, and education still have entrepreneurial room. The best remedy is a working VC ecosystem that keeps producing the next Google. Uber and Airbnb already emerged under the big three. China, with a larger market and more fearless founders, has produced more companies in the $50 to $300 billion range than the U.S. in the previous three or four years, including ByteDance at $75 billion and Alibaba-affiliated but independent Ant Financial at $150 billion.

Ten years earlier, branding Chinese entrepreneurs as copycats would not have been inaccurate. A founder probably could not get funded without describing which American product was being copied. "Who has proven this business model?" was a polite way of asking who they were copying. Lee does not treat that as shameful. China then had lower internet penetration and not enough indigenous experience. Lean startup said to ship a minimally viable product and iterate. Building that MVP on a proven American product was a decent starting point, so long as there was no IP infringement, and in mobile and AI he had not heard of such cases.

Silicon Valley, he notes, cannot hold two beliefs at once: that every idea must be your own, and that lean startup means trying many things and converging on what works. Fridman agrees it sounds like a smart strategy that simply lacks heroic flavor. Lee likes that phrasing: a non-heroic, acceptable, more expedient way to start. If the Valley will not let go of that baggage, it may limit a company's ceiling. Snapchat is the exhibit. Evan Spiegel built great products and was proud of building his own features, not copying others. Facebook was more willing to copy those features. "You see what happens in the competition." Putting that handcuff on a company limits its potential.

Copying, Lee says, was how Chinese founders learned from American masters, the way one learns piano by copying before innovating. About six years before this conversation they moved to a second stage: WeChat better than WhatsApp, Weibo better than Twitter, Zhihu better than Quora. The third stage is products that did not exist elsewhere: Ant Financial and Alipay, VIPKid, TikTok, Pinduoduo, Mobike, now copied abroad. Some rest on Chinese demographics that travel better to Southeast Asia and Africa than to the United States; a few, like TikTok, get traction even in America. The VC flywheel matches the market: Sinovation's first fund was $15 million, its latest $500 million, in a China market Lee puts at three to four times the size of the U.S.

Government support in China, Lee says, is not what most Americans imagine. The entrepreneurial space is left to private enterprise to self-regulate. The state builds infrastructure around it. A mass entrepreneurship and innovation plan built eight thousand incubators, a thick pipeline into VC. For autonomous vehicles, the government is building smart highways with sensors and smart cities that separate pedestrians from cars, so an initially inferior autonomy company can launch without a large jump in casualties.

Local "guiding funds" act as passive limited partners. When a fund makes money, part of the government's share is given back to the GPs and other LPs, raising everyone else's return at the state's expense and outsourcing the choice of founders to people better at it than officials. The idea came from Singapore, which got it from Israel. China added a tweak: cities compete. Officials want the next rung on a political career, so the center gives each a budget that can go to AI, bio, or energy. Whoever gets results, the city shines and the mayor is promoted. Experiments varied: cash awards to smart researchers, money for labs in hope of spin-offs, recruiting overseas Chinese to return. Mixed results. Guiding funds worked best, so every city copied them, a lean-startup story at government scale.

Infrastructure is the other half of the division of labor. Markets cannot afford nationwide 3G and 4G the way a government can, which is why Lee says Chinese wireless reception covers the whole country while American coverage can be spotty. Fridman asks whether full autonomy can be solved without that kind of spend. Lee refuses a yes-or-no. It is a how-long question: 15 years, 30, 45. Infrastructure will accelerate the day of L5. He does not expect America to build a new Chicago-sized AI city, though smaller ones exist. The U.S. highway system was built under Eisenhower. Historians, he suggests, might study how those resources were assembled, because that spend brought prosperity for decades if not a century.

After thirty years in the field, Lee offers a compact history. AI began as the pursuit of understanding human intelligence; the name itself says so. It drifted into the sub-area that worked extremely well: machine intelligence, pattern recognition on limited domains with large amounts of data, relatively simple planning, not very creative. We did not build human intelligence. We built a different machine that is a lot better than us at some problems and nowhere close on others. The public still hears "artificial intelligence" and thinks of replicating a mind. The products out there are closer to having invented the internet, the spreadsheet, or the database, then getting broader adoption.

The job-loss order is counterintuitive. It is routine work that goes first, perhaps fifty or eighty percent of a role, enough that employment falls. Most people picture the assembly line. Replacing a white-collar worker takes software; replacing a blue-collar worker takes robotics, mechanical excellence, dexterity, and unknown environments. The most endangered white-collar jobs are back-office: copy-and-paste, simple programs and data, paper and OCR, no strategic decisions, just keeping the process moving. New-employee orientation, searching past lawsuits and financial documents, reference checks. Simple interaction work follows: telesales, telemarketing, customer service. Same-location, low-dexterity physical jobs such as fruit picking, dishwashing, and assembly-line inspection sit in the same bucket.

Lee sees modest numbers in the next five years, then a rapid increase. Over fifteen to twenty years, dexterity good enough for assembly lines takes a huge chunk. Autonomous vehicles start with truck drivers and eventually reach all drivers. That is the near-term fear he wants taken seriously, not the sci-fi picture of general intelligence.

Fridman raises Andrew Yang, then a U.S. presidential candidate whose platform included automation job loss and universal basic income. Lee thinks the thinking is generally in the right direction and the campaign a little ahead of its time. Displacements will happen, but unemployment numbers are not high yet. He and Yang share a problem: theoretically convincing people of an issue that is not yet visible in the data.

He agrees on displacement and disagrees with vanilla UBI. The main issue is retraining. A monthly $1,000 or $2,000 check to do whatever you want fails if people do not know what to retrain into. Past technology revolutions replaced routine jobs with new routine jobs. AI's point is to replace routine work eventually, so it will not create routine jobs, because if it did, why would AI not do them? People lose routine roles and are offered non-routine ones. The stipend has to fund a six-month to three-year climb onto work that can last a lifetime. Read Yang's document carefully, Lee says, and that layer is there. Campaign shorthand just says "UBI," and simple UBI would not work.

Jobs AI cannot do fall into two groups. The first is creative and complex work. Current systems optimize; they cannot reason, plan, think strategically, handle complex problems, or discover a new problem and solve it. A human has to find the problem and pose it as optimization. An AI would have a hard time discovering a new drug, a new style of painting, or managing a company that is not only about the bottom line but also employee satisfaction and brand. Those jobs also satisfy the human need to accomplish something others cannot. The second, much larger group is compassionate work: empathy, human touch, human trust. AI is cold and calculating; even if it fakes warmth it will err and look silly, and people often simply do not want a robot. An entrepreneur built an elderly-care robot and found the elderly used it for one thing: tapping customer service so a real person appeared, then asking why a daughter had not called and showing photos of the grandkids. People want people.

PwC, Lee says, estimates $16 trillion of value for the world over the next eleven years. That money will buy gourmet meals, travel, and concierge services, and every dollar will grow service jobs around it. Society is already short of compassionate work. The United States will add two million incremental healthcare-service jobs in six years: nurses, hospital orderlies, elderly care, especially home care. They go unfilled because they pay poorly and carry low status, about half a heavy-equipment operator, comparable to an assembly line, while the operator's job disappears sooner. A person repeating one motion on a line gets little satisfaction; a person caring for a sick patient and getting a hug gets a different kind. Fix the pay, and people coming off routine jobs have somewhere to go. Cleaning an arbitrary home is L5-hard. A good plumber is a mini-detective. Assembly lines and auto mechanics are more replaceable. It is easier, Lee notes, to build a world-champion chess player than a mediocre plumber.

Is Lee optimistic about a future of automated repetition and human compassion and creativity? He is, if we start taking action. Five years of no action would make the devastating losses hard to deal with. Low-hanging fruit includes telling vocational schools to train more plumbers than auto mechanics, subsidizing corporate training seats, explaining why retraining matters, and rethinking education for an AI era. Incremental progress plus a wider public that understands the shift is enough, he thinks, because this revolution is comparable to electricity, industry, and the internet. Governments need not even believe unemployment will rise, or that automation will be this fast, to revamp vocational schools. People over eighty need five times as much care as those under eighty. Entitling them to a few reimbursable hours of home care under Medicare or its equivalent would stimulate a whole service industry.

On who better represents the public, companies or governments, Lee sees a double-edged sword. Both can deliver better services with more data and more AI, and both gain power that can become monopoly or corruption. Checks on how much data each holds would help. Fridman, who grew up with the Cold War, asks whether a major conflict could arise around AI and cyberspace. Lee believes in greater engagement. Much of the fear of more powerful AI rests on an arms-race metaphor. Autonomous weapons need some form of control; autonomous decision-making can leave too little time to fix an international crisis.

A Cold War mentality would be very dangerous: two countries relying on AI, not talking, each doing its own scenario planning. Something could easily go wrong. Competition can remain, but with protocols for interaction, as nuclear deterrence once required among the U.S., Russia, and China. What worries him most is that engagement is coming down and distrust is going up, especially from the United States toward China and Russia. Humankind looks foolish, he says, sitting on a chance to create $16 trillion while parts of the world remain in poverty, using the surplus to fuel superpower competition instead of hunger and disease. He worries even more about distribution. That $16 trillion will be made by the U.S., China, and a few other rich countries. Poor countries without the technology get nothing, and the old ladders, China's outsourced manufacturing and India's process and call-center work, disappear in ten or fifteen years. An ordinary citizen of a poor country may become a financial net liability rather than an asset for climbing out of poverty. He has no answer, and hopes people who know geopolitics will find one.

Fridman asks whether American-style freedom of speech correlates with creativity and entrepreneurial success, given Chinese and Russian censorship. Lee is frank: empirically that correlation is not true, and China has been successful. That is not a claim that the fundamental values are wrong, only that the perfect correlation is not there. He has not been good at reading tea leaves on opening. He does believe countries share many long-term fundamental values. China was drafting privacy rules for citizens that did not look that different from American or European ones. People want privacy and a chance to express themselves. The question is execution and timing.

AI needs data to make products feel personal. The popular fantasy, Lee says, is that you can rip all the data out of a provider, deny further access, and still enjoy the same services. Take the data back and the products give you nonsense: rankings break, recommendations break, the experience collapses. The first real step is punishment. Severe penalties for egregious misuse. China, on this point, has strong laws against selling or giving data to other companies, and in the years since those rules took effect they have pretty much eradicated illegal distribution.

The second step is to let technology fix technology. Homomorphic encryption would keep data in the user's hands, encrypted, and still allow training; that problem is unsolved, but the direction is worth pursuing. Federated learning would let each hospital train on patients it is licensed to see, then share models rather than records to assemble a super AI, a path that looks more promising. Do not treat privacy as a policy binary. Let technologists try to have most of the cake and eat most of it too.

The third step is choice and transparency, and here he is sharp about GDPR. Pop-ups asking yes or no on this site, this right, this slice of data, are a granularity no user understands. Most people just say yes. The current implementation lived up to the letter of transparency and choice and missed the spirit. A slider between perfect protection and high convenience with some privacy risk is something a person can grasp; an AI could even customize the slider. Getting that design right is the same heart-versus-profit problem as before. From Lee's perspective, the best way to make a lot of money in the long term is to keep heart and soul intact. Getting the slider right may feel like sacrificing profit; in the long term it begins user trust. He has heard utopian talk of a trusted joint platform that holds licenses to our data and feeds closed-loop suggestions through new social, search, and commerce engines. The timeline to earn that trust sits beyond a venture-capital fund cycle. Who builds it is an open question.

Lee had been a workaholic, nine to nine, six days a week, and had not paid much attention to family, friends, and the people who loved him. Life revolved around optimizing for work. The work itself was not routine, but the optimization made the life mechanical. He got highs from accomplishments he thought were the highest priority. Then came stage-four lymphoma and the possibility of death in a matter of months. Those accomplishments meant nothing. He did not feel like working another minute. If six months remained, he would spend them with loved ones, thanking them, giving love back, and apologizing for living the wrong way.

Society, he says, shapes us to think success is why we exist. Accomplishments bring periodic satisfaction. Facing death shows what is actually important. In remission, he reversed the order. His wife travels with him; when the children really need him he drops work. Family, he found, will not take fifty hours a week, so he still works hard, perhaps ten hours less. The chase had been numbers: more money, more great companies, more funds raised, a higher VC rank every year. Bigger numbers did not make him happier, and facing death made them mean nothing. Fridman pushes back from a life of sports: obsession, giving yourself fully to a task, even when it reveals flaws, is also part of being human. Lee does not deny it. He would adjust in percentages. When younger, a smaller share for family, still high priority; older, a larger share; near retirement, most of it. The point is not to work twenty hours less for a lifetime and spend it aimlessly at home. It is to put down the phone and the PC, a hundred percent present, when a wife is having a baby, a daughter has a birthday, someone is depressed or celebrating, or the family is together. That priority will not wipe out accomplishment. A happier family may even save the time once spent on the aftermath of fights.

Advice to people building the next AI company is colder. Technology waves move quickly. Two or three years earlier you could walk in with a couple of super-smart PhDs, no clear product, and a high valuation. Those days are over. AI is moving from rocket science toward the mainstream, not yet a commodity. VCs fund business and monetary value; scarcity once bought uncertainty, and the scarcity of AI talent is gone. Platforms are emerging, the cost of becoming an engineer is much lower, and there are many more of them. Founders now have to answer ordinary questions up front: whose pain, how that pain becomes value, through what channel you get paid, how much business value is created. Expect more competition and more reasonable valuations. The consolation is TensorFlow, PyTorch, and tools that make iteration faster. Think less like a laboratory. The old talent-premium path still works only if you have a breakthrough no one else has, and that is harder now.

Fridman closes with a hypothetical. Lee does not believe AGI is near, but if one were built and he could ask it a single question, what would he ask? "What is it that differentiates you and me?" A system that can answer how it differs from the person who made it has already crossed a line. One that cannot still does not know what it is. The question sits next to the distinction he has been drawing all hour: a machine that optimizes on limited domains with huge data, and a mind that can find problems, plan, feel, and, in the last six months of a life, finally see what the numbers were worth.