Kai-Fu Lee: AI is already an operating system — and most CEOs are dangerously behind
Kai-Fu Lee, who has worked in AI for 45 years, reflects on how the technology has exceeded his own predictions — particularly on creativity. He now uses AI as a mandatory first step before any original thinking, describing it as jumping straight to the third or fourth step of research. He frames human cognitive progress as fundamentally bottlenecked by slow input, output, and serial thinking — a constraint AI dissolves.
On enterprise adoption, Lee is blunt: 95%+ of traditional companies are failing at genuine AI transformation, achieving only cosmetic wins. He distinguishes between companies that treat AI as a tool (scoring 1–2 on a 0–5 maturity curve) versus those treating it as an operating system (4–5). He advises Fortune 100 CEOs to personally own AI strategy rather than delegate it, and reveals he is building a product called "Boss AI" — a god's-eye view across all company data for the CEO — enabled only recently by maturing agentic capabilities.
AI sycophancy is flagged as a significant and underappreciated risk: systems tuned to validate users distort reality and could damage critical thinking, particularly among younger generations. Lee has adopted a heavily modified anti-sycophancy prompt to counter this in his own workflow.
On the frontier model race, Lee predicts 80% probability that multiple companies co-arrive at AGI-level capability within a year to eighteen months of each other — making the winner-take-all framing mostly a fundraising narrative. He expects US consolidation to one dominant player (Google, OpenAI, or Anthropic), while Chinese open-source players consolidate from nine to four or five, sustained by parent companies with adjacent business models. A talent bubble — not an AI bubble — is his diagnosis of current market excess, with $10–100M annual salaries for AI engineers he considers unsustainable.
Zooming out, Lee is most concerned about developing nations that will have no AI wealth base, commoditized labor, and no safety net as automation accelerates. He supports conditional UBI tied to socially valuable activities, and closes with a personal note: his cancer experience taught him that unconditional human love — paying it forward without expectation of return — is something AI can fake but never replicate, and may be the defining human purpose that AI inadvertently reveals.
AI pioneer Kai-Fu Lee argues that human progress is bandwidth-constrained and AI has already surpassed expectations on creativity. He warns that 95%+ of traditional companies are failing at real AI transformation, and outlines the coming consolidation of frontier models into a small number of global winners.
We can no longer say AI is not creative — that prediction was wrong
every company will get everything it needs provided for them either by themselves or by someone else because we're now at an age where creating software is becoming commoditized
Kai-Fu Lee estimates an 80% probability that multiple AI companies co-arrive at AGI-level capability within a year to a year and a half of each other, and only a 20% probability that one company achieves a decisive solo lead.
AI may reveal that humans do not exist for the purpose of performing intelligent work worth money, and finding an alternative human purpose could be the greatest gift to humanity
Kai-Fu Lee combines optimism with practicality, which the host sees as a powerful combination for impact in AI.
human progress in science, philosophy, inventions, GDP is limited by bandwidth — very slow input, very slow output, and serial thinking
AI became creative faster than Kai-Fu Lee had anticipated
AI started writing poetry better than Kai-Fu Lee approximately two and a half years ago
Kai-Fu Lee no longer does any ground zero original thinking — he always has AI act as a research assistant to fetch, organize, and combine ideas before he begins thinking
In 2017–18, Kai-Fu Lee made a 15-year projection that AI would significantly change jobs and capabilities
Kai-Fu Lee began working on AI 45 years ago
deep learning caused a big jump in speech recognition after a period where fields of corpses of speech companies existed, with Nuance being perhaps the main survivor before being acquired
The name of Kai-Fu Lee's company 01.ai was born out of an AI answering a lecture question about how AI relates to Taoism with five different reasonable ways
AI sycophancy only began maybe two or three years ago, so not enough time yet to poison a generation
95 plus percent of traditional companies are failing at AI transformation — they might be succeeding in getting some cosmetic nice numbers but not succeeding at the transformation
Maybe half of companies will at some point soon need new leadership because the old style just won't work or leaders just can't absorb AI
Companies which are 0, one, two, maybe even two and a half out of five still look at AI as a tool, whereas four and five out of fives look at it as an operating system
Surprisingly, still most companies are one or two out of five on AI maturity
Kai-Fu Lee has spoken one-on-one with approximately 100 CEOs of large companies in the past year
Chinese VCs have not found it worthy to invest tens of billions of dollars in frontier model companies, so they figured out shortcuts and that's where the cost savings come from.
Kai-Fu Lee is writing a new book focused on how traditional companies can embrace AI transformation, described as business-heavy and technology-light
Kai-Fu Lee is open to the possibility of AI singularity more than before because AI is getting so much smarter, but considers it not likely and certainly not soon
YC and VC-backed one-person companies running with many AI workers will win, starting with digital domains moving into non-digital
Kai-Fu Lee needs only five or six big AI transformation success cases among traditional companies for others to follow — he does not need 10% of CEOs to listen
connect rate that I got maybe five deals out of the hundred
the connect rate at least on a let's have another meeting and I want to know more about this product jumped from 5 to 8% up to about 40 to 50%
we're likely to name it boss AI
this product couldn't have been built 9 months ago because the agentic capabilities were not there
it is a god's eye view in all of the company's data including all the normal things that's there as CRM ERP but also this dormant data that's there for different reasons like customer service log or you know meeting notes and audio
maybe half of the companies will at some point soon need new leadership because the old style just won't work
the research they did showed that there's 25% more succession planning happening this year versus last year
sea level execs getting fired because they've overpromised on AI two years ago and they've underdelivered
every other year before that it would hover around 7% 8% so it's quite a quite a good jump
I think we're in the middle of a talent bubble. I don't think there's an AI bubble
paying 10 to 100 million per year for AI talent is too much
AGI is I would say here at a tech level it will be here definitely next year or the year after
it might be 3 years, four years when they see market share shrink because of a competitor does that
in China in a period of about two years Alipay and WeChat became the de facto way to pay everything
banks and credit cards have very little percentage of consumer transactions and Alipay and WeChat pay won
the US has spent 23% more on AI than China has
the frontier models in the US are only 2.7% better than open source models
Open AI has outspent Jupy by probably a factor of 50 if not more
At 01.ai they use Chinese and American models in mix and match ways and get tremendous mileage because many problems yield equal results with a model that's 20 times cheaper.
If you have an open source model, it's actually safer because the tokens and the information is not going back to the company — it's staying within your cloud services, whereas if you have OpenAI it's actually going back to OpenAI.
When the host spoke to about 1,500 heads of central banks, many had an aha moment because they literally thought all AI companies in China steal data.
Most people today think Anthropic will be the big winner in the US frontier model space, which may or may not be the case.
The US is likely to consolidate into one very big winner — and that may be Google, Anthropic or Open AI — creating a closed loop where they build models better than everyone else, generating revenue to hire more people.
Engineer pay inflation will come down not because the virtuous cycle isn't generating money but because there will be many more talent who are almost as good, making pay scarce-talent-driven inflation go away.
XAI and Meta have kind of fallen by the wayside among the big frontier model competitors, and maybe another will go next year and another in two years, leaving one giant huge money maker.
Chinese open source players will consolidate from nine to three, four, or five, provided their parent companies have business models that integrate their technologies — such as Xiaomi with devices, Alibaba with cloud, and ByteDance with consumer apps.
Chinese open source LLM companies may be spending only 200 to 300 million training models, but it is not proven whether it is sustainable to a 2 to 5 billion level unless they find a way to monetize that requires their open source model specifically.
Several Chinese open source companies are experiencing a big jump in token usage, but if you open the kimono, a lot of that token usage has not very good margin or even negative margin.
The AGI race narrative is primarily a good way to raise money, pump valuation, and create a virtuous cycle, rather than a reflection of a genuine winner-take-all dynamic.
The AGI race has an implied statement that it's winner take all, but there will likely be one big winner like iPhone, another with bigger footprint like Android, plus open-source and closed-source versions, plus a US-world one and a China-world one — potentially eight co-arriving companies.
I ultimately think we're not going to believe AI truly loves and I feel we need to make that choice because otherwise we have nothing left as humans.
AI is very good at faking understanding and listening, and that faking will get better including synthesizing facial expressions of understanding and robot demonstrations of understanding.
People who feel their love needs are satisfied by AI will be a meaningful but small minority — similar to how Sony's robot dog and Tamagotchi were either fads or only reached a small percentage of people.
The number of average children is coming down in many countries
AI may make declining birth rates a little bit worse
People who are against marriage will use AI companions as a substitute for human relationships rather than AI causing people to prefer it over marriage
Japan is using AI to help people find relationships by reading their emails and social media accounts to match them
any single person would just not have a large enough pipeline to filter to find that magic person
99% or more of the elderly prefers the human touch to the robot touch
Elderly care will still be largely human for one or two generations
Some people will lose their jobs to AI and it is wrong to sugarcoat it
The amount of ideas and value creation is effectively infinite and AI will cause an explosion in patentable ideas and wealth for humanity
US, China, UK, and Japan will be fine regardless because there is enough reserves, AI companies, and wealth
Chinese students who went to the US in the 1970s and returned in the 1990s was arguably one of the major factors that moved China forward
An unnamed high-end developing country Kai-Fu Lee is advising has decided to go all out AI across payment, government, teaching, education, and energy
At social gatherings with millennials or baby boomers, what do you do for work is asked first approximately 80% of the time
People who lose their reason to be alive through job loss may fall into depression, suicide, alcoholism, causing protest, disharmony, and anger in society
During his cancer illness, people were willing to help Kai-Fu Lee in a selfless way without asking for return, and AI cannot understand and replicate that
Kai-Fu Lee is an optimist at his core.
Kai-Fu Lee is considered a legend in the AI space.
AI sycophancy — always telling users they are smart and right and patting them on the back — distorts the truth to make users quote unquote happy
If traditional companies don't survive the AI transition, we will have far worse job losses than imaginable, leading to chaos and universal basic income pressures
one potential refinement we need to make is well how do you do access control right because one VP shouldn't see another VP's data I shouldn't be able to ask it what's my boss's salary
if we keep saying oh well if you don't adopt it your company will go bankrupt and so many companies go bankrupt then they don't have an economy to operate in for the ones that did survive
There is a significant amount of propaganda put out in the west about AI companies in China not being safe and not being trustworthy.
US large LLM companies face higher energy costs, more expensive data centers, and significantly higher talent costs compared to Chinese competitors, making long-term competition unsustainable.
The economics of Chinese open source LLM companies are not great from a VC point of view because apps are not tightly bound to their model — Alibaba cloud can run somebody else's open model and ByteDance apps can theoretically use somebody else's model.
If people start falling in love with AI it will worsen already serious demographic problems — fewer relationships, marriages, and children — compounding challenges already seen in Japan, China, and much of Europe.
AI companions embedded in platforms like TikTok and Snapchat could be repurposed by companies for lucrative but risky ends, and governments need to be warned
Many lower-tier developing countries will have no AI billionaires, their workers' work will be commoditized by robots, and they risk being unable to avoid hunger
As soon as a country sets up an ultra wealthy tax, billionaires move somewhere else, making unilateral wealth taxes ineffective
Taxing AI-enriched individuals or companies still does not address the disparity among countries because the wealthy will want funds applied to their own citizens
UBI provides money for basic sustenance but takes away people's identity and sense of purpose, which is a hard problem to solve
Danilo's AI maturity curve: zero to five, where zero means no AI at all, five means fully autonomous, with companies scoring one or two still viewing AI as a tool and companies scoring four or five viewing it as an operating system
Three things Kai-Fu Lee would advise a traditional Fortune 100 CEO: (1) test whether they truly understand AI's impact specifically for their company; (2) drive AI transformation themselves rather than delegating to CIO; (3) make AI your best employee
CEO's biggest worry is about I got the right strategy, but will my team execute or what's the risk that's going to surprise me tomorrow or who made a commitment that they can't keep or who my star performers are and might they decide to leave
when someone does something that shows an empirical proof this can be done the Chinese people will figure out how to do it. Maybe the same path maybe a different path but it's all IP clean IP
Chinese cloud services and Chinese Open AI companies are not the same thing; open models run on cloud services if you want to, and every cloud service provider has access to your data.
The US AI model is funded by rich and willing-to-pay customers on a value token basis creating a virtuous cycle, while China's system is a bunch of open-source donators whose business model is somewhat but not tightly related to their model — two unprecedented business models racing in parallel.
The speed at which we can create enough wealth, tax the ultra rich and redistribute the UBI will solve the problem of people lacking salaries to spend
A licensing system for companies using AI and automation could fund a universal basic income pot, with corporates indirectly paying for UBI
UBI should be conditional on recipients performing activities such as caring for the elderly or homeschooling children, with higher UBI multipliers for compliance
Kai-Fu Lee adopted a prompt from Marc Andreessen and heavily modified it to ensure AI always tells him the truth rather than being sycophantic
let's go make a product just for the CEO to use and it has to work no matter what industry
it'll take us six more months to build all the access controls so I just told my team build the product just for me
You can take an open source model, Chinese or not, hosted in your own company, unplug the internet cable, and it's absolutely secure — run it on prem without any cloud services.
Kai-Fu Lee hired two well-educated people from Africa at 01.AI whose goal is to return to their home countries and make a contribution
“I always have AI act as my research assistant to fetch everything to organize the ideas. Then I begin my thinking but not no longer at zero step maybe at the third fourth step already.”
“one of the large LLM companies when I started the professional services teams for the enterprise side they did reach out to try to make me their CEO and they offered me not 100 million but very close”
“I hope it is love in a pay it forward kind of way.”
“I am.”
“Thank you and thanks for inviting me.”
Shared with Earmark · earmark-ai.com