The machines in Kai-Fu Lee’s first laboratory had a modest social ambition: keep up with a person speaking. At Carnegie Mellon in the 1980s, that was a formidable assignment. Speech runs words together, accents vary, and nobody pauses politely between each syllable so a computer can take notes. Lee’s doctoral work produced Sphinx, a system designed for continuous speech recognition across speakers. Nearly four decades later, he is asking a different kind of listener to improve. This time it is the company, and the voice trying to be heard belongs to AI.
That is a useful way to read a career often compressed into a list of employers. Apple, Microsoft, Google: each name marks a different moment in the technology’s public life. At Sinovation Ventures, Lee became an investor in the next generation of Chinese technology companies. At 01.AI, which he founded in Beijing in 2023, he became a builder again. The work began with large language models and has moved toward tools meant to sit inside the messier machinery of an organization.
A model can produce an answer. A business has to decide whether the answer changes anything. Lee’s present argument is that the second problem deserves as much attention as the first.
A voice, a board game, and a long runway
Lee was born in Taipei in 1961 and moved to the United States as a child. He studied computer science at Columbia University, graduating in 1983, then went to Carnegie Mellon for a doctorate. His adviser was Raj Reddy, a central figure in early speech recognition. Lee’s 1988 thesis bore a title that sounds like a dare: Large-vocabulary Speaker-independent Continuous Speech Recognition. The aim was to move beyond small, carefully rehearsed vocabularies toward speech a person might actually use.
He also worked on an Othello-playing program, a neat period detail from an era when games offered compact stages for testing machine intelligence. The game had a board and rules. Human speech had neither the courtesy to stay still nor the decency to arrive one word at a time. Lee kept working on the harder, untidier human interface.
At Apple in the 1990s, he led research and development work on speech and other products. At Silicon Graphics and then Microsoft, his responsibilities widened. In 1998 he became the founding director of Microsoft Research China, later known as Microsoft Research Asia. That institution mattered beyond any one product: a research lab recruits people, creates habits of inquiry, and gives younger scientists a place to do consequential work. Lee’s role was as much organizational as technical. The pattern would return.
The job change that became a public fight
In 2005, Lee left Microsoft to lead Google’s business in China. His move generated a legal dispute over the terms of his Microsoft employment. A court allowed him to begin some work for Google, and the parties settled in December of that year. The particulars of the settlement stayed private. The public drama showed how highly both companies valued someone who could bridge research, product and the Chinese market.
Lee served as president of Google China until 2009. The World Economic Forum says Google’s market share there almost doubled between 2006 and 2009. The country’s internet was expanding quickly, and the job carried more than a technical brief. It required building a team, courting users and navigating a local market with its own competitors and rules. A search engine could deliver results in milliseconds; trust and distribution took longer.
Then Lee left the executive track to found Innovation Works, the investment and incubation business later renamed Sinovation Ventures. It was a change in altitude. Instead of leading one operation, he would study many young ones and decide which founders and technologies deserved a chance. The work also gave him a vantage point on the growing Chinese startup ecosystem that would shape his writing about AI.
“Jobs will get displaced by AI and jobs will get amplified by AI.”Kai-Fu Lee, speaking at Carnegie Mellon
Lee has never confined his interest to what a model can do on a benchmark. In AI Superpowers, published in 2018, he wrote about the different strengths of the United States and China and about the effects of AI on work. In AI 2041, written with novelist Chen Qiufan, he paired imagined futures with explanations of the technology behind them. The pairing was apt: the engineer explained the machinery while the novelist put people back in the room.
The books also show a writer comfortable changing formats. A forecast can make a confident claim and leave its reader with a chart. A story has to give the forecast consequences: someone loses time, gains an opportunity, or makes a choice under pressure. Lee and Chen used ten imagined settings in AI 2041 to make technical possibilities legible through ordinary lives. For an investor who had spent years evaluating what founders might build, fiction offered another way to examine what people might do with it.
Founder, again
The generative AI surge of 2023 pulled Lee from the investor’s chair into a founder’s seat. 01.AI launched in Beijing that year and introduced Yi, a family of open models built for Chinese and English. A technical report described base models at six billion and 34 billion parameters, with later variants for chat, long context and visual tasks. The company released model code and weights for developers. Its early pitch was access: Lee wrote that he wanted better AI available to more people.
A few years can be an epoch in this field. As model prices fell and open alternatives proliferated, 01.AI’s public emphasis shifted toward enterprise use. In March 2025 it released WorldWise, a platform for deploying and adapting models inside businesses. In July 2026 it announced TrueNorth, a decision hub with products called Boss AI, Investor AI and TopSales AI. The names are plain enough to make a manager nervous, but the underlying question is serious: how does a company turn scattered records and repeated conversations into decisions it can defend?
Lee’s answer starts at the top. On 01.AI’s site, the company says AI transformation should be led by the CEO and joined to strategy, data, processes and a measure of return. That is a more demanding prescription than installing a chatbot. It also puts responsibility where it belongs. If a business reorganizes its work around AI, someone must decide what the system may do, what people still own, and how errors get corrected.
There is a continuity here with Microsoft Research China. A laboratory needs more than talented individuals and good hardware; it needs a structure that lets ideas survive beyond a demonstration. An enterprise AI system makes a similar demand of its host. Data must be available, teams must agree on what a useful answer means, and leaders must be prepared to revise a process that has quietly become part of the furniture. Lee’s current products put these organizational questions in front of the customer. They cannot be solved by a model release alone.
He has tried the proposition on himself. In a 2026 LinkedIn post, Lee described using a version of 01.AI’s Boss AI as a CEO coach. It examines patterns across months of meetings and gives him feedback he characterized as constructive and sometimes direct. There is a quietly comic image here: the veteran executive who has spent decades making machines more capable now invites one to critique his management. It is also a practical test. A tool can be admirably blunt only if its user is willing to listen.
The distance between a demo and a decision
In 2026 Lee was still talking about frontier models, the pace of progress and the competitive balance between Chinese and American AI companies. He was also talking about finances. In July, he discussed a possible Hong Kong listing for 01.AI in 2027 and a financing round ahead of it. Those are plans, not outcomes. They suggest that his enterprise turn will face the ordinary tests of a business: customers, costs and staying power.
His next book, AI Native: The Mandate to Transform Your Company, carries the same concern into print. It asks leaders to reconsider how work is organized when machine intelligence grows cheaper. The title is a management argument, but it also sums up a personal transition. Lee began with the science of recognition. He has spent years explaining the economics of adoption. Now he is asking what happens when adoption reaches the meeting agenda, the budget and the chain of command.
There are no clean answers waiting at the end of that chain. A machine may notice a pattern without understanding why people made a choice. A company may ask for insight while rewarding the familiar habit. Lee’s own long career suggests that the difficult work lies in that gap between capability and use. His advice to students at Carnegie Mellon was to follow the work they love. His work has led him from the sound of a human sentence to the texture of an organization.
For now, the former speech researcher has returned to an old problem in a new room. The machine can speak. The question is whether anyone with the power to act is listening.