By Saritha Rai
The Beijing startup founded by technology pioneer Kai-Fu Lee is introducing its first artificial-intelligence application for consumers, a step aimed at helping China capitalise on the promising technology.
Similar to Microsoft Corp.’s Office 365 Copilot, it helps users create spreadsheets, documents and slide presentations more quickly — though it’s mainly tailored for the Chinese market. It can interpret financial reports, take minutes for meetings and speed-read books as long as Elon Musk’s 600,000-word biography to give a quick synopsis. The app works in Chinese and English.
In an interview with Bloomberg, Lee said that China needs its own ChatGPT — OpenAI’s chatbot that was released in 2022 and is banned in the country — to accelerate interest, adoption and investment.
“For Americans, the moment happened 17 months ago,” Lee said over a Zoom call from Beijing. “China’s users didn’t have a ChatGPT moment. Until now, none of the Chinese chatbots or tools have been good enough.”
While US firms such as OpenAI, Meta Platforms Inc. and Alphabet Inc. have taken the lead in generative AI, Chinese players are pressing hard to catch up. In addition to 01.AI, tech players including Baidu Inc. and TikTok-parent ByteDance Ltd. are pouring funds into developing their own AI models and chatbot services. Beijing has also provided financial and policy support. Beijing bars foreign AI models in part because of its strict censorship regime, but the so-called Great Firewall also ensures that domestic players will have an enormous local market without global competition.
The Taiwan-born 62-year-old — who worked for Apple Inc. and Google before starting his own venture capital firm more than a decade ago — became chief executive officer at 01.AI last year. The startup reached a $1 billion valuation, or unicorn status, within eight months on the strength of an open-source AI model that outperformed Silicon Valley rivals on several key measures.
In addition to Wanzhi, the firm is also introducing a bigger, proprietary large language model — the technology underpinning AI chatbots — called Yi-Large, aimed at enterprise users.
Software developers will be able to use the Yi-Large model at competitive prices. Lee says the model’s application programming interface, or API, will cost $2.50 for 1 million input tokens and $12 for 1 million output tokens — about 1 million tokens lets a developer send roughly send 250 queries back and forth. That’s far less than with OpenAI’s GPT-4 Turbo, he said.
Like many Chinese companies, 01.AI stockpiled “graphics processing unit” semiconductors from Nvidia Corp. when it became apparent the US government was planning to ban the export of high-end chips, such as the H100s that are used to train leading AI services. Alibaba Group Holding Ltd., the Chinese e-commerce giant that invested in Lee’s firm, provided an additional supply of H100s, and 01.AI supplemented its needs with slightly slower and less powerful Nvidia H800 processors.
“Our models have been trained on H100s processors legally brought into China,” Lee said. “Necessity is the mother of invention, we squeeze everything we can from the compute available.”
In contrast to many global AI startups, Lee said 01.AI is nearing profitability. After training the models on Chinese and universal data sets, Lee is taking the models and apps global, and signing up domestic as well as overseas customers to boost revenue next year.
After a month-and-a-half of testing on users, Lee’s company is rolling a version of Wanzhi for PC browsers with more comprehensive features, and one for mobiles, accessible through the messaging service WeChat. He’s appearing in video tutorials on Douyin, the Chinese version of TikTok, to coach potential users.
“The year 2024 will be explosive for generative AI applications in China,” Lee said.
Lee said his startup is closing the second tranche of a $250 million pre-Series A round in a few weeks and, by the year’s end, will start looking for investors for its Series A. The company has also streamlined its hardware and software processes to maximise efficiency and keep costs down.
“When GPT-5 comes, we will be steps behind,” Lee said, referring to OpenAI’s rumored next-generation AI model. But 01.AI is focused on making AI affordable rather than creating massive, more expensive models. “You can build a giant amazing spaceship, but can it take you from Sacramento to San Francisco?”
First Published: May 13 2024 | 7:10 AM IST
