How a former Google scientist is trying to ride the wave of China’s ChatGPT-led AI boom
- Beijing-based Mobvoi is one of many Chinese start-ups that have jumped on the bandwagon of large language models this year
- Founder Li Zhifei graduated from Johns Hopkins University with a doctorate in computer science before he joined Google in 2010
Former Google computer scientist Li Zhifei used to wake at 8am on weekdays, but often stayed in bed a bit longer for more sleep. Not any more. “Now I immediately get up and start working,” said Li, chief executive and founder of Mobvoi Information Technology.
His decade-long journey running a start-up has not all been smooth sailing, but the popularity of generative AI has opened up new opportunities.
“In fact, I have been quite hard-working in the past five months,” Li said in a recent interview with the South China Morning Post.
China’s Quora-like platform Zhihu embeds generative AI function on Q&A service
Li graduated from Johns Hopkins University with a doctorate in computer science before he joined Google as a research scientist in May 2010. After leaving Google, he founded Mobvoi, an AI company focusing on advanced voice interaction and hardware-software integration, attracting funding from well-known investors such as Sequoia Capital and ZhenFund.
“I felt I couldn’t stay at Google even one more day because I felt a very strong sense of mission. I felt that the interaction [between human beings and devices] is very important, and it’s something I’m good at [and has] strong market potential. If I got into the field a week later, the market situation could be very different,” Li said.
According to a report published this week by the Institute of Scientific and Technical Information of China, a government research agency, China has at least 79 big AI models with more than 1 billion parameters, and they are mainly in areas of language and visual recognition.
Mobvoi’s model is called Xuliehouzi, which roughly translates as “Sequencing Monkey”. It has tens of billions of parameters, according to the company, which compares with the 175 billion parameters of GPT-3 developed by Open AI.
The Chinese ChatGPT rival submitted its initial public offering plan to Hong Kong stock exchange last week. According to its prospectus, revenue grew from 264.5 million yuan (US$37.3 million) in 2020 to 397.9 million yuan in 2021, and reached 500.2 million yuan last year. However, the company’s losses for 2022 amounted to 670 million yuan, compared with a 56 million yuan profit in 2020.
Mobvoi is working with China International Capital Corporation and China Merchants Bank International to raise up to US$300 million in the IPO, Bloomberg reported last month.
As well as LLMs, Mobvoi has launched consumer products including the Mobvoi mobile app, the TicWatch line of smartwatches, and TicPods line of wireless smart earphones, with the goal of becoming an AI wearable brand. The company also developed an AI voice dub assistant for content generators.
“When we began in 2012, we mostly were working on software and we got funding pretty quickly, but [around] 2014 and 2015 I was at a loss about how to monetise the technology … then we turned to hardware,” Li said.
“We tried many approaches over the past 10 years, and they became building blocks for us today,” he added.
Li compares the GPT model to the feeling about the internet in 2010, when people in China considered it to be the next big thing. “In 2010, you already knew that smartphone hardware was no longer an issue … Now, if you can generate a model as good as GPT-3.5, I think the next thing is to find a good scenario [in which to apply it],” Li said.
“Of course, there will be a process of filtering. Now it’s the explosion of the large language models, and some of the players will be left behind. What’s important is that each [player] has its advantages and limitations, and all of us need to find our own positions,” Li said.
Beijing’s top internet regulator has published draft guidelines that would mandate a security review of generative AI services, as governments around the world move to rein in the rapid development of these tools. Companies that provide generative AI services in China must take measures to prevent discriminatory content, false information, and content that harms personal privacy or intellectual property, according to the proposed regulations published by the Cyberspace Administration of China in April.
Li said Chinese players are at different stages of developing LLMs. “The US doesn’t have that many companies working on large models, but they have an ecosystem, and some of them are working on large models, some are working on the middle layer, and some are working on the application,” Li said. “However, China is still at the first stage … so we are about one or two years behind the US in this field.”
For Mobvoi, the next step is to make its LLM “useful” by connecting it with applications. “When the model gives our users a better experience, we will have more users, more commercialisation, and our model will get more data too,” Li said.
