Dec. 4, 2019
This article was published as part of the Stanford-New America DigiChina Project's first special report, AI Policy and China: Realities of State-Led Development. Read this article and the others in PDF here.
Is an “AI Winter” coming to China? Some leading investors in the Chinese AI space think so. This article looks at Chinese sources from two popular new media outlets to explore why some see stagnation where others see a booming industry.
'DougLong' and the Doldroms of the AI business
“DougLong,” an anonymous investor and entrepreneur in AI, reflecting on Gary Marcus’ recent book Rebooting AI, points to three big issues with AI in 2019. First, good data is hard to come by. Second, test training data doesn’t line up with actual operating environments. And third, “to B” companies having a hard time retaining AI talent for the long haul.
While there is little available as to the DougLong’s true identity, the piece was published by respected tech media outlet and consulting firm Jazzyear, giving it a large audience in a country where pseudonymous commentary is a common tool of discourse. On the first issue, the dearth of good data, DougLong quotes an anonymous old hand in the data industry who just joined an AI startup and has been sorely disappointed with AI’s effectiveness in the marketplace.
“When customers see how many terabytes or petabytes they have in their database, they think they have big data. … But once we get down to work, the data is basically useless. Some fields are mis-entered, others are too sparse. Once you finish cleaning it up, the data leads to totally logically unsound conclusions, with no chance to do any deep learning. …
“For instance, once I did a fault detection project for a Zhejiang tire factory. … Hundreds of thousands of tires were piled up in open air collecting dust, so we had to hire people climb these tire mountains, clear away the dust, and write down the tires’ model and batch numbers and their faults. But on hot days—dirty and tired—some workers just lazed around and wrote up fake data. …
“The fact that some data sources are borderline illegal is an open industry secret. In some industries where information security measures are weak, it’s most cost-effective to find some internal personnel to just copy off full hard drives.”
DougLong used an analogy to Chinese medicine to explain the importance of AI engineers sticking to one project for an extended period of time. “Since the theory around deep learning isn’t complete and it’s impossible to understand the operating mechanisms of algorithms, the success of various adjustments and improvements depends on experience combined with luck, and capabilities are hard to replicate quickly. It’s like learning Chinese medicine. For a novice practitioner to mature into a high-level talent, one must complete many cases and encounter many conditions. Old Chinese medicine practitioners accumulate personal experience of successes and failures, depending on cases and insights to accumulate experience in the ‘four methods of diagnosis.’”
But the way the industry is set up makes it difficult to support top talents, especially in “to B” businesses. DougLong writes that most projects require specialized, non-scalable work, and many senior people already have families and don’t like long business trips. Said one cloud sales rep at a BAT (Baidu, Alibaba, or Tencent) firm:
“One time a customer asked us to do a proof-of-concept for an AI project and wanted some high-level people on it. So I pulled out all the stops and borrowed a few people from AI research institutions. They went on-site for six weeks, but the project didn’t succeed. When I tried to get them again, they wouldn’t answer.
“They don’t like doing client projects, and what’s more, they can’t use that time to publish papers. And with such expensive outlays of human resources, there’s no guarantee of standing out in the year-end results.”
Building Foundations, or Seeking Handouts?
Liu Shui, investment director at the incubator CAS Star, encouraged startups to focus on the tech underlying advances in AI. According to Liu, “AI chips are the main battlefield, regardless of whether you’re pursuing hardware or algorithms.” But he sympathized with the plight of entrepreneurs in the space. “Even if your technology is top notch, it’s very difficult to commercialize a product. First, finding funding is hard. And second, there isn’t a lot of support to help transitioning from tech to a viable product.” He also admitted that China lacks investors with a background in AI to help back the best startups.
Wang Sen’ao, an executive at Lieyun Net, advised AI startups struggling to find cash to “change yourself to let the government back you unconditionally.” It’s possible for a startup after a year of operation to get subsidies approaching U.S. $1 million, but doing so is different than pitching yourself to private sector investors. According to Wang, firms need a model that aligns both with government aims and their own interests. Doing so requires entrepreneurs to “push [government] guiding documents from top to bottom,” because, Wang says, when it comes to national industrial policy, State Council instructions filter all the way down, and their administrative measures and implementing rules require rapid study.
Overall, independent Chinese AI firms are facing similar headwinds to western ones as they struggle to commercialize their technology. It remains to be seen whether government support is a difference-maker. Ultimately, DougLong leaves readers with the following advice: “Give up fantasies, buckle down, and stay strong to get through the winter.”