Everybody’s talking about LLM, but is anyone making sense?

By LI Jingya XIAO Fang

Professor LU Zhiwu develops generative AI models. Since last year, he has met with over 150 venture capitalists. Since ChatGPT came out, he doesn’t have to start at the very beginning every time. But his words still fall on deaf ears. Meetings are no less frustrating. “They are forever asking me how I make money, and what my advantages are against big tech,” the professor said.

‘Why are you trying to invent something?’

The VC world is racing to bet on the next OpenAI, but Lu feels no one understands, or bothers to understand, the underlying technology. And he might be right. After all, investments are often very diverse. Does he really expect investors to understand the underlying technology behind NEVs, e-commerce and the latest ERP suite?

Lu and his students develop multimodal large language to produce images, audio and videos. For now, ChatGPT deals only with text. A regular knuckle-headed question is: “If ChatGPT deals only with text, and everyone likes ChatGPT, why are you trying to invent something else?”

Lu spends most of his meetings answering questions about business models and his team, which to him is beside the point. “There is a set of criteria that everyone subscribes to, even though they don’t really understand anything,” he says.

“If a formula works for one cycle, investors naturally try to apply it to the next one. But for technology as complicated as generative AI, the old way of investing in mobile internet startups is pretty much useless,” says LIU Dawei, partner at Capital O. Investors talk mostly to other investors, and misconceptions become truths if enough people believe them.

Investors who don’t know how to invest

One of the golden rules of the mobile internet age was to vet the people. Visionary founders and competent management can defy unviable business models and immature markets, the thinking goes. Competence, they believe, is universal. That’s why many not-so-technically-minded investors are emboldened to have a stab at chatbots. ZhenFund founder Victor Wang told reporters that hadn’t even used ChatGPT, but he “knows” people.

“Every investor has their own people theory. I call it stereotyping, or pure laziness. When they don’t understand something, they resort to wishful thinking,” says a partner of a venture capital firm. To determine the effectiveness of the approach one only needs to look at the data. Over 70 percent of venture capital investments fail. The number is around 90 percent for angel funds.

Almost everyone who met Lu was impressed but when pressed for a decision, most of them demurred. Unlike dog-walking apps or online groceries, generative AI costs hundreds of millions of dollars and, so far, it has taken years for anything to work.

“Funds that make decisions based on people don’t have the money. Funds with the money don’t get in at such an early stage. If you count who has both the guts and the money, there are not many,” says Zheng Hengle, of Lighthouse Capital.

Why, startled commenters and commentators are asking, has China not produced its own OpenAI? One flimsy hypothesis is that venture capitalists are so used to chasing hype that they never learned how to invest.

“Most people follow the herd. True believers are rare. Plus, visionaries and long-termers are lonely,” says Liu Dawei, of Capital O. “Fund managers educated in business schools now find their knowledge inadequate for hard technology."

Wishful thinking is not science

To evaluate a generative AI project, the eggheads insist that one needs to understand how the data is collected, aggregated and processed, and how the model iterates. In fact, many funds are interviewing STEM background candidates, but they always have. This, of course, presupposes that the academics know something about how investment works, rather than wishing it worked in a way that suits them

“Generative AI rose too fast for people to catch up,” says Zheng of Lighthouse.

When asked who is qualified, LIU Yisong, of PLD Capital, waxed lyrical, without convincing. The team, Liu wistfully insists, must have someone who has: “first-hand experience with convolutional neural networks, recurrent neural networks, and multilayer perception.”

Most projects in China are applications rather than fundamental models, so the decision-maker had better know how to work with suppliers and clients, seek common interests, and benefit everyone. The skillset, says the investor, is pretty much the same as that for the mobile internet.

Professor Lu eventually found an investor, a software company called iSoftStone. The chairman tested Lu’s model himself before the meeting and even knew how he wanted to integrate the model into his products.

Like many na?ve academics turned cold-hearted entrepreneurs, Lu used to believe that good technology spoke for itself. The fundraising experience made him realize that business models and user experience matter too. After many rejections, he now has two apps: a multimodal chatbot for individual users and video-search software for business customers.

Managers over academics

“Univerisity types” are said to be not particularly good at making money, perhaps because they have more important things to do.

“Because of the high computing and hardware demand, only a few can train models adequately. Those who can come up with the tech and the money still have a long way to go. So investors are cautious,” says Liu Dawei, of Capital O.

Venture capitalists are not confident that AI geeks have the business acumen to compete against big tech. Nor are they convinced that out of the tens, if not hundreds of models that all claim to be best, the one in front of them will eventually win out.

Given such concerns, professional managers rather than academics are more likely to get funding or top jobs – the appointment of Peter Deng as OpenAI’s head of product is an example. Among his offspring are Messenger, Instagram, and Oculus, although he has a degree in media. A similar approach is to bet on startups founded by big tech veterans. Lighthouse Capital recently invested in an AI startup founded by the former head of vision technology at ByteDance

Applications are secondary

Founders and venture capitalists will keep complaining that the other side speaks a different language, but both need each other.

“Right now, the products are just not there yet, whether made by startups or big tech. Why? Because at the end of the day, applications are secondary," says Liu Yisong, of PLD Capital. "Chinese companies need their own models."