ChatGPT is good at writing fake news, but is the technology is ready for serious venture capital attention?
Photo from CFP
By LI Jingya
ChatGPT, OpenAI’s machine learning-powered chatbot, is a big deal in artificial intelligence. While laypeople can’t get enough of the fun getting ChatGPT to write high school essays, debug computer code, and answer goofy questions in the style of Shakespeare, the venture capital world is already pondering where the money is in the brave new world.
Most conceivable applications fall under the umbrella of AI-generated content (AIGC), arguably the hottest AI subsector in 2022. Some of the most valuable startups in this category are copywriting apps, which generate marketing material for business subscribers. In the US, Jasper.ai, based on Open AI’s GPT-3, has reached 70,000 paying users and is valued at US$1.5 billion (10.35 billion yuan).
LIANG Junzhang, founder of Kinzon Capital, said startups are more receptive to ChatGPT than tech giants. A marketing firm he invests in, Beijing Zhong Meng, is already using ChatGPT in its products. Tencent, on the other hand, recently banned WeChat mini-programs running ChatGPT. Tech giants may prefer to develop their own AI models for business strategy and data security concerns, he said.
Investors and founders have also pointed to ChatGPT’s potential for long-form content for journalism, TV writing and even literature. They are, however, concerned about content quality. “ChatGPT is as good at writing news as fake news. I don’t think the technology is ready yet for serious content,” said HAN Sanpu, who founded marketing script generator Megaview. XI Xingjie, of China Renaissance, believes that mass adoption will be limited to marketing and entertainment in the short term.
Within these two, possibilities go beyond text. Moviebook, valued at US$2 billion in 2019 and also a Kinzon Capital investee, generates videos according to user specifications. OpenAI’s DALL-E produces paintings based on prompts given by users. Other promising applications include AI-enhanced gaming and social media avatars. “Short videos and avatars are both great investment opportunities,” said Xi Xingjie. There are not nearly enough human artists for the Tiktok-fueled short video boom, he said. AIGC tools will significantly improve productivity for content creators.
So far, chatbots have been nothing but hype. Almost without exception, mediocre applications promise to replace human customer service representatives only leaving users crying for them. Given ChatGPT’s uncanny conversational ability, there seems to be a good reason to believe a truly human-like chatbot is about to materialize.
Not really, said people with development experience. ChatGPT might be good at conjuring up visions for a great product, says Han Sanpu of Megaview, but it has yet to demonstrate the ability to correctly digest actual product information and present it truthfully. It might be fun to watch ChatGPT babble on, but when real money is at stake, companies won’t risk their credibility. LI Di, CEO of Microsoft’s chatbot Xiaoice, said at a recent conference that ChatGPT, far from “being knowledgable,” is optimized to “appear knowledgeable.”
Liang Junzhang, of Kinzon Capital, says it takes more than technology to make a customer service chatbot. For example, a shopping site chatbot must be a fashion guru, while a Saas company needs a sleek user interface. “It’s crucial to understand the market and have big enough training datasets,” he said.
“Large models are like college students, we thought. They know a little bit of everything but nothing in depth. Smaller models, on the other hand, are experienced, highly specialized engineers,” said Xi XIngjie.
ChatGPT ended the debate by demonstrating that logic, consistency, and a firm grasp of sentiment – only possible with large models – deliver a far superior user experience.
“We are very excited about where large models can lead us, be it chatbot or content generation. Layering good use cases on top of breakthrough technology, there will be products that we’ve never seen before,” Xi said.
Although user-facing applications still attract the majority of funding – 77 percent in 2022 by some estimates – investors are increasingly warming up to open-source technologies and fundamental research. ChatGPT may further stoke interest in “foundational projects” such as algorithm training and data labeling.
Typically, tech giants build foundations and startups develop applications, the Open AI - Jasper.ai relationship being no exception. In China, where no niche application is too niche for giants, such a dynamic doesn’t hold. Alibaba’s Ant Group, for example, showed no qualms about parting with unicorn Megvii to build its own image recognition algorithm.
“It’s hard to imagine Google making niche applications, but Chinese tech giants actually do it,” said Liang Junzhang. To gain a foothold, an AI startup needs to find a big enough use case, and generate enough user data to “kick off the flywheel before giants get in.” Every step of it takes technology, money and an incredible amount of luck.
But ChatGPT may change tech giants’ approach to AI. Given the huge R&D cost – OpenAI reportedly poured US$12 million into ChatGPT predecessor GPT-3 and billions into ChatGPT – Chinese tech companies tend to build their AI products to the specifications of big clients to have someone foot the bill.
The drawback, said Xi Xingjie of China Renaissance, is that technologies developed this way are never generic enough for mass adoption.
Although a few companies including Alibaba, Baidu and Huawei have recently produced prototypes for their own large language models, Chinese tech still seems ambivalent about foundational models. In the case of deep learning framework, for example, there’s only Baidu’s PaddlePaddle that is worth comparison with Google’s TensorFlow.
“In areas that require brute-force investment with no sure short-term financial returns, Chinese companies are slower,” said Han Sanpu. “They are subject to more computational power constraints and money concerns.”
ChatGPT may inspire Chinese giants to commit, said Xi Xingjie. Foundational models will lead to brand new applications, the same way as cloud computing gave rise to all kinds of Saas applications.
“Don’t worry about collaboration with startups. The first step is to get the foundation right,” he said.
