Exclusive Interview With Xiaoice CEO Li Di: is "Bad Money Driving Out Good" ...

Image of "Xiaoice" chatbot

BEIJING, January 11 (TMTPost) -- "We are against involution on AI large models and 'open-sourcing', because blindly open-sourcing can lead to a chaotic AI industry dominated by inferior products. I have believed that by the first quarter of 2024, the dust will settle in the industry, especially for large models, and more than 100 large models will become undifferentiated," said Xiaoice CEO Li Di recently to TMTPost App.

Xiaoice Company, formerly known as AI Xiaoice Team of Microsoft Software Technology Center Asia, is Microsoft’s biggest independent R&D team of AI products in the world. It was set up in China in December 2013.

In 2020, Microsoft announced the spin-off of Xiaoice as an independent entity while retaining investment rights. In November 2020, Xiaoice completed its Pre-A round of financing. On November 24th, Microsoft China and Xiaoice company announced a strategic partnership in Beijing to launch a series of AI + cloud computing commercial solutions for key industry clients.

AI large models, represented by ChatGPT, have become one of the hottest topics in the tech industry in 2023. However, contrary to the prevalent views praising "large models leading a new industrial revolution" and "AI reshaping the world," Li is a true innovator as he opposed the involution of large models and computing power, and proposed AI Beings, the AI interactive entities.

Li said that AI technology innovation is still unfolding, with some large model products in an unverifiable state. The industry lacks suitable commercial models, facing challenges such as the vicious low-price competition without generating high-profit value. The Xiaoice company team is actively seeking new and more meaningful business models.

On January 4, 2024, Xiaoice company announced that it had obtained domestic registration for the "Xiaoice Large Model" and officially launched several products, including Rinna, Xiaoice Clones, and Xiaoice digital employees. Meanwhile, some top-tier internet celebrities on the Xiaoice platform have generated monthly incomes surpassing 100,000 yuan by creating their own AI clones, with annual earnings reaching 1 million yuan, marking an initial stage of commercial success.

Recently, TMTPost conducted an exclusive 90-minute interview with Li, discussing the developments in the Chinese AI large model industry over the past year, Xiaoice's overseas development model, and future trends in AI technology development, etc.

Li said that Xiaoice company has always aimed to create an AI partner that can establish a long-term emotional bond with people, which is more valuable in the long run than a useful AI product. This year, the company will launch more original products.

Looking back over the past three years since the spin-off, Li emphasized that the field of AI is by no means a "winner takes all" game. Instead, it requires a long-term, continuous exploration.

Unlike other companies adopting aggressive financing strategies, Xiaoice's last public financing round took place in 2022, reaching a scale of 1 billion yuan. Key investors include institutions such as NetEase and GGV Capital.

"Investors are not your lovers or your partners. The investing by venture capital (VC) is an evaluation of their input-output, and it has independent value judgment." said Li.

In the current environment, Li and his Xiaoice team exhibit a unique attitude: not catering to trends, not participating in meaningless involution regarding computing power, opposing the "open-sourcing" of large models, and the accumulation of high computing costs.

This doesn't mean that Xiaoice is not involved in "large models." According to Li, Xiaoice had been developing various large model technologies back in 2021. Earlier, Xiaoice successfully had launched chat products like "Xiaoice," allowing thousands of users to communicate and interact. And it has been deeply involved in the field of "AI Beings"—emotionally interactive digital entities driven by AI—and has developed products such as its AI clones.

Li admitted that products like chatbots are no longer the most interesting to users. "Since 2018, Xiaoice's entire framework has been emphasizing its comprehensive abilities. The key is not how it converses, but whether it has the ability to communicate and provide feedback. This is what determines the future of AI Beings," said Li. He believes that AI needs to evolve into expressing emotions, making it easier to establish a “connection” with users.

Currently, the creators, after identity verification, can create and release their AI clones, which possess the creator's personality, memories, knowledge, voice, and appearance, allowing for free conversation, generating photos and videos, and forming a community. After six months of testing, Xiaoice clones have attracted 800,000 creators, with top creators expected to have over a million yuan in income, and users can directly interact with the clones on Taobao.

In addition to the Clones program, in the field of AI singers, Xiaoice collaborates with NetEase Cloud to develop AI music creation software, with artists like Luo Tianyi joining. In the realm of digital employees, a new product called Digital Interactive Business Card has been launched, and live streaming solutions have been further expanded to include non-English/Chinese languages, aiding cross-border e-commerce. XiaoAi, Xiaoice's presence in third-party platforms like OPPO, is gradually transitioning to large models. Additionally, Xiaoice's Rinna large model has achieved top positions in multiple professional rankings in Japan.

According to Li, neither Xiaoice clones nor Xiaoice large models will adopt an "open-source" approach.

Li emphasized that large models cannot disrupt all industries. "In fact, over the past period, some things about AI large model technology have been hyped or exaggerated," Li said. He believes that AI large models are not a specific track or industry. There are some misconceptions in people's thinking.

In his view, AI large models are just an expensive"basic capability," but not so scarce, and there is not a high technical threshold.

"First, at the beginning of 2023, everyone believed that the larger the model parameter scale, the better, but that's not the case. Second, many people thought that through models with a large number of parameters, capabilities would constantly 'emerge,' and AGI (Artificial General Intelligence) would arrive quickly, but this goes against scientific laws, so we choose to develop 'hybrid models.' Thirdly, many people thought that computing power would quickly decrease, and business models would be established as computing power decreases, but we think that computing power cost will not decrease. It will remain high for a long time," Li said. He sees these as gaps between fantasy and reality.

He pointed out that the AI large model competition is too fierce, which could easily lead to chaos, such as the phenomenon of "bad money driving out good.” He predicts that the dust will settle in the AI large model industry in the first quarter of 2024.

Regarding industry trends in 2024, Li summarized three aspects: first, computing power costs will not fall, and data training will continue; second, more large model companies will exit the market, and large model company mergers will be difficult; third, the digital human model, such as AI Beings, will also reach commercialization. For example, on January 9, Xiaoice and DingTalk announced a comprehensive collaboration in short videos, live streaming, video conferencing, and AI assistants.

Li admitted that Xiaoice will not directly sell large model APIs by "reducing costs and lowering prices." Instead, they pursue a "packaged product form." Additionally, Xiaoice aims to provide digital human products with high added-value to customers. "I believe everyone should do something meaningful rather than just churning out models," Li said.

Li gave examples of Xiaoice’s prevalence: currently, among the top ten auto brands in China, more than 60% of the conversational systems in intelligent car cabins are provided by Xiaoice. Moreover, 90% of financial text generation in China is also developed by Xiaoice. He believes that the company will continue to be a leading player in domestic AI competitions.

Li emphasized that large models are just one part of the company's many technological iterations, and Xiaoice will not create China's ChatGPT. Additionally, Xiaoice has a complete AI technology framework and it insists on technical originality and independence in key technologies.

Regarding OpenAI's recent technical research, Li said, "I believe that there is a Q* (Q-star) research, and many people are researching in this direction, but has it been achieved? Currently, it does not exist."

Li emphasized that the commercialization of AI technology still has a long way to go, and the widespread application of generative AI technology needs more time for validation.

(This article was first published on TMTPost App, Author: Lin Zhijia)