Li Di, CEO of Xiaoice
"We are anti-'involution' and 'open source', blind open source will make the AI field very messy, and it is easy for the (AI) industry to 'drive out good money with bad money', which is not a benign format. In the past, I have always believed that in the first quarter of 2024, basically the industry pattern (large models) will be settled, and more than 100 large models will become undifferentiated. Xiaoice CEO Li Di recently told Titanium Media App.
In the past 2023, AI large model technology represented by ChatGPT has become one of the most discussed topics in the tech world, with hundreds of millions of people directly interacting with generative AI tools.
However, compared with most people's views such as "large models will lead a new round of industrial revolution" and "AI will reshape the world", Li Di is an out-and-out anti-consensus, anti-"volume" large model and computing power, and an AI Beings (artificial intelligence interactive subject) innovator.
"I'm really not the kind of person who believes in 'big models'. Li Di told Titanium Media App.
Li Di once said that the current AI technology innovation is in a period of shock, some large model products are in a state of unfalsifiability, the industry's suitable business model has not yet been established, and commercialization is facing challenges such as low-price vicious competition, reducing costs but unable to generate high profit value, etc., and has fallen into a "strange circle". The Xiaoice team is also actively looking for new business models with more valuable meaning.
On January 4 this year, Xiaoice announced that it had obtained the domestic filing of "Xiaoice large model", and officially released a number of technical products such as Rinna large model, Xiaoice clone, Xiaoice large model digital employee, as well as the landing and commercial development of self-developed large model products with different parameter scales and uses. At the same time, some top V Internet celebrities use Xiaoice clones to provide digital content, with a monthly personal income of more than 100,000 yuan and an annualized income of 1 million yuan, which has initially achieved commercial verification.
Recently, Titanium Media App and Li Di completed an exclusive conversation of nearly 90 minutes, detailing the development and changes of China's AI large model industry in the past year, Xiaoice's overseas development model, future AI technology development trends, and the relationship between AI and humans.
Li Di said that Xiaoice always hopes to create AI partners that can establish long-term emotional bonds with people, and continues to explore unique innovation paths. This year, the company will increase technology iteration and large-scale commercial verification, and bring more original AI Native products to the industry. He emphasized that emotional companionship and interesting AI applications such as Xiaoice have more long-term business value than useful AI products.
Helping Big V earn 1 million a year, Xiaoice's team opposes the "involution" of large models
Xiaoice, formerly known as the artificial intelligence Xiaoice team of Microsoft (Asia) Internet Engineering Institute, is Microsoft's largest AI independent product development team in the world, established in China in December 2013. In 2020, Microsoft announced that it would spin off Xiaoice as an independent entity and continue to maintain its investment interests.
In November 2020, Xiaoice completed the Pre-A round of financing. On November 24, Microsoft China and Xiaoice announced a strategic partnership in Beijing, joining forces to jointly launch a series of AI + cloud computing commercial solutions for customers in key industries.
On May 16, 2023, Xiaoice announced the launch of the "GPT Clone Project", and the first batch of 300 people was recruited. Today, Xiaoice has developed a variety of technical products such as "Xiaoice clone" and "Xiaoice digital employee".
Xiaoice image photo (source: provided by the interviewee)
Looking back on the more than three years since the spin-off and independence, Li Di has a lot of feelings. He told Titanium Media App that entrepreneurship in the field of AI technology is not easy to do, there are too many temptations and too few things, and the AI field is by no means a game of "winning or losing", but requires a long-term, continuous process of deep cultivation and business exploration.
"People in China are too affected by exam-oriented education, and there are only two types of people who can give real advice: one is a clear 'winner', frankly Xiaoice is not; the other is a clear 'loser', can share his lessons with everyone before dying, Xiaoice is not like this." Li Di said that many people are not suitable for starting AI companies, and one of the reasons is that entrepreneurs are influenced by external investors, which will cause the company to be too "capitalized".
Therefore, unlike the aggressive financing strategies adopted by other companies, Xiaoice's last round of public financing took place in 2022 and reached a scale of 1 billion yuan. Institutions including NetEase and GGV Capital are important investors.
"Objectively speaking, we are not 'tsundere', but we don't want to 'chase' in the AI bubble and don't want to spend more time on it. Li Di told Titanium Media App, "Investors are not your lovers, nor are they your partners. VC's investment behavior is an evaluation of its input and output, and it has an independent value judgment. ”
Li Di and his Xiaoice team present a unique attitude in the current environment: do not cater to, do not participate in "involution", do not compete in computing power, oppose the "open source" of large models, and oppose the "piling" of high computing power costs.
Of course, this does not mean that Xiaoice does not make "big models". Li Di said that as early as 2021, Xiaoice has developed a number of large-scale model technologies, and Xiaoice also made chat products such as Chat "Xiaoice" earlier, successfully enabling thousands of users to communicate. For a long time, Xiaoice has been deeply cultivating the field of "AI Beings" - AI-driven emotional interaction digital humans, and developing products such as Xiaoice clones.
Li Di admits that product forms such as chatbots are no longer the most interesting to users. "It's a shame that it's gone, and most people don't feel that strongly about any chat, or conversation, now. Why ChatGPT is popular because it works. Li Di said.
"Since 2018, Xiaoice's entire framework has emphasized her comprehensive ability, and the question is not how she communicates, but whether she has a 'circle of friends' to communicate and feedback, which is the factor that determines the future development of AI Being." Li Di told Titanium Media App that AI needs to evolve into the expression of emotion, making it easier to establish a user 'connection'. AI cloning should not only be used, but also be able to be mentally "possessive" by people.
At present, the "Xiaoice clone" developed by the Xiaoice team is the creator who clones himself and publishes it through the Xiaoice framework technology after identity authentication, so the clone has the creator's own personality, memory, knowledge, voice and appearance, and can freely dialogue, generate photos and videos, and form a group life. After half a year of testing, the Xiaoice clone has now attracted 800,000 creators, the head creator is expected to exceed one million income, and users can directly interact with the clone in Taobao.
In addition to the clone plan, in terms of AI singers, Xiaoice and NetEase Cloud cooperate with AI music creation software, Luo Tianyi and others will settle in; in terms of digital employees, new products digital interactive business cards are launched, live broadcast solutions further expand small languages, assist cross-border e-commerce; Xiaoice in third-party platforms such as Xiaoai and OPPO gradually switch to large models; in addition, Xiaoice Rinna large models have won the top of many professional lists in Japan.
Li Di said frankly that neither the Xiaoice clone nor the Xiaoice large model will adopt "open source".
"Open source will make a lot of things messy in the AI field, and it's easy to 'drive out good money'. Without my authorization, it is easy to 'go wrong', for example, if I don't authorize, many people's legitimate rights and interests will be harmed. Although the AI open source protocol is 'fuzzy', we don't intend to 'blur' it anymore. Li Di pointed out, "Once you enter the 'deep water area', you will find that whether the model is open source or not does not have a big impact, but at the same time, the computing power cost of model training continues to be high, and it is difficult to 'make profits'." ”
Li Di emphasized that the AI that the Xiaoice team wants to do must not only be useful, but also have "temperature", so that AI digital humans with emotions can change human beings. He hopes that in the next step, human and digital humans can be combined to form a new social network. "I think [that] she'll be able to develop more sustainably. ”
Large models cannot subvert all industries, and the market pattern may settle in the first quarter of this year
"In fact, in the past period, some of the facts [of AI large model technology] have been exaggerated or exaggerated. Li Di believes that AI large models are not a track, nor an industry, and everyone has some misunderstandings.
In his view, the AI model is just a "basic capability", and the cost is relatively expensive, but the "scarcity" is not so strong, and there is no high technical threshold.
Many people imagined that AI could disrupt all industries, but it didn't come to fruition. For example, the Wonderful Duck camera has not subverted the "hippocampus", and the AI large model has not subverted the traditional industrial system.
"First of all, at the beginning of 2023, everyone thinks that the larger the scale of the model parameters, the better, but this is not the case; secondly, many people think that through the model with huge parameters, the ability will continue to 'emerge', and AGI (general artificial intelligence) will come soon, but this does not conform to the scientific law, so we choose to do a 'hybrid model'; third, many people think that the computing power will decline rapidly, and the business model will be established with the decline of computing power, but we judge that the cost of computing power will not decline, and the cost of computing power will continue to be at a high level for a long time in the future. Li Di said that these are the gaps between fantasy and reality.
Therefore, Li Di believes that the AI model cannot subvert all traditional industries, it is just an auxiliary tool.
He pointed out that the current AI large model track is too "involuted", which is easy to cause chaos, and the industry may have the phenomenon of "bad money driving out good money". Li Di predicts that after the process of "big waves and sand", it is expected that by the first quarter of 2024, the AI large model industry pattern may be settled.
For the industry trend in 2024, Li Di summarized three points: first, the cost of computing power will not decrease, and data training will continue; second, more large model companies will withdraw from the market, and it is difficult to merge large model companies, "the merger of two large models is meaningless"; third, the digital human model of AI Beings will also usher in commercialization, such as on January 9, Xiaoice and DingTalk announced that they will jointly explore comprehensive cooperation in short video, live broadcast, video conference, AI assistant and other fields.
Li Di said frankly that Xiaoice will not directly sell large model APIs (application programming interfaces) by "reducing costs and reducing prices", but will pursue "a packaged product form". At the same time, Xiaoice pursues that the digital human products provided to customers should have high added value and make customers feel the real "value". "I think everybody should do something meaningful and not go for the 'roll' model. ”
Li Di once gave an example that at present, among the top ten car brands in China, more than 60% of the dialogue systems in the intelligent cockpit of cars are provided by Xiaoice. At the same time, 90% of China's financial text generation is also developed by Xiaoice. He believes that Xiaoice will continue to become the top player in the domestic AI competition.
Li Di stressed that the large model is only one of the company's many technical iterations, and Xiaoice will not do China's ChatGPT. In addition, Xiaoice has a complete AI technology framework and adheres to technical originality and independence in key technologies.
Talking about OpenAI's recent technical research, Li Di said, "Frankly speaking, I believe that OpenAI's Q* research exists, and many people are also studying this direction, but has he researched it? ”
Li Di emphasized that all of us want to live in a magnificent and thriving era, but the reality is cruel, the commercialization of AI technology still has a long way to go, and the widespread application of generative AI technology will take longer to prove.
"I don't know why we always have a 'time is not me' mentality. When the big models appeared, everyone seemed to think that this would take the industry forward by leaps and bounds, experts even warned that if you don't learn AI now, you might not be able to find a job in the future, and companies were worried about falling behind their competitors and fearing that if they didn't chase it, they would 'fall behind' and lose out at the starting line. However, this is not the case. In my opinion, the status quo is like playing a game of minesweeper, where everyone tries and judges in their own area, but can't tell exactly whether the next brick is a 'mine' or a 'safe grid'. When a new area is revealed, everyone will suddenly open up and understand. This is the nature of the world – the unknowns we face are always beyond what we know about the future. Li Di said.
(This article was first published on the Titanium Media App, author: Lin Zhijia)