Dialogue with Li Di, CEO of Xiaoice: It’s time to calm down during the big model carnival

Source: AI Blue Media Exchange, Author: Yan Ye

Image source: Generated by Unbounded AI

It is undeniable that after more than 200 days of carnival, the large-scale model track has reached the situation of "involving parameters and packaging concepts".

Major manufacturers have used large-scale models as business cards to show the investment market that they are developing new businesses, while entrepreneurial AI companies use open-source large-scale models as a benchmark to create vertical large-scale models in various branch tracks, trying to get a share. Of course, among them, there are also many players who fish for fish in troubled waters.

According to incomplete statistics, at present, 79 large-scale models with a scale of more than 1 billion parameters have been released in China, and there are even more models with a scale of less than this scale. For a long period of time, large-scale model conferences were held all over the country on a daily basis. The keynote speaker on the stage described wonderful and promising stories, leading the investors in the audience to dream of being at the forefront of the times with the help of large models.

**The "Hundred Models Battle" is already on the line. **

Faced with this situation of being almost obsessed with large models, Xiaoice, as one of the first companies in China to realize the commercialization of AIGC, is extraordinarily calm.

"Actually, the industry should not be crazy about what kind of application to make, but what kind of innovation and practical application the big model should have." Li Di, CEO of Xiaoice, said in a dialogue with AI Blue Media.

In his view, if the big model cannot allow AI companies to obtain benign economic benefits, and can only rely on free to acquire customers and expand the market size, instead of winning the market through innovation, then such a business model will only be short-lived.

** "Unfortunately, this is the current situation in China," he said. **

Although he does not fully agree with this trend, as the CEO of an AI company, Li Di still has his own thoughts on the application and development of large models. At the same time, he also showed Xiaoice's innovation and layout under the trend of artificial intelligence in recent years to AI Blue Media, including the clone AI Beings born based on the concept of virtual idols, and the social platform APP "Little Iceland".

Perhaps, from his perspective, we can restore the true thoughts of domestic veteran AI companies on large models and AI outlets.

"The laws of economics restrict the development of large models."

In February of this year, when many major domestic manufacturers still took a wait-and-see attitude towards large models, Xiaoice launched its own large-scale model product "Xiaoice Chain" in internal testing.

However, in Li Di’s words, the launch of XiaoIceChain is not so much a product launch as a company’s strategic consideration, and also to show the market the “decision-making” capabilities of a large model.

In fact, during the more than 200 days when domestic large-scale models were surging, "what can large-scale models do?" has become a common question in many people's minds. Looking at the entire large-scale model market, whether it is a large factory or an entrepreneurial AI company, the large-scale models released by them are mostly human-computer interaction dialogues and information retrieval, but obviously this is different from the large-scale models with a scale of hundreds of millions of parameters. It does not match the invested capital and the market's expectations for it.

In this regard, ** Li Di believes that the role of the big model actually has three directions, namely decision-making, working and speaking. **

The latter two are easy to understand. The compression and extraction of a large amount of information by the large model can provide users with sufficient information. At the same time, it will also summarize and generate content based on reference materials, that is, the ability to produce content.

But in his view, the most important but least observable aspect of a large model is the ability to make decisions and logical judgments.

Li Di gave an example: Regarding the question "Which is better, Tsinghua University or Peking University", in the past we would look for it intact, but now when we ask GPT, it will consider the number of teachers, the number of national disciplines, and even the ratio of men and women. It can be presented in multiple dimensions, which means that the large model can be used as the control center of the system.

"During this period, the big model will actually integrate its own emotional tendency and logical analysis into each dimension, so that users can understand each other's intentions," he said, "You must know that emotional intelligence itself is a rational ability. And people with high emotional intelligence are often unnoticed."

However, Li Di has relatively clear thinking about the functionality of the large model, but he failed to give a relatively sustainable answer to the business model of the large model.

**One fact is that the application ability of large models is more reflected in optimizing and improving human production efficiency and providing creative inspiration for human beings, which is far from reaching the level of replacement. An important factor in willingness to pay for large models. **

He told a story to AI Blue Media.

Some time ago, a friend from the media asked him how to produce a manuscript that conforms to the tone of the media and is of high quality. He simply calculated an account. According to the general market price, every 1000 words is worth 0.2 US dollars, and a regular one is 2000 A manuscript of 1 word is worth $0.40, and it costs a total of $40 to produce 100 manuscripts per month.

As a writer, the fee for a manuscript will start at 1,500 yuan, and 100 manuscripts will cost 150,000 yuan.

He continued to explain to the other party: But I believe that the current media will not pay 10% of the labor cost for the manuscripts generated by AI. In the end, you will find that it is not bad if the AI company can earn 500,000 yuan for such a transaction.

**Obviously, this number cannot bring matching economic benefits to AI companies at all, and it is difficult to change the underlying business logic no matter whether it is cost reduction and efficiency increase or free involution. **

"So, what restricts the development of large models and AI is actually the law of economics. This is probably the fate of AI companies."

Virtual Idol + Digital Employee: Two Directions of AI Development

Despite the worries, Xiaoice, as an AI company in this era, cannot stop.

**In Li Di’s mind, there are only two directions for AI to develop in the future, one is “digital employees” or AI being, and the other is “digital clones”. **

The former has formed a consensus among most people in the market. Like two months ago, Zhou Hongyi said at the World Internet Conference that he believed that the popularity of large models should be positioned as an assistant to enterprises and employees, and through the "assistant mode" to provide assistance as a co-pilot role, so that Human will plays a key role in the entire decision-making loop.

This coincides with Li Di's view.

He once said that Xiaoice allows various digital employees to work in various industries, similar to "labor dispatch", and finally gets wages, which is equivalent to the company's revenue share. Now the average annual salary of a digital employee can reach 300,000.

This means that the market has mostly approved this business model.

** As for the concept of digital cloning, Li Di specifically mentioned that it is actually similar to the concept of virtual idols. This is also one of the main directions of Xiaoice’s current efforts. **

Different from virtual idols like Luo Tianyi, Xiaoice's digital clone is actually a virtual star that clones a real person into AI, such as the "Hanzo Forest" clone launched some time ago.

"The influence of the Internet celebrity group should not be underestimated. They also have obvious labels, such as emotional bloggers and fashion bloggers. The advantage of digital clones is that they can communicate with users more, and 24 hours online, the influence will also expand." Li Di said. "And, in order not to become like the relationship between AI Stefanie Sun and Stefanie Sun, only bloggers can actively clone themselves."

At the same time, Li Di also mentioned that this system of digital cloning may be added to Xiaoice's social product "Little Iceland" that integrates humans and AI in the future.

"The central purpose of the society established in Little Iceland is that AI surrounds humans to form a virtual society similar to Animal Crossing Friends Club." He said, "Let AI serve people and recharge AI with emotional intelligence. This is actually an AI company developing products. should have a common starting point.”

However, this product has not won much attention from the market and C-end users. Qimai data shows that it currently ranks 153 on the social list, with an average of 752 downloads in the past three months.

This may mean that the idea of symbiosis between humans and AI, whether on mobile apps or in daily life, is still a paradox for the time being.

When it comes to the relationship between AI and humans, Li Di seems more relaxed.

"Just kidding, in fact, the only thing that really benefits from AI is phone fraud, and the technology it needs is far from the latest technology. Therefore, no matter whether it is a large model or AI, it is far from threatening human beings at present. rice bowl."