Dialogue with Li Di, CEO of Xiaoice: Large-scale models will cool down by 2024

Original Source: The True Story Research Office

Author: Shi Yuhan

Image source: Generated by Unbounded AI‌

The emergence of ChatGPT is considered to be the last chance for the development of the Internet industry. Large factories, colleges and universities, and individuals have entered the large-scale model market, and the "100-model war" is in progress.

However, calm down and look directly at this wave. With high investment, scarce talents, homogeneous competition and still unclear business model, many doubts about large models have also begun to surface. ‍‍

Recently, "True Story Lab" had a conversation with Li Di, CEO of Xiaobing Company. Xiaoice is the earliest robot with emotional interaction function in China, and Xiaoice is one of the earliest enterprises in China to realize the commercialization of AIGC.

Li Di has a more calm and objective view on the current boom in large-scale models.

The following is the transcript of the conversation:

**Q1: After the launch of ChatGPT, major domestic manufacturers and artificial intelligence entrepreneurs followed up. The industry is very lively, but everyone seems to see no difference. What do you think? **

A1: There are currently at least 70 large-scale models in China, but they are all homogeneous and cannot achieve differentiation. Because everyone's training data is similar, the training method uses the method in the paper published by OpenAI, and they all use GPT for training directly, using GPT as a teacher.

The teacher, training data, and training methods are all the same, how to widen the gap?

All major manufacturers come to make large models, and their starting point is not to seize opportunities and opportunities, but to defend themselves. If you don't have your own big model, you will be suppressed by competitors, otherwise you will have the opportunity to suppress others.

**Q2: Do you think that the current business model like ChatGPT, which uses subscription fees, has development prospects? **

A2: First of all, it is undeniable that compared with previous years, the demand for artificial intelligence has definitely increased. But with so many big models, has it achieved scale effect? I don't think so.

Most of the business models explored by artificial intelligence so far are hardly successful. Just like ChatGPT, its essence is actually "selling words": how much is a word, how much is a sentence. But the value of these replying utterances is not differentiated. For example, when users chat with it, to some extent its words are worthless. But if the user is looking for advice, then its reply is very valuable.

Today we discuss the business model of AI, which involves the value created by **AI and the value it gets, but the gap between the two is very large. **

For example, the face recognition technology that the industry has been involved in in the past has made the country and society safer, and its value is obviously very high. However, after face recognition technology is embedded in various hardware, the value of the technology itself has not been rewarded accordingly. In the end, some companies began to do system integration, software and hardware integration, and sold the hardware stack of equipment.

The current business model of artificial intelligence is to make technology into infrastructure such as water, electricity, and coal, which can be divided into industrial electricity or domestic electricity at most. However, the different values created based on different needs have not received differentiated returns.

Q3: What are the limitations of the business model of charging technology call fees?

**A3: ** Selling infrastructure such as water, electricity, and coal depends on monopoly to keep profit margins. But as far as my experience is concerned, no one can achieve a monopoly in technology. Therefore, it can only rely on cost savings to obtain profit margins, but this is not only limited, but also has no possibility of rising.

Even if the so-called Moore's Law is followed and the price is lowered, competitors will quickly make up for the profit margin. In the end, there will be a price war, and then everyone will start free. This seems to be a good thing, but it actually limits the development of the entire artificial intelligence industry.

**Q4: Has Xiaoice explored different and more effective business models? **

**A4: **Xiaoice's business model is somewhat special, we use revenue share (income sharing). We do not provide a technical interface to sell content calls, but package them into "people" with different abilities based on technology, that is, digital employees. We let various digital employees go to work in various industries, similar to "labor dispatch", and finally get wages, which is equivalent to the company's revenue share. Our average annual salary for a digital employee can reach **300,000. **

For example, in February this year, Japan’s Xiaobing (Rinna) and Netflix jointly launched an AI-generated animation micro-short film "Dog and Boy", which lasted for more than three minutes. Although the share is limited, as a film and television creator, Rinna's productivity is unlimited, and we can get corresponding benefits for every film and television work in the future.

**Q5: Did Xiaoice take a detour before exploring a business model like revenue share? **

**A5: **There must be some in the early stage. In 2017, Rinna began to cooperate with Lawson convenience store to help Lawson sell coupons. Based on Rinna's superiority in emotional interaction, its promotional effect is very good. And consumers use coupons to shop offline, which also helps Lawson get more revenue.

But we did not share these revenues, because at that time our analysis of the artificial intelligence business model was that we should provide APIs and charge money for each call. This income is very small, and the sales brought by Rinna's sales of coupons are huge, and the two are obviously not directly proportional. **

**Q6: The Metaverse has been popular for two years, and it seems that the fever is a little bit lower now. One reason is considered to be that its experience has not yet amazed customers. Will the heat of the large model also fade away due to the lack of good user experience? What do you think are the advantages and shortcomings of my country in the global competition of large-scale models? **

A6: There were many surprises in the AI industry last year. For so many years since the development of artificial intelligence technology, it has been like a tide, and there will be a tide and an explosion every few years. And these so-called "explosions" are only because they have been noticed by the public, which exceeds the public's cognition or expectation of artificial intelligence. They are not the end.

Just like AlphaGo back then, countless people exclaimed that the world would be changed, but after so many years, no great changes have taken place. Whether it is a large model or AIGC, it is a new breakthrough in the technical bottleneck period, and they will have their own bottlenecks in the future. ** The distance from the real AGI (General Artificial Intelligence) will be clearer after several rounds of similar technological evolution. **

There are still many problems in the large model that have not been resolved, such as accuracy and high cost, so Xiaoice uses the concept of hybrid model, and there are nearly 1,000 large, medium and small models organically combined in the framework of Xiaoice Work together to support the operation of AI Being. The advantage of this is that the cost is relatively low, the speed is faster, and it can be guaranteed to be accurate enough and delivered for commercial use.

At the same time, Large model has not been accurately defined until now, how many parameters can be called a large model. In the first half of the year, everyone was guessing that the larger the parameter, the better the effect. Later, it was found that a model with a smaller parameter can also achieve the same effect. So now more and more people in the industry are talking about hybrid models.

If the study of large models is compared to learning a language, it is a bit like learning Japanese. It is easy to get started and difficult to master. Entrepreneurs can see results quickly, but find that there are too many problems to solve.

The current large-scale models are seriously homogenized. ** There are not so many large models on the market. The fever should subside by 2024, and it will be discovered who is swimming naked on the shore. **

Finally, there are actually no technical barriers between large models. Its technical concept has existed for many years, and many people in the industry are applying it, both at home and abroad. It's just that in the process of fine-tuning (fine-tuning), developers have different degrees of concentration and devotion.

I personally think that OpenAI has really achieved this effect with the spirit of craftsmanship, so there is a certain "time barrier", but this is two different things from "technical barriers". **In terms of artificial intelligence technology, there is no huge gap between domestic and foreign countries. For example, chips and operating systems are very powerful in China. The gap lies only in whether they can endure loneliness, dare to grind, and do innovative research. **