AI Sogou

AI Talks with Xiaochuan Wang: Sogou founder's high hope in large-scale model

Thomas Luo

posted on April 22, 2023 0:28 pmEditor : Wang Boyuan

Editor's note: This year, PingWest has launched a series of discussions with China's leading pioneers and trailblazers in generative AI, amidst the rapid growth of ChatGPT-like services in China. Through these exclusive and non-exclusive talks, we aim to provide insight into the potential of Chinese AI and the thought processes of its creators.

In the third conversation, we spoke with Xiaochuan Wang, the founder and former-CEO of Chinese search engine Sogou. On April 10th, Xiaochuan Wang announced his AI company "Baichuan Intelligence (百川智能 in Chinese)" after leaving Sogou for more than a year. Baichuan Intelligence has not yet released an AI product but Wang has set the goal of developing a Chinese equivalent of OpenAI both in fundamental structure of large-scale model as well as upper-level applications. On the day of announcement, the company has already raised $50 million in funding.

The interview was originally conducted in Chinese and can be read here.

During the industry boom of the past two months, Xiaochuan Wang chose a tactical low-key approach. Xiaochuan Wang told PingWest that Baichuan Intelligence has initially built a team including technical backbone from Sogou, Baidu and other companies with $500 million already in pocket.

Even a month ago, Baichuan Intelligence had already started training a 500 billion-level LLM in a situation where they had not formally entered into employment and were completely remote collaboration. The progress was "smooth" and "even the teachers from Tsinghua were surprised", according to Wang.

In the conversation, you can feel the energy that the AGI business brings to this 45-year-old entrepreneur.

This energy comes partly from the "sexiness" of AGI itself, which is large enough and hits the historical civilization process of mankind; and partly from the aggregation of resources, not only top universities and top capital are throwing olive branches at him, but also talents that could not be attracted during the Sogou era are now gathering around him. He frankly told PingWest that he had "never enjoyed such treatment" before.

Xiaochuan Wang is indeed worthy of this kind of attention. He is a top student at Tsinghua University, the soul of Sogou, and a leading role in the generation of AI entrepreneurship. Besides the area of IT, he is also a passionate biology enthusiast and is accustomed to viewing the world from a biological perspective. This coincides with discussions about AGI.

In the interview, he has his own unique understanding of AGI business and maintains a polite frankness towards things he dislikes:

He opposes "platter-style" entrepreneurship, believing that a company, a team, and technology itself are all "living organisms" and therefore require a growth process. He opposes trying to make a big splash and insists that "people" are the core of the project and that all product design should be based on a human-centered approach.

He is against “Idle talk” about AGI start-ups using rough classifications and believes that China will have its own OpenAI. He believes that many people rate Chinese AGI companies as either "too low or too great" and against extreme opinion on the industry.

That’s why he choose to adhere to his past "seriousness."

For example, he chose a very restrained first round of financing; he insisted on placing scenarios and large language model training in equally important positions; he believes that AGI is not a simple reproduction of the industrial revolution, but rather the Renaissance. He also does not believe that China's large language models must pursue the highest level of ChatGPT from the beginning, "the most important thing is to provide valuable services. If I can't make an elixir of life, can't I make a longevity pill first? Is that not good?"

For Baichuan Intelligence, Xiaochuan Wang believes that "by the end of this year, we will be the best in China."

Here is a transcript of the conversation, and conversation was taken before Wang's announcement:

"A big model that is competitive with GPT-3.5 will be launched by the end of this year”

PingWest: Has the company come up with a name yet? It's not going to be "Five Seasons Intelligence", is it?

Xiaochuan Wang: It's called "Baichuan Intelligence", which means "embracing all rivers, gathering knowledge". Because LLM is actually a process of gathering and learning all knowledge in the world based on language as the core.

PingWest: I see you are also betting on your name.

Xiaochuan Wang: Yes. The startup capital is currently supported by friends, and some of our old teammates have also returned. You can understand that "Baichuan" is not only connected to the world, but also fully connected to the team and resources between me.

PingWest: Can you share with us the progress of financing?

Xiaochuan Wang: Actually, it's not really financing. The first round is $50 million. It's the support from friends and personal investment including my own. Financing plan for institutions will be launched in the near future.

PingWest: But many investors are interested in this.

Xiaochuan Wang: At least the money is enough for training a good hundred-billion-level model until the end of this year. If we want to go further on AGI, we will need a great deal of money.

PingWest: We learned that OpenAI has raised $10 billion, but there’s no actual spending on the large-scale model until the GPT-3 emerge, which is also after 2021.

Xiaochuan Wang: Money is crucial, but what's more important is the specific operation of the people inside. The enterprise is a living organism, and it needs a growth process for technology, team, and business. If people and money are forced to be put together like building blocks, it is easy to encounter difficulties.

PingWest: How is your team currently structured?

Xiaochuan Wang: We have the original team from Sogou, as well as leaders from other major companies like Baidu, and core members from other startups. By the end of April, our team will have 50 people. This is actually a great complement. The original Sogou team acheived the first 10 billion-north-parameter model domestically in 2021. With the addition of other leaders and their members, we are a diverse and experienced team.

PingWest: What percentage of your team is from Sogou?

Xiaochuan Wang: It's less than half overall and mainly to quickly establish a foundation. As more people joining us, the percentage will become lower. We have many "young generals" in our team, who are both experienced and young. Actually, I personally have not enjoyed this kind of treatment. You have to understand that it was difficult for Sogou to recruit people in the last few years. The talents may willing to talk to you but not work with you. It's actually because what you were doing at the time wasn't attractive enough.

PingWest: Then let’s talk about “attractive” thing. How is the training going for the large-scale models?

Xiaochuan Wang: We have started training a model with 50 billion parameters, and the progress is going smoothly. When we spoke with professors from Tsinghua University, they were surprised by our progress. We will release a large-scale model comparable to GPT-3.5 by the end of this year, and we are confident that we will achieve the best results in China by the end of the year.

“Some people in the industry have never been right about their views on the future”

PingWest: There is an opinion that after GPT3.5, large language models are no longer a suitable option for startup companies anymore, as it is not meaningful for them to follow up and create another large language model. Because the key would still be pulling all the weight together to achieve a greater goal.

Xiaochuan Wang: Some people in the industry have never been right about their views on the future, and they have always lived in a parallel universe.

large language models require a bit more time and seriousness, and money is not the only important factor.

We see that Tsinghua students are still very outstanding, and their team structure is relatively simple, while many large companies are trapped by their own complicated personnel.

PingWest: So some large companies with accumulated technology actually have inherent problems.

Xiaochuan Wang: You first need a purely simple and focused organizational structure, starting from the vision and continuously upgrading the team, like the process of a child growing up. It is unrealistic to pull a child into stardom on the first day. Some companies have the foundation of a large language model, but if they don't practice well, it will be even harder to go to battle.

PingWest: We see that many companies that have recently launched large-scale model products in China and the US have a background in search, such as Google, Baidu, and 360. You personally are one of the people who developed Sogou. Do search companies really have some innate technological advantages?

Xiaochuan Wang: Because today's focus is on language models, search companies do have inherent advantages in language. Additionally, search companies have comprehensive experience, from data cleaning and model conversion to providing services online and even experience in data security. The organizational structure and evaluation system of the company are also relevant. Therefore, search companies do have an advantage at the beginning.

PingWest: There’s a saying, "large language models are alchemy, prompts are spells". This can be disillusioning for many AI practitioners, because it’s hard to foresee a qualitative change. As a scientist in AI, how do you evaluate your progress and achievements?

Xiaochuan Wang: We still have confidence because this thing is valuable. Therefore, even if we don't stimulate as much intelligence as OpenAI in a short period of time, we can make it the best enterprise in China and provide sufficient services.

It can serve many existing scenarios. For example, integrating it into search would be much more effective than traditional search engines; integrating it with medical data would make doctors more reliable.

We will always pursue the ultimate in intelligence. However, the difference between us and ChatGPT is that we need to pursue whether the landing scenarios are needed or not. You can develop the "Magic Pill", and I might be interested in the "Elixir".

PingWest: Currently, we are still working on a general large language model, and OpenAI rarely mentions scenario-specific issues and provides limited interfaces. So, is there a contradiction between the goals of "Elixir" and "Magic Pill"?

Xiaochuan Wang: In the past, there was a saying that "we should make China's OpenAI, not just China's ChatGPT," which many people thought was advanced, but this statement is incorrect.

Bill Gates called ChatGPT "as important as the internet", while Jensen Huang called it "'iPhone moment of artificial intelligence". However, these two statements refer to different things.

"As important as the internet" refers to the significance of large language models. This large language model "can read, write, and understand" and can generate a lot of imagination. Jensen Huang refers to the changes in human-computer interaction and relationships brought about by ChatGPT. The latter focuses more on changes in people rather than technology.

In fact, ChatGPT represents a new human-computer interface and entry point. Today, it seems that Microsoft has only wrapped it up as one of its tools, just like how Google was just a search box for Yahoo when it first came out. However, in the future, ChatGPT will have even greater development potential, so in my mind, Microsoft Office may not be needed anymore in the future.

PingWest: I agree that Microsoft Office may not be needed in the future, but what about search engines?

Xiaochuan Wang: I think one can be called GC (Generate Content) and the other is IR (Information Retrieval). These two models are like relativity and quantum mechanics respectively. Google represents IR and knows a lot of details but is not abstract. The representative of GC is ChatGPT, which compresses into a model and loses all the details. It is not updated in real-time and integrates all knowledge together.

In China, I think there is an opportunity to integrate the retrieval and search models into Chat and form a base with search capabilities. For us, our goal from day one is to integrate this into Chat. However, we are not trying to create a new search engine, but instead embed search capabilities into Chat.

“Their understanding of China's AI large language model is either too shallow or rate it too high”

PingWest: Some people believe that one of the reasons for OpenAI's success is that it has created a third type of organizational structure outside of businesses and schools (OpenAI Nonprofit). You are a leading figure in the entrepreneurial wave of AI segmentation and are now involved in large-scale model entrepreneurship, connecting with many school resources. How do you view the potential changes in organizational structure brought about by new businesses?

Xiaochuan Wang: I think observing the company's core is secondary because you can't see the complete picture from the outside. It's better to see what it offers externally. For example, OpenAI initially provided AGI cognitive abilities, a type of "knowledge" similar to that in universities. But once it became a company, it began providing products and services.

Therefore, large-scale models do require both research and product capabilities, and engineering capabilities are in the middle. It can only be applied after discovering "knowledge." OpenAI has already broken through the ceiling in discovering "knowledge." So our research pressure is relatively small, and we need to focus more on learning how to apply it.

Secondly, the difference between China and the United States is that we have many excellent talents who are actually fixed in universities. Therefore, companies need to have good cooperation with them. So you can see that many senior teachers at Tsinghua University, even academicians, will provide us with a lot of support, which will also reduce our research pressure.

PingWest: When you release technical achievements such as models, do you also consider publishing papers?

Xiaochuan Wang: We may publish papers with universities but I don't think that's the key to this battle. Of course, we need to do our own research today, but a lot of it will also be done in universities. It is necessary to rely on universities, and many of our talents come from universities.

PingWest: When discussing the competition between China and the United States in large-scale modeling, people tend to talk about the issue of language corpus. Many people think that English corpus is naturally better, while Chinese may be relatively limited. How do you view this issue?

Xiaochuan Wang: Currently, we use both Chinese and English, and there are also many good sources of English corpus. It's like teaching children, ultimately relying on developers to use many mechanisms to adjust the source or proportion of the corpus.

It's like the father speaks English and the mother speaks Chinese, but the content is the same. It doesn't mean that if he learns medical knowledge in Chinese, he won't be able to understand it in English. His knowledge level is the same.

PingWest: What about the issue of computing power?

Xiaochuan Wang: For us, there won't be too much of a problem.

As I said earlier, we don't need to chase after the "Elixir" on the first day. We just need to keep getting closer on the road.

PingWest: This is still a "smart Chinese internet", and we can make it better.

Xiaochuan Wang: I think people who have never been to war have some idealistic things. Their understanding of China's AI large language model is either too shallow or rate it too high, which is unrealistic for China's current AI large language model.

PingWest: Yes, we are facing a complex thing.

“everyone's discussion on the "endgame" is quite superficial”

PingWest: You previously mentioned on Weibo that China should make its own OpenAI. You were the only one who mentioned the phrase "patriotic heart" in this matter. How do you explain these four words?

Xiaochuan Wang: When I was studying biology, I realized that I and the cells are both living organisms. Therefore, I gradually accepted a concept called "Being". Regardless of a country, nation, enterprise, individual, or even culture, they are all "Being". And I am also a part of a larger "Greater Self". The latter has a complete sense of belonging, and even eternity.

"Who you are" is closely related to what you do, and conversely, what you do can also verify who you are.

There is a saying about everyone thinks for themselves, but the world is a bigger place for everyone. This is a feeling of integration. It is actually part of our answer to the ultimate question.

PingWest: Will this make your big model different?

Xiaochuan Wang: We have similarities with the global community, but there are also parts that repel each other. But first of all, we have basic expectations, which is to make China better. You cannot do anything harmful to China.

PingWest: How do you view the "endgame" of AGI?

Xiaochuan Wang: In fact, everyone's discussion on the "endgame" is quite superficial. We should first ask what the world looks like at the "endgame", and if we only talk about AGI as a foundation, there is no visual sense.

The logic of the past industrial revolution was that the more detailed the division of labor, the higher the efficiency and the larger the scale. But in front of the door of intelligence, we may have "reverse social division of labor".

Machines empower people universally, more like the democratization of information access after humans have search engines. And now we have the democratization of knowledge and service acquisition.

People will eventually be liberated from work. So the future of the industry will become flatter, depth will be shallow, and breadth will be expanded. People's work will be more diversified, and they don't necessarily have to be strung together in a production line like they are now to get results. So it is a "reverse social division of labor".

Therefore, AGI can bring more equality to people. It may be more like a Renaissance, redefining what humans are.

PingWest: It seems like you're one of the top Chinese internet entrepreneurs who likes to think and express grand ideas on a civilized level, much like Sam Altman. Have you ever thought about why?

Xiaochuan Wang: This is a big topic, we could talk about it for two to three hours. But I think it's mainly because of my strong belief in the equality of all people. So, I don't just want to satisfy users from a product and market perspective, but I always want to question my relationship with this world.

Interviewer: Thomas Luo, Haiwei Guo