“Today’s distance ChatGPT Posted, it’s day 131. “

Wearing an orange hoodie, Wang Xiaochuan walked into a conference room of nearly 30 square meters, where he announced that the large-scale model product he was about to invest in his business was called “Baichuan Smart”. The word “Chuan” is a pun, which not only symbolizes the process of rivers converging into the sea, like data converging to generate support; it also implies his own name, “I put myself in it. It also means that more talents are needed to join in. .”

This internal exchange meeting can be described as low-key, there is no PPT on site, Wang Xiaochuan finished talking about his entrepreneurial background and vision to an iPad, and the sound of electric drills for decoration came from upstairs from time to time during the process. This is on the second floor of the Sohu Internet Building. The dramatic thing is that Light Years Beyond, founded by Wang Huiwen, who also invested in large-scale entrepreneurship not long ago, also works in this building.

The start-up capital of this venture is also low-key-“Baichuan Smart” was started with 50 million US dollars. According to Wang Xiaochuan himself, part of the source of funds is his personal contribution, and the rest is the personal support of several friends. Compared with the 200 million US dollars in financing (not announced close), in the field of large models, 50 million US dollars is not an eye-catching amount. But Wang Xiaochuan said that the money will be enough in the early stage to run the product from 0 to 1.

butAccording to Geek Park, Wang Xiaochuan could have obtained a larger amount of financing, but he chose a relatively small initial start-up capital. In March, when big news continued in the field of large models, changes in the external environment made him gradually transition from impulsiveness to calmness, and he chose to enter the arena with a very low-key posture in order to “not put too much pressure on the early team.”

In fact, as early as a month ago after Wang Huiwen entered the game with a high profile, Wang Xiaochuan began to communicate closely with the industry, academia and relevant government departments.Recently, he posted a lot of remarks on social platforms, saying that if the machine masters the language, thenstrong artificial intelligenceThe time is coming. At this exchange meeting, Wang Xiaochuan emphasized that the background of starting Sogou is his biggest advantage in leading the team to make large-scale models. “Search is language. In the direction of AGI, language is a great advantage.” According to the observation of GeekPark, Wang Xiaochuan is also the most promising one among the large-scale entrepreneurs in the past month.

In the picture currently announced by “Baichuan Intelligent”, the goal is “to build the best large-scale model base in China”-“best” is reflected in parameters, quality testing and experience. Different from other big models, “Baichuan Smart” will take the lead in enhancing knowledge in the two vertical fields of education and medical care. The goal is to help the public obtain inclusive knowledge and help enterprises provide general services. The base product target will be announced at the end of the year, and the vertical product may be announced in advance.

Talent is the key word in the competition in the large model field.Wang Xiaochuan revealed that by the end of the month, the team of “Baichuan Smart” will be close to 50 people, less than half of which are from Sogou. And half a month later, he himself will fly to the United States to continue recruiting technical talents.

Interestingly, this exchange was inspired by ChatGPT, when Wang Xiaochuan asked about ChatGPT: How to recruit more and better talents in the field of large models? The answer given by ChatGPT is that you must first explain your logic to the outside world. Hence the meeting.

Regarding the financing pace of this venture, Wang Xiaochuan’s performance did not match the fanatical expectations of the outside world. When asked by Geek Park whether this is still the style of a professional manager, not the status of an entrepreneur—he replied with a smile, “I am used to announcing when everything is ready. I admit that this has the inertial thinking of a professional manager, but I have come out, so I choose to announce to everyone first.”

01 From frenzy to calm

Behind Wang Xiaochuan’s low-key official announcement is a microcosm of the large-scale model business from fanaticism to calmness in the past few months.

Speaking of Wang Xiaochuan, the first label that pops up is often “science and engineering genius”. He has the characteristics that the world has in the stereotype of “science and engineering man”: he is cautious in doing things, and he doesn’t want to do anything.But when it comes to large models, Wang Xiaochuan also fell into a frenzy at first.

Just when Wang Huiwen announced in the circle of friends to make the Chinese version OpenAI The day before , the entrepreneur community “Frontier Society” held an online exchange meeting on general large models.“OpenAI is showing the beginning of a revolution, which will be an industry change of the same magnitude after the Internet and the mobile Internet.”At the meeting, Zhang Peng, the founder of Geek Park, put forward this point of view and initiated a vote among the participants.

A total of 13 votes were cast for approval, and the first voter among them was Wang Xiaochuan. Since participating in many activities, that night was the one he expressed the most and was the most excited in more than a year.

In this exchange activity, Wang Xiaochuan was ChatGPT One of the staunchest supporters.When several people who participated in the exchange AI When entrepreneurs in the field tried to define the progress of ChatGPT only in terms of function, Wang Xiaochuan even made it clear that everyone “considered this matter small.”

“I see the dawn of AGI (General Artificial Intelligence)!” Wang Xiaochuan said that this is what he started to experience at the end of last year ChatGPT The strongest feeling in my heart. And in this discussion activity, for the first time, he implicitly revealed his attitude that he might also participate in the transformation of the large model.

On February 12, 2023, at an “AGI Believers Gathering” held by Geek Park Frontier Club, Wang Xiaochuan clearly expressed his determination to “be determined to join the game” for the first time. In other words, Wang Huiwen’s sudden high-profile entry actually disrupted Wang Xiaochuan’s original steady rhythm, helped him get out of the “manager’s thinking” to some extent, and began to activate the “adrenaline”.

From left:Li ZhifeiZhou Yuan, Wang Xiaochuan, Zhai Guanglong, Zhang Peng

Geek Park learned from people close to Wang Xiaochuan that within a week or so after Wang Huiwen announced his entry in a high-profile manner, Wang Xiaochuan also prepared to announce his entry to the outside world. However, during that period of intensive public opinion upsurge, the FOMO mentality that the outside world would be left behind if they did not board the ship, made him finally start to brake and reflect on whether his hasty announcement was really necessary.

Wang Xiaochuan’s most familiar “old place”

During the period of Wang Xiaochuan’s silence, other teams that released the news first started intensive preparations. Wang Huiwen established Beijing Light Year Beyond Technology Co., Ltd., which attracted subscriptions of hundreds of millions of dollars from top VCs.

Another founder of Mobvoy who announced his commitment to the field of large-scale modelsLi Zhifeiwho expressed his opinion to Geek Park, he believes that the general AI Large models are like nuclear weapons, and there is indeed a time window for entry. “Once the barriers of talent, time, data, and capital are established, a small team will be useless.” Wang Xiaochuan and Li Zhifei are also very familiar friends. He also agrees with this point of view, but he also put forward a point of view—— “This matter can’t be discussed above, and we still need to think about it.”

In the past period of time, there are indeed some things worth thinking about.For example, take Wang Huiwen, who was the first to announce his entry into the game, as an example. One month later, even if he has a large amount of money and recruits enough talents, especially a chief scientist or CTOis still a challenging task.

And from the acquisition AI Architecture company’s first-class technology Oneflow, to the intention to acquire two Tsinghua University NLP The team’s start-up companies – Shenyan Technology and Facing Smart.also reflectsWang Huiwen hopes to use the shortest path to gather talents as soon as possible, which just reflects the complexity of large-scale entrepreneurship. Capital is of great significance, but it is not the only decisive factor.

And Wang Xiaochuan AI The accumulation in the field and his personal brand and experience in the field of technical products did allow him to choose another way to enter the game after thinking about it.

02 Why did financing ideas change?

The first different thought that deserves attention is why Wang Xiaochuan set the start-up capital at 50 million US dollars.

exist ChatGPT When it came out, Wang Xiaochuan once posted a circle of friends saying:OpenAI The victory of China is the victory of technological idealism.

From founding to launching ChatGPT，OpenAI It’s been eight years, and before that, they invested a lot of money and resources. But it is precisely this kind of belief in technology that allows them to achieve world-shocking achievements. On the eve of AGI, they can earn themselves a leading edge that is difficult to match today.

Initially, domestic entrepreneurs all wanted to be in China OpenAIthat is, to make the underlying general large model, the logic behind this is very simple:

1. OpenAI Can not enter China, the country must need its own large model.
2. Big Internet companies will make large models, but they can’t eat all the markets. Companies with overlapping businesses will not hand over their lifeblood to big manufacturers, so entrepreneurial teams have room to survive.
3. OpenAI You have embarked on a successful path. If Columbus had already discovered the New World, you are in this direction. It is possible to do the same thing domestically in a shorter time and with fewer resources.

A few months have passed, and today there is no problem with the first two.but copyOpenAIpath, but it looks more and more challenging.

• GPT-3 released 3 years ago is just a basic base;
• half a year ago AI robot ChatGPT(GPT-3.5), is already a complete product, and has the higher-level technical ability of dialogue in general-purpose scenarios;
• GPT-4 a month ago has stronger reasoning ability and multi-modal transformation ability;
• a few weeks ago,OpenAI declare support ChatGPT Plug-ins, the latter is not just a product, but an ecology, and even an operating system.

Based on large model capabilities,OpenAI Product capabilities, commercialization capabilities, and ecological capabilities have been built, which has accumulated a large number of users for them, and the data flywheel effect has further accelerated OpenAI’s iteration of the basic large model.

If the difficulty factor of making a large model is 1, if you want to quickly make a ChatGPT The difficulty of the horizontal, “fully open dialogue product” is 100.

Even if Wang Xiaochuan is determined to be China’s OpenAI, and even eventually surpass him. So the core question he has to face is:

1. what year are you going to do OpenAI? 2015’s ? 2020’s? 2023? Or in 2025? The capability system of OpenAI is different every year. How can new entrepreneurs establish their own characteristics, differentiation and advantages?
2. From catching up to surpassing, how should a start-up company take the first step in its long march? Where is the first milestone set? What is the first business cycle that can be closed?

face to face OpenAI In the past few years, billions of dollars have been burned to build a complete ecology, and the resources required to replicate the same path, whether it is 50 million dollars or hundreds of millions of dollars, are actually a drop in the bucket if you don’t consider the issue of momentum.

More pragmatic thinking is how to start quickly, set reasonable milestones, and reach the milestones with the highest efficiency and unexpected results, so that follow-up resources will continue to flow.

According to friends who participated in Wang Xiaochuan’s current round of financing, Wang Xiaochuan changed his previous financing plan out of this consideration. At the internal exchange meeting, Wang Xiaochuan was asked whether the 50 million US dollars had been calculated in detail. He revealed,“From 0 to 1, 50 million start-up capital is enough. But to continue to do the best, I estimate it should be between 300 million and 2 billion US dollars.”

03 What does Wang Xiaochuan want to do?

At the exchange meeting, Wang Xiaochuan gave several goals of Baichuan Intelligence at this stage:

1. Build the best large model bases in China.
2. It is strengthened in several directions: enhanced search, enhanced multimodality, enhanced education, medical and other knowledge levels.

compared to OpenAI Focusing on a general model, BCI will develop vertical models and general models at the same time.

As for the general-purpose large model, Wang Xiaochuan said that Baichuan Intelligent has already started training a model with 50 billion parameters, and the progress is going well. It is expected to release a large model that is benchmarked against GPT-3.5 before the end of the year. It is confident that it will be the best in China by the end of this year.

In contrast, vertical models are expected to come out faster.Although the generality of vertical models cannot be compared with OpenAI However, by optimizing for specific scenarios, it can achieve an effect similar to that of OpenAI in subdivided scenarios, thereby accumulating users, building an ecology, and running through a small closed loop.

As for the several scenarios selected by Wang Xiaochuan, in addition to searching for his old business, he chose the two vertical tracks of medical care and education because these two industries “have the highest requirements for knowledge”, and this feature will also be a major factor in the industry. The direction in which the model first lands.

Even Wang Xiaochuan once said to friends who participated in this investment that besides the large model, the most energy-consuming thing is the thinking and practice of creating super applications. “The largest and most successful companies must be the ones that think the most clearly company.”

Apparently, China OpenAI Today, this matter is not just a competition of large models themselves, but a “new triathlon” competition that combines large models, super apps, and industry applications in one.

Attachment: Full text of Wang Xiaochuan’s open letter

Today I officially announced the establishment of “Baichuan Intelligence”, a Chinese company that develops and provides general artificial intelligence services.

We are so lucky to live in the early 21st century. The magnificent Internet revolution has not yet come to an end, and the era of general artificial intelligence is roaring again. Many years ago, I asserted that the era of general artificial intelligence will come when machines master language; I also imagined that the future of search is question and answer.ChatGPT The sky was born, the earth was shaken, and all this began to become a reality. It has only been 131 days since ChatGPT was released, and there are new progress and breakthroughs coming every day. 131 days seems like a lifetime away!

For more than half a century, scientists have worked hard to create general intelligent machines.The day finally came, and we found an ingenious way to use AI Successfully transformed language into a mathematical model. Language is the carrier of knowledge, thinking, communication and culture. Wittgenstein said that the boundaries of language are the boundaries of the world. When AI masters the language, it begins to recognize the world and become an intelligent agent with knowledge, thinking, expression and learning. I call such an agent “linguistic AI”.

Individual human beings will die one day, but they can record their lifelong knowledge through language for new life to learn and inherit, so civilization continues. Human collaboration also requires verbal communication between individuals.Master the language of communication and world knowledge AI No longer just a “tool” for humans, but will become our partners and super assistants, and become a part of human civilization.

The age of intelligence will rewrite the paradigm of the Internet age. “Connection” will be replaced by “companion”, “information service” will be replaced by “knowledge service”, and “free and advertising” will be replaced by “paid and high quality”.

The age of intelligence will also rewrite the paradigms since the industrial age.In the past, the specialization of social division of labor brought economies of scale and improved efficiency, but in the future it will be de-specialized division of labor: the work that previously required the cooperation of many companies will be due to AI The empowerment of “end-to-end” completion.

The era of intelligence will once again open the “Renaissance”.everyone will because AI become stronger with help. With good use of AI, one person can build a company, solve problems that were difficult to solve before, and become the person he wants to be more easily. The society will also be flatter, and the intellectual services that only a few people could enjoy before will be more inclusive.

After bidding farewell to Sogou, I have started a new expedition: I am determined to contribute to the development of life science and medicine in the next 20 years, and make a little contribution to public health. The core path is to build a mathematical model of life and health, and I have paid Actions.ChatGPT The emergence of will also become an important boost for the new expedition.

Not only that,ChatGPT It is the continuation and sublimation of my past experience. In 2019, I upgraded Sogou’s mission to “make it easier to express and acquire knowledge”, and defined the vision as “helping everyone to talk and write easily. Access knowledge and services conveniently, improve work efficiency and quality of life”.

The data preparation and product form of Sogou Input Method and Sogou Search are also related to ChatGPT near. They all collect, clean, and compress all language data on the Internet into databases and build super products. With the development of deep learning, these products have been iteratively upgraded many times, and technologies such as Transformer/Bert have been fully used. The input method guesses what you want to say, and the search guesses what you want. And ChatGPT has pushed these capabilities to a whole new level. It is the third super product that converts the whole network data into language services besides the input method and search. Sogou’s unfinished mission can finally be realized after the machine has mastered the language.

Many friends recognize my technical idealism and technology-driven entrepreneurial experience, encourage and support me to do Chinese business OpenAI. There are also many Sogou and non-Sogou technical experts who learned that I have the idea of ​​​​making a large-scale model, and actively apply to participate in the team, and even “bring money to join the team”.

What fascinates me the most is that I can work with a group of people with ideals to create subversive smart products driven by extreme technology.This time, Baichuan Smart will use language AI With breakthroughs, it builds the best large-scale model base in China, and enhances it in search, multi-modality, education, medical care, etc., to help the public easily and inclusively obtain world knowledge and professional services.

It has been less than two months since the establishment of Baichuan Smart, and there are many technicalpartnerand leading figures. By the end of this month, there will be an elite team of nearly 50 people. The training of the large model has also started and is progressing smoothly. According to the current progress, we will strive to release the best large-scale models and disruptive products in China within this year.

We are a Chinese company, but we don’t want to be just “Chinese OpenAI“, but to establish a world-class vision and technical quality, so that China can enjoy world-class products, produce technological breakthroughs with world influence, and then serve overseas. The original meaning of Baichuan is that many rivers converge to the ocean, which symbolizes the convergence of many data and industry knowledge into a powerful intelligent system, which is endless. Baichuan also symbolizes the gathering of wisdom from hundreds of schools of thought, and more people will go to the mountains and seas with me. We call for leaders in China, industry product managers and colleagues in the United States to join us.

Comrades, please contact us via the following email:

80@baichuan-inc.com

90@baichuan-inc.com

95@baichuan-inc.com

00@baichuan-inc.com

oier@baichuan-inc.com

The era of general artificial intelligence has just begun. As the first batch of human beings to enter the new era, we embrace it with anxiety and curiosity, think and explore “Who am I?” We can also inject our own wisdom into it to create a new era The pioneers of the future generations will have a better future, prosperity and continuation of human civilization.

Baichuan Smart CEO Wang Xiaochuan

April 10, 2023