Wang Xiaochuan officially announced Baichuan Intelligence: What is the endgame of AGI?
Author| Dong Zibo
"These 131 days, to me, feel like several years have passed." Wang Xiaochuan said when recalling the days since the birth of ChatGPT.
A few months have passed since the big model entrepreneurship started. Wang Huiwen landed "Light Years Away" in the Sohu Building, Li Zhifei returned to go out to ask for a two-front operation, and Wang Changhu named the new company "Love". Poetry "Recruiting troops and horses, Jia Yangqing is still waiting for success after leaving Ali..."
And Wang Xiaochuan was not idle either. Downstairs in Wang Huiwen's new office, Xiaochuan held a media communication meeting and officially announced his new brand - "Baichuan Intelligence". At the meeting, Wang Xiaochuan wore an orange-red hoodie and did not use PPT. He just talked in front of the crowd. Someone close to Xiaochuan told Leifeng.com: "He has changed a lot in the past few years and has become more relaxed." A lot, and a lot of humor.”
As early as the previous article on Leifeng.com - "ChatGPT heroes compete: Lu Qi slays the dragon and commands the world; if Xiaochuan doesn't come out, who can compete with him" - it has been mentioned in terms of academic appeal, engineering ability, political and business ability, In the four dimensions of willingness and momentum, Wang Xiaochuan is not lagging behind others, and his comprehensive strength is first-rate.
By the end of April, Baichuan's team will reach 50 people. "By the end of this year, the training of the model benchmarking ChatGPT3.5 will be completed." Wang Xiaochuan said.
Recently, Leifeng.com had a conversation with Wang Xiaochuan, and they had an in-depth exchange of views on the possible forms of generative AI and even future AGI, the final outcome, and how entrepreneurs can find the right position to enter the game.
During the exchange, Wang Xiaochuan mentioned that in the past two months, he has been repeatedly thinking and iterating about the future AI large model strategy and decision-making, and "basically the path has been figured out."
In the later stages of Sogou, Wang Xiaochuan increasingly felt the difficulty of recruiting personnel; and after ChatGPT, as people's enthusiasm for AI increased, Wang Xiaochuan said that it was easier to gather talents than a few years ago. By the end of this month, the team The scale can be expanded to 50 people, and some people even "bring capital into the group."
It is reported that Baichuan’s team has recruited a large number of his old Sogou employees. On the one hand, they understand search and NLP better and are on par with the big models. On the other hand, they agree more with Xiaochuan’s values and have a better sense of purpose. On the other hand, they also know better. How to cooperate with him and know what the other person wants.
As for the team, Wang Xiaochuan's goals don't stop there. He told Leifeng.com that after clarifying the structure of the existing team, he will immediately go to the United States to "recruit people", which is quite like "the Duke of Zhou spits out his support and the world returns to its heart." meaning.
We mentioned in the original article: Wang Xiaochuan was born in Tsinghua University and co-founded the Tiangong Artificial Intelligence Research Institute with Tsinghua University, which has high academic appeal. Behind Ogawa, who received a lot of help, there was also the support of Tsinghua University, one of the top universities in the country.
Commenting on Wang Xiaochuan’s large-scale entrepreneurship, Zheng Weimin, an academician of the Chinese Academy of Engineering, professor of the Department of Computer Science at Tsinghua University, and Wang Xiaochuan’s master’s tutor at Tsinghua University, said: “Wang Xiaochuan dares to innovate and think, and has rich experience in system engineering... He studies parallel computing and related The architecture is an important work of the High Performance Computing Institute of Tsinghua University, where I work. I also have rich experience and will definitely cooperate and support Wang Xiaochuan.”
Zhang Bo, academician of the Chinese Academy of Sciences, dean of the Institute of Artificial Intelligence of Tsinghua University, and Wang Xiaochuan’s doctoral supervisor, also said: “The newly created Baichuan Company has a very strong team, and I believe he (Wang Xiaochuan) can complete this mission and continue to work on it.” We will give you our full support in future development.”
In addition to academicians Zheng Weimin and Zhang Bo, professors such as Yin Xia, Ma Shaoping, and Liu Yiqun of the Department of Computer Science at Tsinghua University also had kind words for Baichuan Intelligence and expressed their full support.
In terms of funding, Wang Xiaochuan also said that he can rest assured about the current financial situation. The US$50 million in start-up capital in hand can already support his current team and computing power costs. As for the large-scale model from zero to one, Wang Xiaochuan estimates that the scale of the cost is about 300-200 million US dollars.
From a product perspective, Wang Xiaochuan is determined to make progress and bluntly said: We want to make the best AI model in China.
How to do it best? Many people blindly believe in model parameters - the larger the parameters, the stronger the model.
But Wang Xiaochuan disagrees with this. He said that blindly pursuing big parameters would be a bit exaggerated. At present, Baichuan has started model training with a parameter level of about 50 billion, and will benchmark the GPT3.5 model by the end of the year.
Wang Xiaochuan’s core understanding of the outcome of AGI
Language is the key to opening the door to AGI
Twenty years later, Wang Xiaochuan left Sogou and said boldly: "In the next twenty years, if I can contribute to the development of life sciences and medicine." Two years later, Wang Xiaochuan entrusted the life sciences to Yang Hongtao. , fully devoted to large-scale entrepreneurship. Regarding this, it was inevitable that he would have some confusion in his heart.
To find out the secrets of life sciences, 20 years is too long and too idealistic. Wang Xiaochuan knows that idealism and reality need to be balanced. As early as a few years ago, AlphaFold developed by Google realized the function of predicting the 3D structure of proteins based on their genetic sequences. Although the function is far from perfect, this allowed Wang Xiaochuan to see the "curve to save the country" from AI to life sciences.
To build AI, Wang Xiaochuan did not choose to start with life sciences. Instead, he chose a language he was more familiar with.
Why start with language? Wang Xiaochuan’s thinking starts from human epistemology: Only through language can we understand the world.
Many people have long asked him, the wave of AGI is so strong, and the opportunities are not limited to large models, why not do it in areas with more mature technical paths such as Vincentian graphics, CV, and intelligent driving?
But Wang Xiaochuan’s position is quite firm. Language is the carrier of knowledge, thinking, communication, and even culture. If the target is AGI, the “crown jewel”, then we must start with language—language is far closer to AGI than images and The vision must be closer.
At the same time, language has always been Wang Xiaochuan’s advantage. Regarding Leifeng.com, Wang Xiaochuan said bluntly: "The input method guesses what you want to say, and the search engine guesses what you want - and ChatGPT is a one-stop solution to these two needs."
Big model X big application, both are indispensable
Drawing on the "old paths" of WeChat and Taobao in the past, if you want to break through, in addition to having technology, the key is to create China's own killer application.
Wang Xiaochuan understands this very well: Sogou's achievements are not reflected in the search engine technology itself; it was only after the two popular applications of search engine and input method that Sogou's value was truly seen.
This is also the reason why Wang Xiaochuan entered the business and quickly accumulated a number of old Sogou departments. Presumably, Wang Xiaochuan's "three-level rocket strategy" guiding ideology and experience in building killer applications will once again be of value in this entrepreneurial venture.
"We will definitely always pursue the ultimate in intelligence. But the difference between us and ChatGPT is that we also pay attention to whether the scene where the large AI model is implemented is real; whether the productized AI is really useful." Wang Xiaochuan expressed this.
From input method to search, and in the future to Chat, Wang Xiaochuan believes that the king of tomorrow will be the "Chat Pro" form of "Chat + search".
"I think that today's OpenAI is a bit 'arrogant'. It has strong AI capabilities but does not attack the search field; New Bing takes search as its core and adds the capabilities of ChatGPT, but it is still not pure enough."
Wang Xiaochuan believes that Chat is just an upgrade of experience and an ability to enhance the core of the product. To really fall into actual scenarios, it should be oriented to professional fields - such as health and law, and these professional fields should be integrated together to create a "big chat".
To complete the integration of many fields and create a "super APP" in the AI era, it must be supported by a large AI model with powerful capabilities.
"Accompanying X Knowledge" - How does AI achieve inclusive information?
Wang Xiaochuan told Leifeng.com that the paradigm of the last era can be called "connecting X information". Taking Google as an example, whether it is portals, searches, or recommendations, all knowledge is connected through the Internet.
In Wang Xiaochuan's view, there are three types of connections in the previous paradigm - point-to-point "Portal connections", "search connections" from keywords to massive related information, and from user habits to "Recommended links" for recommended content.
Douyin has taken "recommended connection" to the extreme, and has become one of the kings in the mobile Internet era.
In today's era of rapid AI development, Wang Xiaochuan believes that the new paradigm should be called "accompanying X knowledge" to make it easier to express and acquire knowledge.
Under this paradigm, the "accompanying" role of Chat ability can be maximized - for example, in current hospitals, doctors cannot devote all their energy to every patient; with the help of Chat ability, AI can complete personal tasks The role of a doctor is to provide one-on-one care to patients.
In similar scenarios, teachers, lawyers, and doctors can all use companionship to complete the transfer of knowledge.
In the old days, the swallows in front of Wang Xietang flew into the homes of ordinary people. With the empowerment of the "Companion X Knowledge" system, society will become flatter, and private legal, medical and other services will become more inclusive.
Two possible endings for generative AI
AI's To B business is already a red ocean today, and can even be called a "dead sea."
It is true that To B business can earn more stable revenue, and as mentioned above, business decisions must not be made just for the immediate moment. Wang Xiaochuan believes that when looking at the general direction of AI, we must see the “endgame” of generative AI.
Throughout the history of computer development, from mainframes to WorkStation, to PCs and mobile phones, the ultimate path of technological simplification will always fall on individuals.
"After the service industry is replaced by machines, people can be liberated to innovate, do things that go into the universe, and hand over simple services to robots." Wang Xiaochuan said.
He also said in his circle of friends: “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, thinking and exploring “Who am I?” "We can also inject our own wisdom into it and be the pioneers of a new era, so that future generations will have a better future and prosper and continue human civilization."
Regarding the vision of Baichuan Intelligence, Wang Xiaochuan told Leifeng.com that there are currently several:
First of all, we must build the best large model in China. Currently, Baichuan Intelligent’s large model is being trained step by step, striving to release it before the end of the year;
In response to the hallucination problem of "nonsense" in ChatGPT and similar products, Wang Xiaochuan intends to use Sogou's past accumulation in the search field to increase the accuracy, detail and timeliness of answers;
From the perspective of product implementation, we will enhance the knowledge accumulation of large models in vertical fields such as education and medical care, so that large models can shine in the professional field as early as possible;
The end result of all this is to allow the public to easily and universally obtain knowledge and professional services, and use the evolution of AI technology to promote the improvement and transformation of social productivity.
京公网安备 11010802033920号