The fifth anniversary of the dormant model of Tsinghua University: 1 pioneer strung together half of the entrepreneurial history, and 3 alumni competed for "first brother"

Source of this article: Times Finance Author: Xu Xiaoqian

Source: pixabay

After only a year, the competition for large models has become divided.

Recently, some media reported that Baichuan Intelligence, founded by Wang Xiaochuan, has carried out a new round of financing of hundreds of millions of dollars, which will become one of the largest AI financing in China in 2024. Investors from the participating confirmed that Baichuan Intelligence has been valued at more than 1.8 billion US dollars (about 13 billion yuan) since the beginning of this year. In this regard, Baichuan Intelligent responded to Times Finance and Economics that there is a certain error in the report, and the specific information needs to be further announced by the official.

If the above financing news lands, it means that another Tsinghua large-scale model company will enter the 10 billion valuation club.

According to incomplete statistics from Times Finance, from 2019 to 2023, at least 30 Tsinghua alumni will start their own businesses, invest in the large-scale model industry, establish more than 15 companies, and raise more than 15 billion yuan in total. These enterprises, founded by Tsinghua alumni, are known as "Tsinghua Department" in the industry.

With the rise in the valuation of Tsinghua large-scale model enterprises, a "Three Kingdoms Killing" in the industry was silently staged.

The large model cannot be separated from the Tsinghua system

Three years ago, two first-generation Internet entrepreneurship bigwigs announced their retirement from the circle.

In December 2020, Meituan issued an internal letter announcing that Wang Huiwen, co-founder and senior vice president of Meituan, had successfully completed the handover and officially exited as planned. Half a year later, Sogou was acquired by Tencent, and Wang Xiaochuan, who was once a leader in the search field, gradually moved away from the center of the Internet stage.

At this time, on the other side of the ocean, a dazzling new star is rising in the AI academic community. OpenAI's GPT-3 has made a lot of waves in the academic world, with a parameter scale of up to 175 billion, which can not only do language generation, but also greatly improve the ability to understand language.

To the outside world, GPT is far from bursting with potential energy. But a group of top scholars at Tsinghua University quickly made a big decision: to catch up with GPT3 and invest in large model training.

Three years later, GPT-4 took the wind of large models around the globe through a conversational product. Wang Huiwen took the lead in initiating the recruitment order for AI large model talents, and then Internet star entrepreneurs such as Wang Xiaochuan and Li Kaifu appeared one after another. Wang Huiwen and Wang Xiaochuan, former Tsinghua alumni, once again met in the arena because of large models. Only this time, after they leave their comfort zone, they will face more opponents from the "academic circle".

In the core storm of the 100-model war, the Tsinghua Department has undoubtedly become a pivotal force, they are like a rushing river gradually diverging into streams with their own characteristics, some are veteran mages who have experienced a hundred battles in the Internet era, and some have been dormant in the field of large-scale model research for many years.

"If you make large models or chips, then you'd better be from Tsinghua. A chip industry practitioner told Times Finance. Many scientific research talents, start-ups and technical reserves have jointly built a network of contacts called "Tsinghua Department", which seems to have become a kind of industry consensus.

In the past year, technicians of large-scale unicorn companies have stayed up late every day to catch up with OpenAI.

Light Years Away and Zhipu AI both selected the basis in the Sohu Building outside the East Gate of Tsinghua City, across the street is the Sail Building, where Wang Xiaochuan founded Baichuan Intelligence, and there are also two Tsinghua start-up companies on Chengfu Road on the south side of the Sohu Building, namely Lingxin Intelligence and Shenyan Technology.

A new wealth-making movement is being carried out vigorously, and the "Tsinghua Department" has become an inescapable historical seal.

1 pioneer strung together half of the history of entrepreneurship

Years later, when the outside world recalls this boom in large-scale model entrepreneurship, KLCII will undoubtedly become an iconic symbol. It is only a 5-minute walk from the south gate of Tsinghua University, and now, within three kilometers of the main campus of Tsinghua University, it is the base of large-scale model entrepreneurship.

In 2020, artificial intelligence has generally fallen into a dilemma, the market value of the AI four tigers has fallen to the bottom, and huge R&D investment and inefficient landing output have made a group of investors retreat. According to a research report by Southwest Securities, the number of financing of AI companies has declined sharply since 2019, and ninety percent of AI companies are in a state of loss.

Contrary to the strong pessimism of the outside world, KLCII is quietly brewing a change, bringing together a group of large model pioneers such as Sun Maosong, Tang Jie, Liu Zhiyuan, Yang Zhilin, etc., who agree that large model training can bring about a disruptive change.

Hundreds of people put their enthusiasm and energy into the alchemy furnace of the large model, opening a long period of land reclamation.

At the beginning of 2021, a large model named Wudao 1.0 was launched, which presented the results of the multimodal model BriVL 1.0 trained on 30 million graphic data pairs, the GLM-10B bilingual model with 10 billion parameters, and the multimodal model CogView 1.0.

On this basis, Tang Jie, as the leading teacher, made a bolder attempt, he called on external resources to train several trillion-level large models. Three months later, the Wudao 2.0 series of large models were unveiled. As a domestic large-scale model system benchmarking GPT-3, Wudao 2.0 created the world's highest record at that time with 1.75 trillion parameters, nearly 10 times more than GPT-3, which also became a landmark milestone in the initial stage of China's large-scale model.

Tang Jie's other identity is one of the founders of Zhipu AI, the company was founded in June 2019, and the Knowledge Engineering Laboratory of the Department of Computer Science of Tsinghua University is its cradle, and almost all the core members of the team are of Tsinghua blood.

After the Wudao series of large models became famous, the atmosphere of Tsinghua entrepreneurship was pushed to a small climax. A group of Tsinghua alumni gradually came out of the laboratory to explore the path of commercialization of large models.

In November 2021, Huang Minlie founded Lingxin Intelligence, in March 2022, Sun Maosong's doctoral student Gui Fanchao founded Shenyan Technology, in August 2022, Liu Zhiyuan founded Facewall Intelligence, focusing on the legal field, and in 2023, Zhu Jun, who also studied under Sun Maosong, established Shengshu Technology.

At this time, the large-scale model entrepreneurial team is once again facing a dilemma: they can't find the connection point between the large-scale model and the market, there are not many companies willing to eat crabs, and they don't even know how to schedule the ability of the large-scale model.

It wasn't until a nuclear explosion point emerged that these early pioneers grabbed the ticket to the next technological change.

Academic VS Entrepreneurial

In November 2022, ChatGPT launched by OpenAI was born, and the function of automatically generating text shocked the global technology circle, with the number of users exceeding 100 million in just 2 months, and the large model began to enter the public eye.

Like a spring breeze coming overnight, the Tsinghua model, which has been working silently for several years, has gained the favor of capital and is the first to seize the outlet. Times Finance found that in 2023, the financing of Tsinghua large models will gradually usher in the fast lane, and more than half of the Tsinghua large models will have their first financing of hundreds of millions of yuan in 2023.

"In the past few years, AI investment has been basically zero, and when everyone felt that their eyes were full of haze, suddenly ChatGPT came out and brought light to people. Zhou Ming, the founder of Lanzhou Technology, made such an exclamation.

This is a bottomless investment of resources, training a high-quality large model, requiring tens of thousands of chip clusters and cloud computing scheduling, and also to catch up with the industry indicators stacked by OpenAI, once opened, it means that a huge money shredder is started.

According to a research report by Guosheng Securities, GPT-3 can cost up to $1.4 million for just one training, and for some larger LLMs, the cost of a single training can range from $2 million to $12 million. Zhang Peng, CEO of Zhipu AI, also said in public that the financing of 2.5 billion yuan is simply not enough, no matter how much money is raised and how much money is earned, it is a entanglement on the road to AGI.

In October 2023, Zhipu AI announced the completion of more than 2.5 billion yuan of financing (with a valuation of more than 10 billion yuan), and at the same time, Baichuan Intelligent ushered in a new round of financing with an amount of more than 300 million US dollars, and the dark side of the moon completed two rounds of financing totaling nearly 2 billion yuan last year.

Compared with the difficult and low-key dormant of the academic school, Internet star entrepreneurs are usually more decisive and resolute, and they have a sharper judgment of applied products and markets. At the beginning of 2023, after witnessing the advent of ChatGPT, Wang Xiaochuan, a serial Internet entrepreneur, smelled the breath of technological change, and he excitedly posted a circle of friends: "The victory of OpenAI is the victory of technological idealism. ”

The path of setting up a company first and then recruiting a technical team is the best way for entrepreneurs to quickly seize the attention of the market. After thinking about it for two weeks, Wang Xiaochuan put down the smart health project at hand and turned to the big model.

In less than half a year, the team of Baichuan Intelligence was set up, and Wang Xiaochuan got $50 million in start-up funds and officially announced that he had entered the AI model entrepreneurship track. In addition, he also brought in his old subordinates, Ru Liyun, former COO of Sogou and alumnus of Tsinghua University in 1998, and half of the members of the new startup company are from the original Sogou team.

After half a year of rampage, the list of unicorn large models has gradually surfaced - Zhipu AI, Dark Side of the Moon, Baichuan Intelligence, MiniMax and Zero One Everything, whose publicly disclosed financing in 2023 is about 10 billion yuan, accounting for 40% of the country's AI investment, and the Tsinghua large model has once again won half of the country.

However, judging from the overall financing scale, the popularity of the large-scale model concept has not triggered an investment boom. According to a later report, the total financing amount in the AI field in 2023 will decrease by 4.5% year-on-year to 25.7 billion yuan, less than half of that in 2021.

"This market is currently in a situation where it is difficult to raise funds, invest and exit, so for large models, most institutions are still looking at it, and if they really make a move, it is not like FOMO (fear of miss out) that we often said before, it is out of the mood of 'fear of missing out'. Wu Shichun, founding partner of Meihua Venture Capital, said in an interview with Times Finance.

Who is the first brother of the Tsinghua department?

Without the explosion of the intelligent assistant Kimi, the dark side of the moon may not be able to stand in the spotlight quickly. In the past year, the founder Yang Zhilin has almost hidden in the huge pursuit of his fellow brothers, and has not gained much exposure.

At the beginning of 2024, the Wensheng video application Sora and the concept of long text became popular, and the investment in large models that was briefly extinguished was able to recover again. The dark side of the moon was the first to be raised by capital, and it was rumored that a new round of financing of more than $1 billion was completed, which was the largest financing in the large model track after the emergence of ChatGPT, with a valuation of $2.5 billion (about 18 billion yuan), and investors included Alibaba, Sequoia China, Xiaohongshu, Meituan, etc.

Yang Zhilin, who was born in 1992, has become the focus of the entire large model field, and before entering the large model, Yang Zhilin has already conceived two problems, one is the first capital of entrepreneurial preparation, and the other is to build the direction of entrepreneurship - to find application opportunities through interaction with humans.

In October 2023, the dark side of the moon launched the intelligent assistant Kimi, which became the first popular project of a large model startup, and on March 19, the dark side of the moon announced that it would increase its long text capacity from 200,000 words to 2 million words, and its weekly visits reached 3.94 million, slightly higher than Wenxin Yiyan's 3.38 million times, and once jumped to the AI application with the highest weekly visits on the domestic web side.

In the eyes of some industry insiders, the weapon of overtaking in the dark side of the moon lies in the label that binds long text, and the positioning is relatively vertical. "C-end oriented products are easier to win market attention, but at the same time, they will be easy to be surpassed because of the lack of technical barriers. AI entrepreneur Zhang Wei told Times Finance that he believes that the development direction of large models is still relatively vague, and everyone is crossing the river by feeling the stones, facing a completely unknown world.

As an evangelist in the large model industry, Zhipu AI, founded by Tang Jie, has frequently invested abroad and appeared in the financing teams of 13 large model companies, including computing power, software and hardware, applications, medicine, insurance and other directions.

However, when technology shines like fireworks, it's hard to tell if it's a flickering illusion or a real world that can be touched.

"I thought I got an all-purpose hammer, and I thought everything I saw was a nail, but in fact, I found that it was difficult to cut into the business scenario. There are not enough resources for the B-side to do the enterprise, and the choice to do the C-end promotion and cost are all problems. Zhang Wei said to Times Finance.

Even Baidu, which has a more mature ecosystem, has not yet found a bright outlet in the field of large models. According to publicly disclosed data, in the last quarter of 2023, the large model will only bring incremental revenue of 660 million yuan to Baidu's cloud business.

"The team that masters the core technology of the model will stand out from the competition, launch the corresponding application products, and carry out multiple iterations and optimizations to find the right route. If you only develop application products and do not have the core technology you have mastered, there is basically only one chance for trial and error. Zhou Zhifeng, a partner of Qiming Venture Capital, recently said in public that in his view, companies with large model technology will eventually move towards companies with integrated model applications.

If we focus on the core area of the Tsinghua Arena, the "Three Kingdoms Killing" has also begun to get into the right topic.

Zhipu AI, the dark side of the moon and Baichuan Intelligence quickly entered the $10 billion valuation club, but the weapons used by the three are not the same: Zhipu AI, which has an academic background, is looking for a fusion scenario between enterprises and large models; Yang Zhilin, the founder of the dark side of the moon, is the ultimate technological idealism, he believes that the "aesthetics of violence" of technology can change the world, and Wang Xiaochuan, who hesitates between the two routes, wants to find another perspective in idealism and realism.

Like most practitioners, chaos and hesitation have almost run through Zhang Wei's entire year of entrepreneurship, and he doesn't know where the industry will go next, and he is still waiting for the emergence of a coordinate system.