Baichuan Intelligent Launches 13 Billion General-purpose Large Language Model with Parameters to catch up with ChatGPT this year?

Reporting by reporters Li Kunkun and Li Zhenghao from Beijing

Recently, Baichuan Intelligence, founded by Wang Xiaochuan, former CEO of Sogou, officially released the general large-language model Baichuan-13B-Base with parameters of 13 billion, the dialogue model Baichuan-13B-Chat and its INT4/INT8 two quantitative versions.

Wang Xiaochuan said: "We look forward to the domestic large-scale model industry and vertical fields to develop more excellent products and industry applications on this basis, so that technology can quickly iterate and innovate in real and rich application scenarios, and we are willing to work with many enterprises and developers to contribute to the ecological prosperity of the domestic open source community." ”

Strengths and weaknesses

"There are many companies that make big models now, but it costs money to make big models, especially general artificial intelligence big models." An artificial intelligence practitioner told the "China Business News" reporter that the current large model track is very hot, but in the end it is estimated that there are only 3~4, and other estimates are related to the ecology and application landing around these few.

The above people said that for large model startups, data is a big problem, as well as computing power problems, and training costs are also very high. OpenAI has a peculiarity, it was supported by several giants in the early stage, and now few giants support startups like this, and many have to do it themselves.

On April 10, 2023, Wang Xiaochuan officially announced the establishment of Baichuan Intelligence, a Chinese company that develops and provides AGI (General Artificial Intelligence) services, aiming to build a Chinese version of OpenAI basic large models and disruptive upper-layer applications. Baichuan Intelligent claims that it will build "China's best large model base" with the help of language AI breakthroughs, and enhance it in search, multi-modality, education, medical and other aspects to help the public easily and universally obtain world knowledge and professional services.

Speaking about the origin of the company's name, Wang Xiaochuan explained: "Baichuan originally means that many rivers converge to the ocean, symbolizing the convergence of many data and industry knowledge into a powerful intelligent system. Baichuan also symbolizes the wisdom of a hundred schools, and more people go to the mountains and seas with them. ”

Wang Xiaochuan's logic is that the "knowledge-intensive" field is his core breakthrough point, and education and medical care are the knowledge-intensive direction in his eyes. He has always pursued "the meaning of life".

It is understood that the general large language model Baichuan-13B-Base is the second general large language model released by Baichuan Intelligence, and on June 15, not long ago, Baichuan Intelligent has launched the first Chinese and English language model Baichuan-7B with 7 billion parameters, and won the first place in the same magnitude test list of multiple world authoritative benchmark lists.

Talking about the advantages and disadvantages of Baichuan Intelligence, Li Zhe, chief analyst of Ai Analytics, told this reporter: "Baichuan Intelligence is currently doing basic large models, and the future should be to do general large models and B-side/C-end applications." We are currently evaluating large model manufacturers according to the four dimensions of computing power resources, data sets, AI engineering and ecology, the advantage of Baichuan Intelligent is the dataset and AI engineering capabilities, the dataset is the accumulation of corpus data in Chinese scenarios, AI engineering is mainly reflected in the advantages of its algorithm framework layer and model layer, computing power resources and ecology have disadvantages compared with other manufacturers, Baichuan 'tunka' (referring to GPU boards) is too late, and the time to launch the basic model is also lagging behind other large model manufacturers. ”

Li Zhe said that the C-end is limited by regulatory and other factors, and the short-term development will not be particularly fast, and the focus is definitely on the B-side. The B-side generally starts from the future business model of large model manufacturers to think about the future competitive pattern of the large model market. The future business model of large model manufacturers includes: one is API call, the other is large model license/all-in-one machine, and the third is industry application (application products and solutions). API calls will definitely be market opportunities for cloud vendors in the future, and the opportunities of Baichuan Intelligent lie in large model licenses/all-in-one machines and industry applications, which is more optimistic about the market opportunities of Baichuan in industry applications.

Opportunities and challenges

Previously, Wang Xiaochuan told the media: "Catching up with the ChatGPT level, I think it may be achieved within this year, but for GPT-4 or GPT-5, I think it may take about three years, which should not be less than two years." ”

Wang Xiaochuan said frankly that there is a big gap between the products of large domestic manufacturers and ChatGPT, and we must continue to chase OpenAI. "Now the first thing Baichuan Intelligent has to do is China's best big model, and then go to the United States to recruit talents and then chase it (OpenAI). This is more realistic, people make 'immortal pills', let's make a 'longevity pill' first. ”

Baichuan Intelligent has many opponents, and the domestic large model "arms race" is constantly updated in terms of "days", and Alibaba, Huawei, JD.com, and SenseTime are all vying for seats on this general artificial intelligence (AGI) train.

Shenyang, senior consultant of the cabinet think tank, told this reporter that compared with other AI companies, the advantages of Baichuan Intelligent are: First, the technical ideal, the company has advantages, Wang Xiaochuan is currently in the best state, but also the domestic in the AGI field of the deepest cognition, the most capital and talents, the most international vision of the person, his cognition surpasses peers, at least the current most leading. In particular, his technical cleanliness can stay away from business and pursue breakthroughs in AGI. Second, there are many infiltrates of Chinese traditional culture, and the current leading companies are mostly overseas companies, and Wang Xiaochuan has the heritage of traditional Chinese culture. The third is the deep understanding of AGI, which is the biggest highlight of Wang Xiaochuan's Baichuan Intelligence, and may also be the height that other AI companies cannot reach, Baichuan Intelligence is the most likely company to break through AGI that can be seen in China.

Shenyang believes that the reason for cautious optimism is: on the one hand, Wang Xiaochuan has been relatively smooth, has not experienced too much tribulation, may not have encountered major obstacles and bottlenecks in self-cultivation and improvement, so he may not have a huge breakthrough ability, which is indeed a problem of personal cultivation; On the other hand, the current medical and education fields may be the best breakthrough areas of AGI, but the solution path is too "materialized", more in the fields of capital, talents, computing power, resources, etc., and does not pay too much attention to the high-dimensional energy fields such as "consciousness" and "cognition", which may be a defect.

Wang Xiaochuan said: "Baichuan-13B is a gift from Baichuan Intelligent for a scientific and technological power. ”

(Editor: Wu Qing, Proofreader: Yan Jingning)