Wang Xiaochuan, CEO of Baichuan Intelligence: Partnering with Alibaba to Foster a More Prosperous Large Model Ecosystem
-
On October 31, at the 2023 Apsara Conference, Wang Xiaochuan, founder and CEO of Baichuan Intelligence, shared his thoughts and practices on large models in the era of AI. Recently, Baichuan launched Baichuan2-192K, the large model with the longest context window globally, attracting media attention. Wang Xiaochuan noted that Baichuan has released seven large models in just six months, with rapid iteration supported by cloud computing.
In April 2023, Wang Xiaochuan, former CEO of Sogou, announced the establishment of Baichuan Intelligence, an AI large model company. Within months, Baichuan became a tech unicorn. Notably, the company released seven foundational large models in just half a year, far exceeding industry standards in R&D speed, with open-source models performing exceptionally. By late August, Baichuan became the first large model company founded this year to pass China's "Interim Measures for the Management of Generative AI Services."
Wang Xiaochuan believes the large model wave will surpass all previous expectations, with a technological explosion underway as humanity enters the era of artificial general intelligence (AGI). In 2018, he stated, "When machines master language, the era of AGI will arrive." He argues that language defines the boundaries of cognition; mastering language rules means mastering knowledge, reasoning, communication, and culture. In the AI era, large models will "ignite intelligence" across society, much like chips did in the internet era.
However, objectively, China's large model development still lags behind OpenAI's "cost-no-object moon landing" approach. Wang Xiaochuan describes China's current state as "one step slower in vision, three steps faster in implementation." While the U.S. leads in innovation and vision, China excels in practical application. Baichuan's experienced search team has successfully built language model-based super apps, with market-validated tech, products, and commercialization capabilities. Their search expertise aids large model development and addresses issues like hallucinations.
Through "small, fast steps" leveraging past experience, Baichuan has become a leader among China's large model startups. Its open-source models surpassed 6 million downloads in three months, with Baichuan2 outperforming LLaMA2 in general and vertical domains. The October 30 release of Baichuan2-192K, supporting a 192K context window (350,000 characters per input), surpassed Claude2-100K globally.
Beyond team and experience advantages, cloud computing enables Baichuan's monthly model iterations. Wang Xiaochuan emphasized that rapid iteration and deployment rely on cloud computing. Baichuan and Alibaba Cloud collaborate closely on pre-training and deployment, successfully completing large-scale training tasks, reducing inference costs, and improving deployment efficiency.
Baichuan's open-source models on Alibaba Cloud's ModelScope community have attracted many developers. Leveraging the community, developers fine-tune models for vertical domains and integrate them into AI applications, enriching China's large model ecosystem.
Wang Xiaochuan stated that Baichuan will deepen collaboration with Alibaba Cloud to explore cutting-edge innovations, apply large models to more scenarios, and jointly foster a thriving ecosystem.