Chinese AI has the potential for significant growth by leveraging larger models and more data, says Stepfun founder Jiang Daxin. Despite facing challenges in investment and access to advanced chips, Jiang remains optimistic about the future of artificial intelligence development in China.
During the World Artificial Intelligence Conference (WAIC) in Shanghai, Jiang highlighted the importance of scaling laws in the development of large language models (LLMs). These laws indicate that larger models, more data, and increased computational resources lead to improved performance, although with diminishing returns. Major tech companies have been investing heavily in advanced technologies like Nvidia chips to enhance AI model performance.
Jiang noted the success of OpenAI’s GPT series and the investments made by tech giants like Amazon, Microsoft, and Meta in supercomputing centers as evidence of the effectiveness of scaling laws. He emphasized the need to address challenges related to data availability, human resources, and return on investment to accelerate the pace of scaling laws.
In China, numerous Big Tech firms and start-ups have launched their own LLMs following the release of OpenAI’s ChatGPT. However, many Chinese companies have struggled to match the level of investment in LLMs seen among US tech giants. Stepfun, founded in 2023, has taken a different approach by focusing on developing foundational models like Step-2, a trillion-parameter LLM, along with multimodal and image generation models.
Jiang highlighted the importance of multimodality in constructing comprehensive world models that can process different types of data. Stepfun aims to integrate generative and comprehension capabilities in a single model to enhance AI performance. The company also offers consumer-facing products like Yuewen, a ChatGPT-like personal assistant, and Maopaoya, an AI companion with unique personalities.
With global AI investments reaching $22.4 billion last year, the focus on companies developing large models has been significant. According to Alex Zhou Zhifeng, managing partner at Qiming Venture Partners, future investments will likely prioritize AI applications benefiting from decreasing token costs. Peng Wensheng from China International Capital projects the AI model market in China to reach about 5.2 trillion yuan by 2030.
In conclusion, Chinese AI development can thrive by embracing larger models, more data, and advancements in multimodal capabilities. With the right investments and strategic focus, China has the potential to become a leader in AI innovation on a global scale.
