China's Large Models Lag Behind ChatGPT-4 by at Least Two Generations

Currently, there are two prominent perspectives on the commercialization of large language models. The first advocates for creating a universal large model similar to the Android operating system, while the second emphasizes using large model technologies to develop AI-generated content (AIGC) applications in specialized fields. These include dedicated writing platforms like Notion, intelligent design tools like Midjourney, and AI programming assistants like Copilot.

In the domain of general-purpose large models, OpenAI released the multimodal capabilities of ChatGPT-4.0 in March, establishing it as the most advanced model to date. Compared to its predecessor, ChatGPT-3.5, it exhibits substantial improvements in processing complex long texts, converting between images and text, and enhancing accuracy and emergent intelligence. Major Chinese companies like BAT, ByteDance, Huawei, iFlytek, and Kunlun Wanwei are also debuting their large-scale models. However, Zhang Peng, CEO of Zhipu AI, asserts that the overall performance of China's large models is on par with GPT-3, possibly slightly lagging.

ChatGPT is an upgraded version of GPT-3.5, showcasing notable performance enhancements that lead to superior outcomes. The recently introduced ChatGPT-4.0 is said to surpass GPT-3 by two generations, reflecting its robust capabilities. Consequently, Chinese models appear to be at least two generations behind ChatGPT-4.

Large models function similarly to operating systems, offering a diverse range of utilities. ChatGPT-4 is particularly distinguished for its ability to process text while also understanding images. It can manage texts of up to 25,000 words, making it ideal for writing scripts or short stories. OpenAI has demonstrated GPT-4's capability of automatically generating code for webpage development by identifying website wireframes.

Following GPT-4's release, Microsoft integrated its functionality into the latest version of Bing, initially focusing on AI text capabilities. CEO Satya Nadella has claimed that their search engine has surpassed Google since incorporating GPT.

However, large models' generalization capabilities may not adequately address every query. For example, when predicting stock price movements, GPT-4 tends to offer vague and broad insights, which may lack practical value for professional investors. Cheng Hao, founder of Yuanwang Capital, argues that each industry contains specialized technologies and insights often confined to private databases or restricted to a few experts.

To effectively handle high-value, specialized tasks, vertical models are crucial. BloombergGPT serves as an example, tailored specifically for the financial sector, assisting professionals in interpreting and analyzing financial data. Cheng believes the potential for vertical models is vast and that their development will spawn numerous entrepreneurial opportunities across industries. Creating vertical models is often less costly, as companies can utilize a well-pretrained large model and fine-tune it for specific applications. Unfortunately, many Chinese firms focused solely on the ChatGPT model tend to prioritize general-purpose models, which Cheng argues creates limited opportunities.

OpenAI's success with GPT-4 can largely be attributed to Microsoft's support. Estimates indicate that ChatGPT attracted an average of 13 million unique visitors in January, generating demand for over 30,000 Nvidia A100 GPUs. The initial investment cost was around $800 million, with daily electricity expenses reaching approximately $50,000. Therefore, it is imperative for companies in the GPT landscape to concentrate on developing vertical models tailored to their specific business needs to fully leverage the potential of AIGC.