Two Key Directions for the Future of General Artificial Intelligence

The initial excitement surrounding ChatGPT's launch appears to be declining. Web analytics firm Similarweb reports a 9.7% decrease in global visits to the ChatGPT website in June 2023—the first month-on-month drop noted. In the U.S., the decline was even sharper, at 10.3%. This suggests that engaging in conversation is just one dimension of what large language models can offer; merely focusing on entertainment like poetry and art will not revolutionize society.

Since their introduction, large conversational models have increasingly infiltrated various professional fields, transforming daily operations. A study in JAMA Internal Medicine found that doctors typically provide answers to patient inquiries averaging 52 words, whereas chatbots offer about 211 words per response. These chatbot interactions are not only lengthier but also demonstrate superior quality and empathy, with 78.6% of participants preferring chatbot responses over those from physicians.

In China, numerous companies are exploring the potential of large models across diverse industries. During the recent World Artificial Intelligence Conference (WAIC) in Shanghai, many businesses, including JD.com, highlighted their innovative solutions, addressing specific industry needs. Some companies are customizing general large models for niche applications, while others are building vertical models from scratch.

Looking forward, the evolution of AI is poised to extend beyond enhanced contextual understanding. He Xiaodong, President of JD Technology's Intelligent Services and Products Institute, advocates for a multimodal approach for genuine AI advancement. He emphasizes, "Humans remain central in this evolution; all technologies must ultimately serve humanity." Effective communication through language, vision, and voice will be vital to achieving this goal, underscoring the necessity of mastering multimodal capabilities for improved service delivery.

In today’s competitive landscape, “grounding in context” stands as a primary objective for large models. One key discussion at this year’s WAIC centered on practical implementations of these models. Large models should extend beyond simple conversation; they inherently improve productivity through intelligent interaction. Thus, their successful application must be tailored to specific industry requirements to enhance productivity.

While there are challenges, China possesses unique advantages in certain sectors that may foster global competitiveness in large model technologies. Data from the China Academy of Engineering reveals that as of late 2022, the U.S. held 36% of global computing power, while China accounted for 31%. Although a gap exists, it is smaller than it seems; for instance, in 2021, the U.S. represented 15% of the global intelligent computing market compared to China’s 26%.

Despite these opportunities, obstacles remain. Advancements in deep learning frameworks are necessary, as well as the integration of various large models for practical use. The availability of training data in Chinese is also insufficient. Additionally, reliance on NVIDIA's A100 chips, which face export restrictions, poses significant hurdles.

The current global AI race reveals that effectively leveraging large models presents considerable challenges, yet it offers China a pathway to catch up. While data, computing power, and financial resources are crucial, the role of "context" is paramount. Ultimately, large models are transforming how we access information and services. They must ensure not only information accuracy but also enable a deep understanding of human intent to efficiently complete tasks. Achieving this level of precision relies heavily on comprehensive contextual understanding.