AI Boom Sparks Sobering Reflections on Fierce Model Competition

The recent World Artificial Intelligence Conference, held in Shanghai, has concluded, but the fervor ignited by AI models like ChatGPT continues to shape the landscape of technology and business. The consensus among industry leaders appears to emphasize the importance of early involvement in AI development, coining terms like "getting on the board early" and claiming the "first move advantage." This reflects a broader ambition to propel the industry into what some are calling the "Battle of a Thousand Models."

In early July, the air buzzed with excitement as the AI industry gathered in Shanghai, marking a record high in both the number of participating companies and the exhibition space devoted to AI innovations at the 2023 World Artificial Intelligence Conference. Despite the sweltering heat and strong winds, attendance remained robust, with many families attending to catch a glimpse of the latest advancements. Scalpers were even spotted outside the conference hall, capitalizing on the disparity between demand and supply.

Advertisement

However, amidst this enthusiasm, it is essential to acknowledge that large models still face considerable challenges, notably in areas like robustness, compliance, and trustworthiness. In contrast to developed nations, China continues to grapple with issues regarding chip technology, computational power, and data infrastructure. One of the pressing hurdles is the scarcity of high-quality Chinese language data, which significantly constrains the applicability and development of these large models.

As China continues to address these critical issues, the question arises: what developmental path should the country's AI sector pursue? Throughout three days of forums and interviews with numerous industry experts, the most common answers revolved around the ideas of "vertical integration" and "practical applications." Experts suggest that early experimentation with applications in specific vertical sectors and the construction of composite AI systems reflect current trends within China's AI realm.

"Getting on the Board Early"

The development of the digital economy has emerged as a global consensus. AI, as an essential emerging technology, is increasingly recognized as the cornerstone of industrial upgrades and enhancements in productivity. The introduction of OpenAI's ChatGPT in November 2022 sparked a new wave of innovation within the AI domain, leading to a national surge in AI model announcements and rapid iterations by several Chinese firms.

The World Artificial Intelligence Conference in 2023 spotlighted these large models, showcasing an impressive array of innovations. From Baidu's Wenxin Yiyan to Alibaba Cloud's Tongyi Qianwen, from Huawei’s PanGu to iFLYTEK's Xinghuo, the range of models unveiled left attendees dazzled.Notable figures, like Lin Jianming, founder and chairman of Samoyed Cloud Technology Group, shared insights on this new industrial trend, highlighting that major players like Baidu, Alibaba, and Huawei are engaging in a "four-in-one" strategy, targeting advances from computational resources to applications.

According to Lin, the current landscape of AI models in China largely consists of those with parameters in the hundred billion range and above. Initially, enterprises are focusing on internal applications, gradually expanding their services toward B2B clients. The pressure to capitalize on the monumental shift towards large models is palpable, with companies of all sizes racing to be players in this evolving game. The mantra of "getting on the board early" underscores the urgency to grasp the rules of this burgeoning field.

Furthermore, Zhou Bowen, an esteemed academic and founder of Xianyuan Technology, emphasized the importance of an innovation-driven, self-sufficient ecosystem for developing large language models and generative AI technologies. He stressed the need for diverse commercial applications and academic innovations, advocating against a monolithic approach to AI solutions.

Complex Challenges

While the AI boom may be driving excitement, substantial challenges loom over large models, including robustness and compliance. Lin candidly pointed out the significant gaps that exist between China and the international community when comparing AI chips, patents, and thriving ecosystems for innovation. Major obstacles confronting AI model development in China include the following: first, the high demand for significant computational power to train large models, an area where China remains behind; second, the lack of high-quality Chinese language datasets; third, a limited pool of qualified professionals who can drive foundational research.

The finance sector, in particular, presents unique challenges, insisting on elevated levels of risk management and security. Lin noted that developing financial AI models entails navigating numerous risks, including trust, stability, and compliance. The intricacies of creating reliable systems capable of maintaining stability during unpredictable events are particularly daunting.

Jiang Ning, Chief Information Officer of Mashang Consumption, echoed these sentiments by pointing out that inconsistencies in dynamic decision-making and adaptability remain critical concerns for AI models. The ability to minimize noise and distractions during unexpected occurrences is of utmost importance.

Furthermore, Zhou noted the ongoing lack of original breakthroughs within China's AI models. Challenges in model inference abilities persist, primarily due to the scarcity of high-quality Chinese-language corpora, particularly in domains that demand strong privacy and compliance standards, such as finance.

Furthermore, Zhou highlighted more obstacles related to the commercial viability of large AI models, including divergent data quality, training complexity, and high performance expectations. Thus, the feasibility of widespread commercial application of large models rests on algorithmic and data-driven support systems that can bolster their development.

Vertical Integration

With numerous core challenges still ahead, what pathway should China's AI sector pursue? Jiang emphasized the significance of building composite AI systems that incorporate various specialized models while also harnessing the adaptive and generalized skills of generative models. This duality can lead to a more efficient application of AI technologies across various industries.

Lin suggested that AI developments have vast potential across numerous sectors, including smart cities and enterprise digitization. Investments in AI models should enhance central competitiveness by focusing on computing power, algorithmic advancements, and talent acquisition tailored to national strategies and sectoral demands, allowing for an exploration of industry-specific challenges.

Additionally, he noted the importance of utilizing existing technologies and industry-specific knowledge to create vertical models. By leveraging the combination of general models alongside specialized intelligence tailored for specific sectors, there lies an opportunity to generate unique competitive advantages.

In Zhou's view, a transformative approach to large model commercialization involves an end-to-end focus, gradually refining business models that work in conjunction with core competencies. Training large models specific to vertical industries is vital to advancing technology and application scenarios.

The future of AI in China indeed lies in navigating fundamental disparities, while simultaneously striving for positions of application-level innovation and leadership. This vision underscores the need for ongoing vertical integration that spans from self-developed general models to their implementation in real-world scenarios, aiming to align generated intelligence with tangible commercial outcomes.

Zhou proposed that competitive pathways for startups can be categorized into three distinct strategies: the first involves building proprietary general models with comprehensive capabilities; the second includes training business-specific models based on broad architectures like GPT; finally, the third path focuses on applying existing models for practical use while inherently possessing lower barriers to entry.