The rapid advancement of artificial intelligence (AI) technology, particularly through the emergence of sophisticated large models, is significantly reshaping industries around the globeThese advanced models, endowed with powerful digital processing capabilities and deep learning potential, are merging with various fields, becoming pivotal drivers of innovation and the catalyst for a new era of productivityAs industries grapple with these changes, a defining question arises: How will large models transform our lives, empower various sectors, and what direction will their future development take? Experts convened at the 2023 China Artificial Intelligence Industry Annual Conference to delve deeply into these pressing topics.

The conference opened with an intriguing address from Professor Xu Mai of Beihang University, who showcased a speech generated by the domestic large model Kimi

Advertisements

He presented the notion that large models and general artificial intelligence represent a vibrant frontier leading into the future, enriching interactive experiences and marking a true evolution of artificial intelligence capabilities.

The scope of these large models is impressive, not merely due to their extensive parameter scales but primarily because of the vast potential and diverse application scenarios they promiseBeyond standard applications such as content generation, these models are making inroads into a plethora of sectors including autonomous driving, smart healthcare, and the industrial InternetThey are actively fostering empowerment across various industries, clear indicators of their transformative power.

Zheng Qinghua, an academician of the Chinese Academy of Engineering and the president of Tongji University, noted the revolutionary heights reached by large models in AI, likening their capabilities to the monumental shift in efficiency experienced in biochemical research

Advertisements

In the past, predicting the precise structure of proteins could take months and require teams of thousands; now, with the aid of large model analytics, results can be generated in mere minutes.

The surge in generative large models like Moon's Dark Side Kimi, Baidu's Wenxin Yiyan, and iFlyTek's Spark has catalyzed rapid advancements in the application landscapeLi Cong, Vice President of iFlyTek and Director of its Research Institute, highlighted that Spark, as an exemplar of domestic general large models, has notably partnered with Chery Automobile to develop an intelligent cockpit, thereby becoming a pioneer in applying local models in that sector effectively.

However, it is essential to recognize that not every business requires the comprehensive functions of a general large model; specific scenarios often demand precision insteadZi Ran, head of the Security GPT business at Sangfor Technologies, expressed that achieving operational efficiency and market readiness might be better suited to vertical large models that can cater to specialized needs effectively

Advertisements

He noted that current cyber attackers have turned to leverage these large models to refine their methods, posing challenges to traditional detection mechanismsThe Security GPT model, in contrast, excels in recognizing malicious code and understanding cyber adversarial tactics more effectively than general models, offering lower inference costs and higher accuracy as it serves over 130 companies.

The effective application of vertical large models extends into environmental pollution management, another critical arenaA team led by Qiao Junfei, Vice President of Beijing University of Technology, reported significant innovations in intelligent feature modeling, self-organization control, and dynamic optimization for pollution preventionHistorically, environmental management struggled to create analytical models and relied heavily on human decision-making; however, the implementation of vertical large models facilitates a more objective basis for pollutant treatment decisions, enhancing practices in scientific and targeted pollution control.

While large AI models bear the promise of transforming production and everyday life, challenges remain in data utilization, computational power, and algorithmic efficiency

• Boosting Growth: The Role of Efficient Circulation
• Challenges in Reforming the EU Electricity Market
• Integrating Innovative Resources
• Is a 4.8% U.S. Treasury Bond Worth Buying?
• Interest Rate Cuts, But Funds Are Unavailable

On the computational front, Li Cong explained that the sheer capacity of AI models is fundamentally linked to power availability, which serves as the 'fuel' for these technologiesTo democratize computational power, much like electricity usage, issues around data security and result transmission must be addressedCollaborative efforts are underway to establish a comprehensive "China Computational Power Network," spearheaded by the Pengcheng Laboratory, which has exhibited remarkable results in building powerful computational nodes crucial for developing the domestic ecological framework.

Moreover, the importance of high-quality data as the Continuous 'raw material' for large models cannot be overstatedZheng Qinghua warned that high-value data is akin to mineral resources, finite and not endlessly extractableSome experts forecast that by 2026, the exploitable value of vast large-scale corpuses would be largely exhausted, posing difficulties in training new valuable data from existing big data

Presently, Chinese data constitutes a minimal part of training data pools; statistics reveal that less than one-thousandth of the data used in training models like ChatGPT is sourced from Chinese content.

The industry is also grappling with the phenomenon known as catastrophic forgetting—a situation where training on new tasks can adversely affect the performance of models on previously learned tasksIn practical applications like autonomous driving, human drivers draw from memory to navigate roads, while AI-powered systems lack this capability and must recalibrate every time, leading to excessive consumption of computational resources and energy.

Addressing how to enhance large model technology from weak to strong, Zheng Qinghua proposed three essential technological pathwaysFirstly, the advancement of models through the synergy of vast data, robust computational power, and sophisticated algorithms requires extending existing technological trajectories

Secondly, integrating 'neural + symbolic' approaches could combine the learning robustness and universality of neural networks with the interpretability and composability of symbolic reasoningThirdly, machine memory models inspired by human memory might resolve inherent limitations observed in current models.

As the discussion progressed towards future trends for large models, many experts highlighted their potential integration with embodied intelligence, with humanoid robots serving as prime examplesThe relationship between large models and robotics is particularly intriguingLi Cong illustrated this with a scenario where retrieving a bag of chips from a drawer would require the model to understand and plan the sequence of actions involved, a function termed as 'disembodied intelligence.' However, coupling with embodied intelligence allows for the decomposition of these tasks into executable commands that a robot can perform, enhancing practical agency.

Researcher Zhao Jian from the Northwestern Polytechnical University indicated that continual expansion of AI models into multimodal fields presents vast possibilities for understanding, decision-making, and perception, paving the way for effective responses to the myriad complexities faced in the real world

This convergence of embodied intelligence and multimodal models could act as a vital bridge connecting AI with everyday experiences and applications across diverse sectors.

The establishment of a robust digital industry cluster that embraces research and development in big data and AI is crucialZheng Qinghua posits that the "AI+" initiative, aiming at innovative development, should actively implement holistic measures concerning education and talent cultivationIt is not only essential to convey principles and methodologies but also vital to engage in practical actionsSuggestions include integrating AI capabilities into educational curriculums and reengineering traditional experimental platforms to enable hands-on research and application experience in the entire lifecycle of AI technology.

Industry members were encouraged by Zi Ran to embrace large models dynamically across sectors, exploring pathways for more profound integrations to harness their potential effectively for growth