From Small Models to Large Models: Analyzing the Technology Pathways of Specialized and Generalized Artificial Intelligence

Future Development Pathways for General Artificial Intelligence

The future of general artificial intelligence (AGI) is shaped by two primary technological paths. The first is the continuation of large model approaches, which utilize enhanced computing power to expand the scale and capabilities of these models, enabling deeper integration into various industries. The second path involves the exploration of emerging technologies such as reinforcement learning, knowledge computation, symbolic reasoning, and brain-like computing.

During the 2024 Industrial Technology Innovation Conference in Shanghai, Qiao Yu, assistant director of the Shanghai Artificial Intelligence Laboratory, highlighted that AI is currently at a crucial transitional phase from specialized intelligence to general intelligence. He noted that since the advent of deep learning in 2010, the AI landscape has experienced multiple transformations. Initially focused on developing small, specialized models to meet specific application needs, the field has shifted dramatically with the introduction of large models driven by Transformers, big data, and self-supervised learning by 2020, allowing AI agents to tackle a variety of tasks.

The rise of large models, particularly under the leadership of OpenAI, has led to centralized industrialized research and development, resulting in groundbreaking innovations like ChatGPT and GPT-4. Looking ahead, future advancements will not rely on isolated successes but rather on collaborative innovation across various domains, including chips, internet infrastructure, frameworks, data, models, evaluation, and deployment, to achieve comprehensive optimization.

Qiao emphasized that while scaling up models can provide certain advantages, it also presents challenges, such as efficiency, reliability, and safety. Addressing these issues will require exploring knowledge and symbol-based reasoning methods, which can offer greater interpretability and enhance safety.

Innovation in the future will increasingly depend on a systematic, multi-faceted approach. With Shanghai's rich B2B landscapes in finance, urban development, and manufacturing, Qiao advocated for collaboration with leading enterprises to develop specialized models in vertical domains. This strategy aims to accelerate the creation of an open ecosystem and an ethical evaluation framework, fostering healthy development and standardized applications.

The future of general artificial intelligence not only calls for technological innovation but also necessitates a focus on optimizing the environmental ecosystem to address the multifaceted challenges we face today.