On June 14, 2024, the Beijing Zhiyuan Conference featured Baidu's Chief Technology Officer, Wang Haifeng, who delivered insights on the evolution of artificial intelligence (AI). He discussed how large models signal a new era for artificial general intelligence (AGI), emphasizing two key aspects: the universality of AI technology and its comprehensive capabilities.
Wang noted that the ongoing trend in AI technology supports the continued relevance of the scale law in the coming years. There remains significant potential for improvement in large language models (LLMs), and as multi-modal models become more user-friendly, agent technology is also progressing, leading to groundbreaking application advancements.
New Perspectives on General Artificial Intelligence
AI aims to replicate and enhance human intelligence. Over the years, AI has transformed from manually crafted rules and statistical methods to today's deep learning, where cohesive architectures can solve a range of challenges. The introduction of large models has further generalized algorithms and applications.
Today’s large models excel in handling various tasks, languages, modalities, and contexts. In natural language processing, previously separate areas like tokenization, syntax analysis, semantic matching, machine translation, question-answering, and dialogue are now largely managed by single large language models. These models bridge gaps between human languages and structured formal languages, enabling integrated multi-modal applications across diverse industries.
Comprehensive Capabilities of AI
Wang identified four foundational capabilities of AI: understanding, generation, logic, and memory. Additional functions—like creativity, problem-solving, coding, planning, and decision-making—arise from these core abilities. The enhancement of these capabilities brings us closer to realizing general AI.
Insights into Wenxin Large Model Technology
Wenxin Yiyan is Baidu's next-generation knowledge-enhanced language model, designed using advanced platforms, top-tier data, and optimized algorithms. This model learns from extensive datasets to enhance knowledge retrieval and conversational capabilities. Key innovations in foundational model training, data enhancement, alignment technology, prompt optimization, and agent mechanisms represent major advancements.
Agents built on foundational models refine cognitive processes through supervised fine-tuning, preference learning for decision-making, and reinforcement learning for outcome evaluation, resulting in a robust thought model. Code agents utilize these thought models to interpret user needs and translate them into executable code, streamlining complex tasks.
Since Baidu began investing in AI in 2010 and launched the Wenxin Model 1.0 in March 2019, the company has continually evolved the model, with Version 4.0 released last October. This quick progress is fueled by Baidu's comprehensive strategy encompassing chips, frameworks, models, and applications, with a focus on the PaddlePaddle deep learning platform.
Industrial Production and the Future of AI
Wang asserted that the scale law will remain significant in the coming years. The power and rapid advancement of large language models point to an exciting future full of potential for enhancement. The practicality of multi-modal large models is poised to improve, and agent technology is expected to continue its growth.
He drew parallels with past industrial revolutions—powered by mechanical, electrical, and informational technologies—which have universal applications across industries. As these technologies move toward standardization, automation, and modularity, they pave the way for industrial-scale production. AI, underpinned by deep learning and large model engineering platforms, shares this universal applicability and is marked by trends of standardization, modularity, and automation. This positions deep learning and large model platforms at the forefront of ushering AI into a new era of industrial production, advancing us towards general artificial intelligence.
Ranking
