Recent comments from Baidu’s CEO, Li Yanhong, have shed light on the misconceptions surrounding large-scale AI models. In an internal meeting that has now come to public attention, he expressed concerns about the growing disparities among these models as they evolve. According to Li, while the potential of large models reaches significant heights, there remains a considerable gap between current capabilities and ideal performance. He emphasized the need for continuous and rapid updates to these models to keep pace with user needs and to enhance efficiency while reducing costs.
Li addressed the prevailing industry notion that there are no longer barriers separating the capabilities of different large models, countering this perspective with his own. He noted that every new model release typically invites comparisons to GPT-4, often boasting competitive scores or even surpassing it in specific areas. However, Li cautioned that such comparisons do not accurately reflect the overall gaps that exist between leading models.
He elaborated that many models engage in benchmarking post-release to validate their performance, yet while they may appear similar based on scores, practical applications reveal clear differences in capabilities.
Highlighting the multi-dimensional nature of these disparities, Li pointed out that while the industry often emphasizes skills such as understanding, generating text, logic, and memory, it tends to overlook critical factors like cost and inference speed. Some models, although they may achieve similar outcomes, do so at a higher cost and slower speed, making them less viable than their advanced counterparts.
Li also discussed a shift in perceptions about open-source models. Historically, the term “open-source” was synonymous with low costs or free access, as exemplified by Linux. However, he argued that in the realm of large models, inference is costly, and open-source models do not come with built-in computational resources, which must be purchased separately. This makes efficient utilization of computational power more challenging.
He stated, “Efficiency-wise, open-source models fall short,” and clarified that closed-source models, referred to as commercial models, benefit from shared development costs and optimized resource use among numerous users. Baidu’s large models, such as Wenxin 3.5 and 4.0, reportedly achieve GPU utilization rates exceeding 90 percent.
In educational and research contexts, he acknowledged the value of open-source models but maintained that in commercial settings—where efficiency and cost-effectiveness are paramount—open-source options do not provide a competitive edge.
Li envisaged the evolution of AI applications, starting with tools like Copilot that assist users, progressing to intelligent agents that exhibit a degree of autonomy in utilizing tools and self-evaluation, ultimately leading to the development of AI workers capable of independently managing diverse tasks.
Despite the growing excitement around intelligent agents, Li noted that a consensus on this direction has not yet been widely established. Companies like Baidu that position intelligent agents as a crucial strategic focus are relatively few.
Moreover, he explained that the barrier to entry for creating intelligent agents is relatively low, as many do not yet know how to leverage large models for practical applications. Building intelligent agents on top of these models offers a straightforward and efficient path to harnessing their capabilities.