TMTPOST — Robin Li, the founder, chairman, and CEO of Baidu Group,   on Wednesday   refuted   the argument of   no longer   existence of barriers to the capabilities of large AI models, citing   computational power as   a critical factor in determining the success of these models   and inadequate current efficiency of open-source models for future development.  

Li made the remarks during in a closed-door   address, predicting   that the gap between AI large models from China and OpenAI's GPT will likely widen in the future.

Since the launch of OpenAI's ChatGPT on November 30, 2022, Baidu has been active in the AI space with its Ernie Bot   model, which debuted on March 16, 2023. The model has since undergone multiple iterations, with versions 3.5 and 4.0 enhancing its capabilities. Baidu ’ s AI model now generates approximately 10% of the traffic on its search engine and supports around 250 million users daily.

Earlier this year, Li voiced concerns about Baidu ’ s financial strategy, criticizing the company's approach to short-term losses and emphasizing the need for improved business acumen among employees. He highlighted the importance of revenue quality and criticized the slow internal response of large companies, which he believes hinders innovation.

In November 2023, during an event in Shenzhen, Li expressed skepticism about the proliferation of large models in China, noting that many of the 200-plus domestic models have low usage rates. He suggested that Baidu ’ s Ernie Bot   model was far more frequently used compared to others, implying that excessive duplication of large models   represents a waste of resources.

In December 2023, at a Beijing event, Li reiterated that the proliferation of large models is a significant waste of resources. He argued that more effort should be directed towards creating impactful applications rather than producing numerous models. Li also stressed   the need for companies to utilize large models to drive core business metrics, rather than merely focusing on model performance in isolated benchmarks.

At the April 2024 Shenzhen AI Developers Conference, Li highlighted the inefficiencies of open-source models, particularly in terms of computational power and cost. He pointed out that while open-source models may be cheaper initially, they often lack the efficiency and effectiveness of commercial models.

According to Li, open-source models do not provide the necessary computational resources to compete with commercial offerings, which are designed to share resources and costs among multiple users.

Regarding AI large model agents, Li mentioned that this is a non-consensus area. AI agents provide a very direct, efficient, and simple way to build intelligence on top of models. Meanwhile, he believes that Baidu's AI agents are at the forefront, and there are not many companies that see agents as the most important strategy and development direction for large models, as Baidu does.

"Why do we put so much emphasis   on AI agents? Because the threshold for agents is indeed very low. Last year, we talked about diving into applications and everyone started working on applications, but many people still said they didn ’ t know how to proceed or whether this direction would work out," said Li.

"There are countless uncertainties about what capabilities are needed to create value in a given scenario. People don ’ t know how to transition from models to applications. However, agents provide a very direct, efficient, and simple way to build intelligence on top of models. That ’ s why every week, tens of thousands of new agents are being created on the Ernie Bot   platform,"   Li added.

Li emphasized that the development of large models inevitably goes through several stages. Initially, models assist humans, and at the end, humans need to give the final approval to ensure that the outcomes are satisfactory in all aspects before they are released. This is the Copilot stage.

The next stage is the "Agent,"   where agents have varying definitions but mainly possess a degree of autonomy, including the ability to use tools independently, reflect, and evolve. Further automation progresses into what is termed an AI Worker, which can perform various cognitive and physical tasks independently, like a human.