AI News

Li Yanhong in Dialogue with Yang Yuanqing: What Pitfalls Has Baidu's AI Transformation Stumbled Into, and How Can Enterprises Build Intelligent Organizations

YicaiUpdated Nov 12 12:13

AI summary

At the 2026 Baidu World Conference, Li Yanhong and Yang Yuanqing held a dialogue. Li Yanhong reviewed the pitfalls Baidu encountered in its AI transformation: business departments independently conducted SFT based on open-source models, and as soon as the foundation model was upgraded, they fell behind; later this was centralized under a dedicated model department for unified training. Various departments chased the agent hype, leading to scattered resources; Baidu then changed to holding monthly meetings requiring business units to focus and do fewer things. Li Yanhong proposed three principles for building an "intelligent organization": breaking down job boundaries, accumulating tacit knowledge, and designing closed-loop processes. Both agreed that it is still too early to discuss AI input-output ratios in the early stages of revolutionary technology development.

Body · AI translation

“Right now, AI still can't replace the CEO on a business trip.”

On November 12, at the 2026 Baidu World Conference, the conversation between Baidu Chairman Li Yanhong and Lenovo Group Chairman and CEO Yang Yuanqing began with a light topic.

But aside from CEO business trips, what other things can't AI replace? As employees use AI more and more, how should companies manage it? In the process of rapid AI technology iteration, should companies encourage employees to experiment boldly, or concentrate resources and plan in a unified way?

The two leaders of tech companies shared their respective experiences in the conversation. Li Yanhong noted that over the past few years, Baidu also stumbled in some areas during its AI transformation. From various business units training their own models to chasing agent application trends, Baidu gradually discovered that employees' enthusiasm for using AI does not necessarily mean organizational efficiency will improve.

Yang Yuanqing, on the other hand, began with Lenovo's supply chain management practices. In his view, even if AI can predict market demand more accurately than humans, it is still people who ultimately make judgments and decisions. As for the AI input-output ratio currently hotly discussed in the market, both believe that in the early stages of a revolutionary technology's development, it is too early to talk about input-output ratio.

What pitfalls has Baidu stumbled into?

Over the past decade, artificial intelligence has gradually formed an independent technology stack, including a four-layer architecture of chips, cloud infrastructure, models, and applications. Li Yanhong believes that if energy is the upstream of AI technology, then the downstream of AI technology is organizational capability—an AI-native organizational system different from the traditional bureaucratic hierarchy, which he calls an “intelligent organization.”

“AI is not just a technological revolution; it is even more an organizational revolution,” Li Yanhong said. In the coming years, AI will determine how companies divide labor, how they collaborate, and how they create value. Companies cannot treat AI merely as a tool for individual productivity; they must let agents and digital employees truly enter business links such as sales, customer service, R&D, and operations.

Regarding how to build an intelligent organization, Li Yanhong proposed three principles for organizational evolution: breaking job boundaries, accumulating tacit knowledge, and designing closed-loop processes.

“Jobs are the unit of the industrial age, while tasks are the unit of the intelligent age,” Li Yanhong believes. In the future, companies will shift from being job-centered to task-centered. One person can simultaneously take on tasks from multiple jobs, and their cognition, experience, and judgment can also be reused multiple times. At the same time, companies need to accumulate the experience and knowledge originally residing in employees' minds onto platforms so that AI can understand and invoke them, thereby transforming individual capabilities into organizational capabilities.

However, moving from a traditional organization to an intelligent organization is not achieved overnight. Over the past few years, Baidu has also gone through a process of continuous trial and error.

Li Yanhong gave an example: when large models first emerged, many Baidu engineers wanted to train models themselves, performing supervised fine-tuning (SFT) based on open-source models. These models initially worked well when applied to business, but after a while, once the foundation model was upgraded, the original models fell behind.

“Later we required business units not to train models themselves; only a dedicated model department could be responsible for model training,” Li Yanhong said. This effectively avoided “entropy increase,” and work efficiency also improved.

Similar problems also appeared at the agent application level. As various agent applications rapidly emerged, Baidu's business units also wanted to chase new trends—whatever agent became popular, they wanted to build that kind of agent.

“Later we also found that this would spread resources very thinly, and the success rate in each direction would become even lower,” Li Yanhong said.

To this end, Baidu holds meetings every month, requiring business units to continuously focus and reduce the number of directions they pursue from many to few. “We require everyone to do fewer things, so that efficiency can improve and the success rate can increase.”

In the AI era, how do CEOs manage companies?

Compared with Baidu, Lenovo, founded in 1984, has gone through more rounds of technological change. The first question Li Yanhong posed to Yang Yuanqing was: over the past three-plus years, under the AI wave, what changes have occurred in Lenovo's management philosophy, internal organization, and processes?

Yang Yuanqing called the fifth decade of Lenovo, which began in 2024, the “fifth entrepreneurship.” This time, Lenovo's entrepreneurship revolves around AI, and its three major businesses—intelligent devices, intelligent infrastructure, and intelligent solutions and services—will all change because of AI.

When talking about the intelligent devices business, the two company leaders coincidentally mentioned the importance of PCs in the AI era.

Yang Yuanqing said that PCs and phones are both where AI data and application scenarios occur, and they also have on-device computing capabilities. After matching with corresponding models, they can become the starting point for personal intelligence.

Li Yanhong also observed a similar trend from Baidu's business. He said that some of Baidu's agent projects have much higher user activity on PCs and also greater value.

In Yang Yuanqing's view, future personal devices will be used not only by people but also by agents. And when a person owns multiple devices, how to connect the AI experience across different devices, integrate scattered data, form short-term and long-term memory, and make the personal super agent understand the user better is a direction Lenovo is exploring.

Inside the company, Lenovo employees are also highly enthusiastic about using AI. Yang Yuanqing listed a series of scenarios: R&D staff use AI for programming, sales staff use AI to find customers, marketing staff use AI for creative work, service staff use agents to serve customers, and even the human resources department uses AI to match suitable candidates from tens of thousands of applicants, while the legal department uses AI to review contracts.

Li Yanhong was more interested in Lenovo's experience in supply chain management.

“After AI arrived, for companies like ours in the cloud business, one particularly personal experience—even a pain point—is that in the past we were unfamiliar with supply chain management and did not have much experience with it,” Li Yanhong said. Lenovo has years of accumulated expertise in this area. In the AI era, how can the supply chain be managed well? And can AI accurately predict this kind of explosive growth in demand?

Yang Yuanqing said Lenovo already uses AI in multiple areas, including demand forecasting, production scheduling, quality inspection, and logistics and transportation. Enterprises can forecast demand based on market data, user demand, supply conditions, and past transaction records. Compared with the past, when decisions relied more on human gut feeling, now with data, AI can better help enterprises analyze and make decisions, and demand forecasts will be much more accurate.

But forecasting demand is only the first step. Yang Yuanqing said supply chain management affects every part of the operation: once demand rises, whether procurement, production, and supply can be adjusted in a timely and rapid manner is equally important.

Li Yanhong said that during Baidu's transformation, one fairly obvious lesson was that one cannot rely too heavily on bottom-up momentum. Frontline employees are all highly enthusiastic and want to use AI tools, but at the organizational level there still needs to be overall consideration of when human intervention is needed and when AI technology should be used.

Yang Yuanqing believes top-down and bottom-up approaches should be combined. Enterprises need an overall plan that clarifies what can be done and what cannot be done, and draws boundaries. But in the early stage of a new technology's development, employees should still be encouraged to practice boldly and to use AI.

At the end of the conversation, Yang Yuanqing also shared an observation from recent exchanges with global customers. CIOs and CTOs at many Fortune 500 companies have almost the same view: in the early stage of a revolutionary technology's development, it is not appropriate to emphasize input-output ratios too early.

“Don't worry too much about input-output for now.” Li Yanhong, picking up the thread, agreed with this view. In their view, enterprises' exploration of AI is still at an early stage, and it is not yet fully clear from which links new productivity will emerge and in what way it will be released. Rather than rushing to measure returns on investment, what matters more at this stage is encouraging practice and discovering, through exploration, the ways in which AI creates value.

Source

Yicai