AI News

Robin Li Proposes "Intelligent Organization" at Baidu World 2026: The Key to Realizing AI Value Lies in Organizational Transformation

TMTPost · AGIUpdated Nov 13 19:02

AI summary

At Baidu World 2026 on November 12, Robin Li proposed the concept of the "Intelligent Organization," arguing that AI's failure to generate the expected economic benefits is like "swapping the motor but keeping the old factory building," and laid out three principles: breaking down job boundaries, accumulating tacit knowledge, and designing closed-loop processes. Baidu's Digital Human Innovation Business Unit merged product, R&D, design, and testing into the Builder role; in the third quarter it completed 564 product iterations, accumulated 130 reusable Skills, and increased per-capita output by 125%.

Body · AI translation

By Hu Shuochengli

In 2026, the AI industry faces an increasingly sharp question: money has been invested, technology is iterating at an astonishing pace, but where the real value growth it brings to the world will land, no one can say clearly.

This is a worldwide problem that seems to have no accurate answer. But at Baidu World 2026 on November 12, Robin Li offered his own diagnosis—the motor has been replaced, but the factory is still the old one.

Ten years ago, on the same stage, he was the one who said "AI is the next act," and he turned out to be right. Being right once is not hard; what is hard is that over the following decade, at every new milestone in AI's development, his judgments came first—and he delivered on them himself first.

Why is it always him?

01 Why is it always him?

November 12, 2026, Baidu World Conference. Robin Li stood on stage and, with a word that seems to have little to do with technology—"organization"—set the tone in advance for the next act of AI value realization.

Why organization? Robin Li offered his judgment: AI has now reached a new hurdle—the motor has been replaced, but the factory is still the old one. AI has not produced the expected economic benefits; it is not that the motor is bad, but that the "factory" has not kept up. He set three principles for the "new factory": break down job boundaries, accumulate tacit knowledge, and design closed-loop processes.

Whether this judgment holds still needs the test of time. But I have noticed a very interesting phenomenon: why is it that at every such milestone, one can always hear Robin Li raise a question worth the industry pondering further?

To find the answer, we must start with The Silicon Valley Business Wars. This is a chapter-style little book Li wrote in 1999; it was written in the United States but published in China. When he wrote it, it was precisely the eve of Li's return to China to found Baidu.

I once asked him to his face why, on the eve of starting his business, he still took the time to write a book. His answer was very specific: after returning to China, he repeatedly found that the domestic IT industry paid great attention to drawing on Silicon Valley's experience, but many of the Silicon Valley stories that circulated back were told in a plausible but inaccurate way and might mislead domestic entrepreneurs, so he wanted to write it clearly himself.

Then, he did not publish a book for many years. But as the AI era arrived, he seemed to enter a peak creative period—Intelligent Revolution in 2017, Intelligent Economy in 2020, and Intelligent Transportation in 2021.

These books do not feature him as the protagonist, nor are they startup memoirs; they are all judgments about industry trends. Among entrepreneurs of his generation, someone who continuously writes such "big books" of trend judgments is almost unique.

My understanding is that this is a kind of intellectual responsibility: things one has come into contact with first and figured out first should be spoken out, to give peers some reference and to leave later arrivals with a little less misunderstanding.

More importantly, he dares to "sing a different tune."

In 2016, when the mobile internet was still a bustling business, he had already pointed the "next act" toward artificial intelligence; in 2023, when the "hundred-model war" was in full swing, he said that without applications, foundation models are worthless; in May this year, Silicon Valley was keen on Tokenmaxxing. He proposed DAA, daily active agents; and this time, he has brought out the new topic of the "intelligent organization."

In other words, his views come from the practice he deeply loves.

In 2018, Baidu released China's first full-featured cloud AI chip. Li judged at the time that the explosion of AI applications would significantly increase demand for algorithms and computing power. In the same year, Amazon AWS launched its first self-developed CPU. Global tech giants successively deepened their layout in AI's underlying computing power.

At Baidu World 2024, Li demonstrated no-code generation capabilities on site and judged that "agents are the most mainstream form of AI applications." As a result, half a year later, Anthropic disclosed its Research multi-agent architecture, and the first year of agents began, with multi-agent division of labor and collaboration becoming one of its prominent features.

In May 2026, at the Create conference, Li proposed DAA (daily active agents) as the ruler for measuring AI value; afterward, Meta, Amazon, and others successively took down their token leaderboards: in September 2026, Anthropic disclosed that about 30,000 agents participated in R&D (daily active R&D DAA).

And this time, regarding how organizations can become intelligent, Baidu has already been practicing internally for a long time. For example, after employees use AI to complete tasks, how individual efficiency becomes team efficiency, how experience stays in the organization, and how the next piece of work can build on the previous accumulation—these are all problems one keeps encountering only by doing it oneself.

So, when he spoke last November about "internalizing AI capabilities," he had already made clear that for AI to deliver value, true internalization would first happen at the organizational level; in May this year, he went further and spoke of the organization's "self-evolution," just short of piercing the window paper.

And by this Baidu World, he finally gathered these understandings into the four words "intelligent organization."

He makes predictions not to shock people with startling words, but to help the industry break through.

Entering November, market expectations for Anthropic's IPO valuation and OpenAI's valuation target for a new funding round were adjusted one after another. The market is still willing to spend huge sums on AI, but questions about infrastructure investment, growth costs, and time to returns have become increasingly specific and cautious.

Li's proposal of the intelligent organization at this time is meant to explain what still needs to be added for AI investment to turn into growth—Baidu has already found in practice that individual efficiency gains have not yet fully translated into organizational efficiency gains, so it is necessary to change the way tasks are organized, the way knowledge is accumulated, and the way processes operate.

In this sense, others are consuming expectations, while Li Yanhong is the one who issues them—and that is not easy.

Moreover, he has placed the possibility of the next round of growth on something that all enterprises can change and also verify. He seems very happy to let more people know where the road to the other shore can begin—and that is even less easy.

And over these years, he has never had it easy.

Indeed, the intelligent era is racing toward us, but "intelligence" cannot replace everything. Li Yanhong is very firm in his view on this. He believes that under an intelligent organization, people are still the key factor, especially the CEO—the pressure and responsibility borne by this role cannot be replaced by AI.

02 Ten Years of Drinking Ice, Hard to Cool Hot Blood

Liang Qichao called his study "Drinking Ice Studio," a reference to Zhuangzi: "Today I receive my charge in the morning and by evening drink ice water—is it not that I am burning within?" To receive a heavy responsibility in the morning and need ice water by evening, because one's heart is burning with anxiety.

Those who play the prophet probably all understand this feeling.

When he said in 2016 that "AI is the next act," Baidu's situation was not easy. As for the mobile internet wave, Li Yanhong later admitted frankly that they "did not catch the timing." The outside world's tone when talking about BAT quietly changed in those years. But Baidu had long since placed its chips on things whose returns could not be seen in the short term—from 2012 to 2021, cumulative R&D investment over ten years reached 125.2 billion yuan.

125.2 billion yuan, invested bit by bit in those years when clear returns could not be seen. The directions invested in were infrastructure such as chips, cloud computing, autonomous driving, and large models—all asset-heavy, long-cycle matters that require patience.

Criticism followed, and from increasingly sharp angles. In the early years, people asked why Baidu wanted to make robot cars and chips. In recent years the question changed: having started so early, why did it not catch the liveliest gathering?

Li Yanhong rarely argued back about this.

But I think he is using his own way to "argue"—in 2020, for Baidu's twentieth anniversary, a documentary called Twenty Degrees was made. In it, there is a passage where he talks about his love for plants, saying that plants grow slowly but accumulate more.

Once, in a media interview, he was unusually direct: "I want to use technology to change the world. If one year of trying does not work, OK; if two years do not work, OK; if five years do not work, OK; if ten years still do not work, if it is meaningful, I will keep trying. That is the kind of person I am, and that is the kind of company we are. The outside world only needs to understand my products. If you use my product and it creates value for you, I am truly very happy."

With this sentence in mind, looking again at this year's Baidu World, one thing will stand out: there are new things, but there are more iterated "old things."

The new things are important. For example, "Ciyuansuo" made its first appearance—you cannot simply understand it as Baidu having launched yet another trading platform, but rather, against the current background in which "the number of model variants increases by the day," enterprises need a reusable access rule and continuous monitoring mechanism, not just another model catalog.

To put it even more deeply, Ciyuansuo and the intelligent organization are actually different sides of the same thing: when quality and fulfillment on the model side can be measured, the effects of AI have, for the first time, a cross-enterprise comparable scale, and only then does "measurability" at the organizational level have a foundation.

But I must say that more of the highlights were reserved for those "old" products that have been iterated for many years.

There is Kunlun Chip.

Baidu began making AI chips in 2011, when only a handful of companies worldwide were doing this. In 2015, the chip-making "Fairchild" team received Baidu's highest award.

"At the time no one understood. The first ten-plus years were very hard, and many outside did not understand why we wanted to make chips," Li Yanhong recalled in September this year at the highest award ceremony.

Fifteen years have passed. Kunlun Chip has become an independent company and submitted a listing application to the Hong Kong Stock Exchange in January this year. By this conference, the Tianchi super node based on Kunlun Chip had reached 1,024 cards, with the entire stack domestically produced, from core components to the liquid cooling system.

There is Apollo Go.

Baidu has been working on autonomous driving for thirteen years. In July this year, the Hong Kong Transport Department issued it the first local fully driverless testing license—the world's first fully driverless test under a "right-hand drive, left-hand traffic" road system. In Dubai, users can hail its cars by opening Uber.

There is the ERNIE large model.

Baidu's large model started in 2019, the earliest in China's industrial sector, and by 2023 there was ERNIE Bot. A week before the conference, Baidu had just open-sourced ERNIE 5.5 Flash Base.

Other companies keep switching tracks, chasing whatever is lively. Baidu has spent more than ten years guarding "chips, cloud, models, and agents," continuously iterating this system. The overall direction of the products has not changed; what has changed is the core that drives them.

Maintaining strategic focus while also staying at the technological frontier—doing one of these is not hard; doing both together, without letting go for more than ten years, is the hardest thing.

03 The Story of the Eagle

To say that Li Yanhong is the only one in the world turning his gaze to organization would be false—it would be flattery that kills.

In July 2024, Bloomberg disclosed an internal roadmap from OpenAI that divided the path to AGI into five levels, with the highest level called "organization"—AI capable of taking on the work of an entire organization. McKinsey is studying the "agent organization." Microsoft is even more direct, spending $2.5 billion to build a "forward-deployed engineer" team of about 6,000 people, embedded directly in customer companies to redesign processes.

The difference between Li Yanhong and them is this: what Microsoft talks about is what an organization should look like; what McKinsey talks about is how an organization should be designed; what he talks about is how an organization grows itself.

He dared to say this because Baidu first took the knife to itself. And it wasn't a light cut—it truly tore down the old factory building.

Anyone who has done corporate management knows how heavy those six words—"breaking down job boundaries"—really are. A job is not a line drawn on the ground; it is a person's professional identity, their reporting relationships, their evaluation methods, the coordinate by which they know what they should do each morning. If you tell an engineer who has done testing for three years that from today on he is also a product manager, also a designer, also a developer—his first reaction will not be excitement.

Baidu's Digital Human Innovation Business Unit was one of the earliest teams to act. They did something extremely rare in a big company: they tore down all the boundaries among the four roles of product, R&D, design, and testing, and turned everyone into a single role—Builder. In the past, a requirement went from analysis, to proposing a solution, to drawing prototypes, to writing code, and finally to testing—five people in a relay, each handling one segment. Now a single Builder, with the team's self-developed AI tool BiuBiuAgent, goes through the entire process alone.

And this is not a one-off sprint. As the experience from each round of tasks is accumulated and fed back into the AI, the entire team's capabilities keep rolling upward. In the third quarter of this year alone, this team completed 564 product iterations, accumulated 130 reusable Skills and 729 organizational knowledge assets, raised per-capita output by 125%, and shortened the delivery cycle by nearly 30%. This is no longer as simple as "a group of people using AI tools"—tools can be swapped; the way of working, the way knowledge accumulates, and the collaborative relationship between people and AI have all changed.

Similar restructuring is happening simultaneously in several other Baidu teams. Baidu's Yijian team, which does visual intelligence, originally had more than a hundred people and transformed into small end-to-end delivery teams; the first pilot had three people plus an AI Agent and launched an initial version in two weeks, improving efficiency by 60 to 70 percent. The Aether team, which works on autonomous driving data, is even more extreme—two people formed an AI squad and in one month finished work that previously took ten people three months.

There is another story that best explains what "growing itself" means.

Baidu AI Cloud's past model was a three-stage relay—frontline sales understood the industry, the delivery team went to the customer site to implement, and the product and R&D team collected feedback in the rear. Three groups of people, three sets of language, three kinds of KPIs, and information loss in transmission was inevitable. No executive ever planned a new role to solve this problem. But on the front line, a role grew out on its own—FDE, Forward Deployed Engineer.

One person understands the industry, can deliver, and can also drive product iteration; the three-stage relay became a one-person closed loop. It was the frontline people who, with AI's help, discovered that they could do all three things—and so they did. What management did was to name, confirm, and promote this wildly growing species. At present, FDE has delivered more than a hundred scenarios in industries such as finance, energy, and manufacturing.

There was an image at the main forum of the conference that left a deep impression on me. A speaker from Baidu's Dazi team walked onto the stage—half a year ago, this person was still researching quantum computing in a physics lab. He said one sentence: "Many times, the so-called job boundary is not entirely the boundary of ability itself, but the boundary of how knowledge and experience flow." A quantum physicist standing at an AI product launch, talking about how he simultaneously takes on the roles of product, R&D, and project management—this fact itself is the most vivid footnote to "breaking down job boundaries."

And the foundation that strings all this together is dodo, built by ten people. This AI assistant, seven months after launch, covered 99% of Baidu's regular employees and pushed more than three hundred internal platforms to undergo AI-friendly transformation. A company of tens of thousands of people was changed from the ground up by a product made by ten people. This year, this team won Baidu's highest award, one million US dollars.

English text is being prepared.

That an intelligent organization could grow on Baidu also has a foundation-level reason—chips, cloud, models, and agents; it calls this "chip-cloud-model-agent," all four layers its own, and feedback from each layer can flow back to another layer. This itself is a Loop. Solutions pieced together by purchasing third-party capabilities cannot achieve such a closed loop.

But these are all things Li Yanhong said publicly at Baidu World in November. Earlier, he had said something else to his own people.

At the beginning of this year, at a Baidu directors' meeting, he told management a story.

The story came from an article he had read years earlier, "The Rebirth of the Eagle" by TCL's Li Dongsheng. That was 2006, when TCL had fallen into huge losses after international mergers and acquisitions, and Li Dongsheng wrote a fable on the company's internal forum:

When an eagle reaches about forty years of age, its beak grows longer and longer, almost curving into its chest. Its talons and feathers also age day by day. At this point it flies to the mountaintop, fiercely strikes off the aged beak, then plucks out the old feathers and talons one by one. After about 150 days of extremely painful transformation, a new beak grows, new feathers fill in, and it flies into the sky again, winning itself another thirty years of life.

This story accompanied TCL through the most difficult stretch of its transformation. When Li Yanhong retold it, the circumstances were different—Kunlun Chip had just submitted its listing application, and Apollo Go's robotaxis had already driven beyond the border—but it was precisely the most demanding moment. I think what he wanted to tell Baidu's managers was this: self-evolution means actively breaking inertia, passing through discomfort, and embracing reconstruction. Change the beak while you can still fly.

A few months later, "self-evolution" became a keyword at the Create conference. After that came this November's "intelligent organization."

Organizational change at the spiritual level and organizational evolution at the management level met each other, and Baidu was very lucky.

From a fable about an eagle at the directors' meeting, to "self-evolution" before several hundred managers, to the "intelligent organization" aimed at the entire industry—each step stood on the foundation of the previous one. It was not that inspiration struck one day and a new concept was proposed; it was that an idea ran through the organization for more than half a year, and only after it worked did they dare to present it. This is the cognitive efficiency that can only emerge from a high-speed closed loop in practice.

This may be the final layer of the answer to "why is it always him."

Foresight, persistence, and actively changing the beak are not three things, but different sides of the same thing. He sees the direction because he immersed himself in the field earlier and more persistently than others; he holds the direction because he is willing to spend ten years of investment waiting for a judgment to be fulfilled; and he dares to actively smash and start over before it is fulfilled because he knows the next turning point will catch this change.

Source

TMTPost · AGI