Today's Create 2026 Baidu AI Developer Conference delivered a substantial amount of information. At the conference, Baidu founder and CEO Robin Li proposed a definitive metric for gauging the prosperity of a platform or ecosystem in the AI era: DAA (Daily Active Agents). Simply put, DAA measures "how many agents are genuinely helping humans with tasks and delivering outcomes each day."
In the mobile internet era, we looked at DAU (Daily Active Users), with global leader Meta having approximately 3.4 billion. However, in the agent era, as every individual and enterprise possesses their own "digital employees," the global DAA could easily surpass 10 billion. Based on this new metric, Li directly pointed out a prevailing misconception in the industry: an excessive focus on token consumption. He noted that the industry often boasts about being "token billionaires," but in his view, tokens merely represent computational costs. They measure input, not output, and do not signify the final outcome. Centered on this shift in fundamental logic, Li proposed the "self-evolution" theory for the AI era. This evolution encompasses three layers: first, the evolution of the agents themselves. They are no longer passive, waiting for queries, but can proactively learn from their environment and execute tasks. Second, the evolution of the individual human. An individual equipped with an agent team can become a "super individual." Finally, the evolution of corporate organizations. Future companies will no longer rely solely on human collaboration but will become "hybrid teams" of humans and agents, evolving into super organizations. Beyond these core concepts, Baidu also shared several distinctive insights regarding corporate strategy and macro trends: First, the leading role in AI has officially shifted from "competing on models" to "competing on applications (agents)." Users are not concerned with model parameter sizes; they care only about whether the technology "can help me get things done." Second, the software industry will be reshaped into a "fast-moving consumer good." As the cost of AI-generated code approaches zero, there will be an emergence of numerous "disposable software" solutions tailored for specific, minor needs—used once and discarded. This trend is expected to expand the overall software market by tenfold. Third, workplace management theory is being completely overturned. Traditional theory suggests a manager can effectively oversee up to 10 people. However, within AI hybrid teams, the span of management can flatten to 30-50 people. The core function of a manager will shift from "supervision" to "alignment"—ensuring everyone is focused on the right tasks. Next, let's examine the most concrete aspects of today's conference regarding product and industry implementation:
One Person as a Full Team: General-Purpose Agents Begin Handling Complex Operations. Baidu's newly launched general-purpose agent, "DuMate (Baidu Partner)," can directly "see" and understand screens to operate software. A live demonstration showcased a realistic startup scenario: with a single instruction, the agent simultaneously acted as customer service, sales, and marketing. DuMate automatically logged into an email account to handle a customer complaint (even considering emotional安抚 and offering free shipping compensation), cleaned up a messy raw sales spreadsheet to calculate inventory forecasts, and concurrently generated promotional posters and图文 content, embedding a flash sale QR code within the article. Development Barriers and Costs Approach Zero: The Era of "Disposable Software" Arrives. Previously, developing an app required maintaining a team. However, the "Miaoda App" was developed by a team of just one person, as 90% of its code was written by the Miaoda agent itself. An 8-year-old primary school student from Wenzhou demonstrated a school-wide umbrella-sharing mini-program he created effortlessly using simple instructions. Analyzing this, Li stated: when code becomes inexpensive, software turns into a "fast-moving consumer good." In the future, one could instantly generate software for a minor, one-time need and discard it after use. This will not harm the software industry but will instead expand the market scale tenfold. Short Video Generation Breaks the "Tens of Seconds" Barrier: Digital Humans Can Produce Hour-Long Productions. Baidu upgraded its former Huiboxing to the full-scenario digital human platform "Baidu Yijing." Now, users can employ "Baidu Yijing" to create interactive content lasting several minutes or even hours. A live demonstration featured an AI Pamela promoting protein bars, with incredibly realistic details such as the crunch of grains, skin texture, and eye sparkle. For cross-border e-commerce or brand marketing, a single person paired with an AI can now accomplish video production volumes that previously required an entire professional team. Decision-Making Agents Dive into Deep-Water Industries: Directly Calculating "Optimal Solutions" for Experts. The previously launched "Famou" strategy tool has been upgraded to version 2.0. The key change is that it is no longer an exclusive tool for algorithm engineers but is now directly accessible to frontline business experts. For instance, in manufacturing production scheduling or chemical formula optimization, business personnel do not need coding knowledge; they simply input data, and the agent can repeatedly simulate scenarios within actual production environments. After implementing it in Qingdao Port's A-TOS system, absolute production metrics were boosted by 10.21% on top of already high efficiency, with an estimated capacity for nearly one million additional TEUs annually. Advanced Intelligent Driving Passes the Tipping Point for Adoption; "Cabin-Driving Integration" Becomes Reality. Horizon Robotics' CEO Yu Kai shared an intriguing statistic: for vehicle models equipped with their HSD (end-to-end autonomous driving) system—offered as a premium option—the consumer take-up rate reached 77%, indicating willingness to pay for intelligent driving. Their newly launched "Starry Sky" chip further integrates the intelligent cockpit and autonomous driving functionalities. Supporting this significant industry evolution is a massive computing cluster, co-built by Baidu Intelligent Cloud and Horizon Robotics, comprising over 5,000 nodes capable of supporting ten-thousand-card level training. Real-Time Robot Control Becomes Feasible; Embodied AI Faces Commercial Tests in the Second Half of the Year. Xinghai Tu released the world's fastest world model, Fast-WAM, drastically reducing single-step latency to under 190 milliseconds. This breakthrough means robots can finally react swiftly like humans. They are currently collaborating with Baidu Intelligent Cloud to push latency below 90 milliseconds. According to their forecast, the second half of this year will be a crucial period for validating whether embodied AI can effectively replace human labor in industrial settings. The Fundamental Logic for Cloud Providers is Changing; Domestic Computing Infrastructure is Ready for Demanding Tasks. Baidu Intelligent Cloud's Shen Dou stated that in the next phase for cloud services, competition is not about who consumes more tokens, but who achieves a higher "token conversion rate." Their upgraded Agent Infra (Agent Infrastructure), through optimizations in long-context management, helps clients save over 23% on tokens in typical office scenarios compared to OpenClaw. Notably, over half of China Merchants Bank's AI application scenarios now run on Baidu's domestic Kunlunxin P800 chips. Furthermore, a significant model within the Wenxin 5.1 series has been trained entirely on a domestic cluster, achieving an effective training rate exceeding 97%. To some extent, this signifies that domestic computing power has become a reliable mainstay capable of supporting core business operations.
Comments