Summary of Key Points
Baidu uses DAA (Daily Active Effective Agents) as the primary metric for measuring success in the AI field, rather than the commonly used Token call volume. It has established three types of agent matrices: general-purpose, industry-specific, and sectoral, and competes by leveraging a comprehensive “core-cloud-model-agent” full-stack approach. Shen Doud predicts that the AI industry will see five major trends over the next one and a half years: accelerated iteration of agent-driven models, scaled implementation, cognitive upgrades, reduced costs of embodied intelligence, and the importance of full-stack closed-loop capabilities. Baidu's AI revenue has already exceeded half of its total revenue, but it faces competition from rivals like ByteDance. To stay ahead, Baidu promotes the use of AI among its employees through internal initiatives such as benefits and skill development, and uses the resulting data to improve its products.
Why Does Baidu Use DAA Instead of Tokens?
Tokens represent the number of “characters” processed by AI models, similar to electricity consumption; in contrast, DAA represents the actual number of agents that are actively working to solve problems (not just being online). For example, users care about the cooling efficiency of air conditioners, not how much electricity they consume; companies use agents to reduce costs and increase efficiency, not the number of Tokens used. Unlike Daily Active Users (DAUs), which measure the number of users using a service (e.g., an app), an operation and maintenance agent from the Shenzhen Power Supply Bureau may be used by just one employee but can significantly reduce error rates in fault detection—DAA better reflects the value in such scenarios. This approach encourages the industry to focus on specific, smaller use cases rather than developing general-purpose large models.
Baidu’s Agent Matrix: Three Types of Agents for Different Purposes
Baidu categorizes agents into three types to suit various use cases:
1. General-Purpose Agents: Such as Baidu Zha Zi and GenFlow, which handle everyday tasks like information retrieval, writing copywriting, and scheduling.
2. Sector-Specific Agents: Like Miaoda (for office work) and Famo (for optimizing computing power), which are tailored for specific industry scenarios (e.g., data center resource allocation and fault troubleshooting).
3. Industry-Specific Agents: Customized agents developed by partners using Baidu’s infrastructure, such as the fault detection agent used by the Shenzhen Power Supply Bureau.
These agents work together in a coordinated manner: general-purpose agents act as “dispatchers” and call on sector-specific agents when specialized skills are needed, allowing users to complete complex tasks with a single interface.
Baidu’s Full-Stack Closed Loop: What Is Its Competitive Advantage?
Baidu’s full-stack approach consists of four layers: the Kunlun Chip, Intelligent Cloud, Wenxin Large Model, and Agents, enabling it to control the entire process from “growing wheat” to “making bread”:
- Benefit 1: Closed data loop—agents solve real problems, generating data that feeds into model improvements, creating a self-reinforcing cycle.
- Benefit 2: Higher efficiency—Baidu’s own-developed Kunlun Chip can optimize computing power according to the needs of large models, avoiding waste of resources on general-purpose chips.
- Three-Year Window: Chip development takes 2–3 years; companies trying to catch up now will need three years to match Baidu’s lead.
- Critical Factor: The ability to effectively manage and control these agents—currently, only 30–40% of AI systems can complete tasks independently; Baidu aims to increase this figure.
Shen Doud’s Five Industry Predictions for the Next One and a Half Years
1. Iterative Model Development: The focus will shift from comparing model parameters to task completion rates and efficiency improvements, with customers paying for actual value.
2. Scalable Implementation of Agents: By the second half of 2026, there will be numerous AI-powered applications, with 90% of work involving AI (human decision-making and creativity, with AI handling repetitive tasks).
3. Cognitive Upgrades: People will no longer view AI as merely chatbots; human-machine collaboration will become the norm, leading to the era of “super individuals” (a person plus multiple agents).
4. Reduced Costs of Embodied Intelligence: Robot prices will drop to a level affordable for households and small businesses, becoming as common as cloud services.
5. Full-Stack Closed Loop as a Competitive Factor: Companies that focus on only one aspect (e.g., providing computing power) will fall behind; comprehensive full-stack capabilities will be decisive.
How Baidu Builds Internal Strengths
Baidu encourages employees to use AI through various strategies:
- Incentives: Offers a monthly allowance of 1000 yuan to try AI products, lowering the barriers to experimentation.
- Skill Development: Encapsulates repetitive tasks into shareable skills (e.g., creating a skill for checking meeting link permissions).
- Internal Testing: Baidu tests AI products internally before releasing them to the market.
- Building Trust: Focuses on improving agents’ autonomy—ensuring they can complete tasks efficiently without human intervention, making their use more appealing.
Conclusion
Baidu’s core AI strategy is to measure value using DAA and accelerate iteration through a full-stack closed loop. In the next one and a half years, the AI industry will shift from focusing on technical parameters to the effectiveness of practical applications. Comprehensive capabilities and scaled use of agents will be crucial. Despite competition, Baidu’s early adoption and internal efforts give it a competitive edge in this field.