Is AI Entering a Self‑Evolving Era? Baidu’s Robin Li Introduces the Daily Active Agents (DAA) Metric

1. Goodbye Token and DAU: DAA Becomes the AI‑Era Metric

In the mobile‑internet era, DAU measured traffic, while early large‑model AI used Token to gauge compute cost; both miss AI’s core value. Li states that Token measures input cost and DAU measures user visits, but the true standard is how many agents continuously execute valuable tasks and deliver tangible outcomes.

DAA (Daily Active Agents) counts "what AI does for humans" rather than who uses AI.

It measures task‑completion efficiency instead of raw compute consumption.

It focuses on value‑closed loops rather than traffic scale.

Market evidence shows Anthropic’s enterprise‑value ARR surpassing OpenAI in 2026, illustrating DAA logic: higher DAU does not beat focused agent‑task delivery.

Li predicts global DAA will exceed 10 billion, representing 10 billion digital tasks efficiently executed, not conversations.

2. AI Evolution Theory: Three‑Layer Self‑Evolution

2.1 Agent Self‑Evolution: From Reactive to Autonomous

Traditional AI follows a "command‑response" model. Evolved agents gain four capabilities—environment perception, self‑learning, path planning, and error correction—transforming from tools into digital employees.

Ba​idu’s DuMate and the decision‑making agent FaMo 2.0 exemplify this: DuMate handles cross‑application workflows autonomously; FaMo 2.0 improves port automation efficiency by 10.21 %.

2.2 Individual Self‑Evolution: Everyone Becomes a Super‑Individual

AI lowers development barriers so anyone can create a personal agent fleet via natural language, turning ideas into deployable applications.

At the conference, an 8‑year‑old built a campus‑umbrella‑sharing app with MiaoDa 3.0, showcasing rapid agent‑driven development for both personal and enterprise scenarios.

2.3 Organization Self‑Evolution: Human‑Agent Hybrid Teams

Traditional organizations rely on hierarchical human‑to‑human division of labor. In the AI era, enterprises become "human‑agent hybrid" super‑organizations where agents handle repetitive, high‑throughput tasks and humans focus on innovation and decision‑making.

Li emphasizes that CEOs must define a "agent‑first strategy" and adopt four organizational principles:

More delegation, less control – unleash agent and talent creativity.

Fast alignment, fewer layers – direct information flow and rapid decisions.

High‑density talent, fewer headcounts – combine elite human talent with strong AI capabilities.

Task‑centric, less siloed – convey intent via prompts to maximize AI’s full‑spectrum value.

3. Chip‑Cloud‑Model‑Agent Full‑Stack Evolution: Baidu’s AI‑Native Infrastructure

Ba​idu’s "Chip‑Cloud‑Model‑Agent" architecture underpins the AI‑era evolution, integrating hardware, cloud, models, and agents.

Chip: Kunlun P800 supports hundred‑thousand‑agent concurrency; a 10 k‑card cluster and Ti​anChi 256 super‑node (launch June) break single‑point limits.

Cloud: Baidu Intelligent Cloud transforms into an Agent‑native stack, offering memory, task orchestration, security isolation, and compute scheduling tailored to agents.

Model: Ernie 5.1 pre‑training cost is only 6 % of comparable models, leading domestic LMArena rankings and empowering agents with complex instruction understanding, multi‑step reasoning, and self‑correction.

Agent: A matrix of agents—from generalist DuMate and code agent MiaoDa 3.0 to enterprise decision agent FaMo 2.0 and full‑scene digital‑human platform—covers personal, corporate, and industrial use cases, turning DAA from concept to concrete value.

Conclusion: AI’s Endgame Is Value Evolution, Not Technical Arms Race

The AI industry lacks lasting logic, not concepts. Li’s DAA‑centered evolution redirects focus from parameter bloat and traffic anxiety to agent deployment, individual empowerment, and organizational upgrade. When 10 billion daily active agents become reality, AI will shift from a technical notion to a productivity infrastructure reshaping individuals and enterprises alike.