Robin Li sets the tone for the second half of AI: The focus shifts from "large models" to "applications," DAA will become the new benchmark, and "disposable software" reshapes productivity.

```

Today I attended the Create2026 Baidu AI Developer Conference, and the amount of information was overwhelming.

At the conference, Baidu founder and CEO Robin Li gave a definitive metric for measuring whether a platform or ecosystem is truly prosperous in the AI era — DAA (Daily Active Agents).

Simply put, DAA measures "how many intelligent agents are actually working for humans and delivering results every day."

In the era of mobile internet, we look at DAU (Daily Active Users), with Meta topping out globally at about 3.4 billion. But in the era of intelligent agents, as every person and business has their own "digital employee," global DAA can easily exceed 10 billion.

Based on this new metric, Robin Li directly pointed out a common misconception in the industry: focusing on Token consumption.

He said that nowadays, people often brag about being "Token billionaires," but in his view, Tokens only represent computational costs. They measure input, not output, and they don't reflect the endgame.

Around this fundamental logic shift, Robin Li introduced the concept of "self-evolution" in the AI era.

This evolution has three layers: first, the evolution of the intelligent agent itself, with agents no longer passively waiting for questions, but able to proactively absorb from the environment and execute tasks; second, the evolution of individuals, where a person with a squad of agents can become a "super individual"; lastly, the evolution of business organizations, with future companies forming "hybrid squads" of humans and agents, becoming super organizations.

Aside from these two core concepts, Baidu also has some unique judgments about company strategy and macro trends:

First, the main player in AI shifts from "competing on models" to "competing on applications (agents)". Users don't care how big your model parameters are, just whether you "can get the job done."

Second, the software industry will be reshaped into "fast-moving goods". Since AI makes coding costs nearly zero, there will be a surge of customized "disposable software" for small needs — use it once and discard it — expanding the software market tenfold.

Third, workplace management theory will be transformed. Traditional theory says a manager can directly oversee at most ten people, but with AI hybrid squads, management span will flatten to 30–50 people. The manager’s core function will shift from "supervision" to "alignment" — making sure everyone is doing the right thing.

Next, let's look at the most hardcore points from today's conference in terms of product and industry implementation:

One person becomes a whole team, general-purpose agents start to take over complex tasks. Baidu’s general-purpose agent "DuMate (Baidu Partner)" can directly "see" the screen and operate software. The live demonstration showed an authentic startup scene: one sentence assigns tasks, and DuMate morphs into customer service, sales, and marketing. DuMate automatically logs into email to handle complaints (even soothing emotions and offering free shipping compensation), cleans up chaotic sales sheets to forecast inventory, then generates new product posters and articles, embedding flash sale QR codes within.
Development thresholds and costs drop to zero, "disposable software" era arrives. Previously, making an app meant maintaining a team, but "MiaoDa App" has just one developer, as 90% of its code was written by the MiaoDa agent itself. At the conference, an 8-year-old student from Wenzhou showed how he easily built a school umbrella-sharing app with just a few plain sentences. Robin Li analyzed: when code is worth nothing, software becomes "fast-moving goods." You can generate an app for any minor, one-time need, and discard it after use. This won’t kill the software industry, it will instead expand the market tenfold.
Short video generation breaks the "10-second" curse, digital humans can make hour-long blockbusters. Baidu upgraded Huiboxing to the new all-scenario digital human platform "Baidu Yijing." Now, you can use "Baidu Yijing" to create highly interactive content spanning minutes or hours. The demo showed AI Pamela promoting protein bars—grain bite details, skin texture, and eye glimmer were incredibly lifelike. In the future, cross-border e-commerce or brand marketing can be done by one person plus an AI, matching the output of a whole professional video team.
Decision agents dive deep into industry, directly helping experts find the "optimal solution." The previously launched "Famou" has been upgraded to 2.0, with its biggest change being it's not just for algorithm engineers but business experts on the frontlines. For manufacturing scheduling and chemical formula optimization, users just feed it data — no coding needed — and it iterates in real production environments. Qingdao Port’s A-TOS system used it to boost already high efficiency by another 10.21%, expected to handle nearly a million more containers per year.
Advanced smart driving crosses the adoption inflection point, "cockpit-driving integration" becomes reality. Horizon CEO Kai Yu revealed an interesting data point: vehicles equipped with their HSD (end-to-end autonomous driving) — despite being a high-priced upgrade — have a 77% selection rate; consumers are ready to pay for smart driving. Their new "Starry Sky" chip merges smart cockpit with autonomous driving. Supporting this massive industry shift is a super large compute cluster co-built by Baidu Cloud and Horizon, with over 5,000 nodes and support for tens of thousands of cards in training.
Real-time robot control becomes possible, embodied intelligence faces a commercial test in the second half of the year. Xinghaitu released the world’s fastest world model Fast-WAM, pushing single-step delay to under 190ms. This means robots can finally react as quickly as humans. They are working with Baidu Cloud to push delay below 90ms. According to their projection, the second half of this year will be a verification period for whether embodied intelligence can truly replace manual labor in factories.
The underlying logic for cloud vendors has changed; domestic compute power base is ready for tough battles. Baidu Cloud’s Shentao said the second half of cloud computing isn’t about who consumes more tokens, but whose "token conversion rate" is higher. Their upgraded Agent Infra (agent infrastructure) uses optimized long-context management to save clients over 23% token use in typical office scenarios compared to OpenClaw. Of note, over half of China Merchants Bank’s AI use cases are now running on Baidu’s domestic Kunlun P800 chips; and a key Wenxin 5.1 model has finished training on a fully domestic cluster, with over 97% effective training rate. In some sense, this means domestic compute power is truly ready to shoulder core business tasks.

This article is from WeChat public account "Hard AI". For more AI news, visit here.

Risk Warning and DisclaimerThe market has risks, investment requires caution. This article does not constitute personal investment advice, nor does it account for the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions herein are suitable for their particular circumstances. Invest accordingly, at your own risk. ```