Live from WAIC: Radish Run Enters the 'National Team', AI Digital Human Tech Upgrades, Baidu's Full-Stack Self-Research Goes Crazy}

Recently, Ultraman revealed in an interview that after experiencing GPT-5, he was shocked by its powerful capabilities. It could answer questions he couldn't even understand, making him feel somewhat powerless even in his expertise.

Despite rapid AI progress, some still question its practical deployment, suspecting many so-called new technologies are just hype.

However, at the just-opened WAIC, we see many real AI applications in practice.

Autonomous driving is a hot topic, with many unmanned vehicles busy both inside and outside the conference, serving as official shuttles.

Inside the exhibition, digital human anchors interact naturally with visitors, demonstrating top-tier broadcasting skills with quick responses and rich knowledge.

This scene feels like stepping into the future. What's more impressive is that all these useful AI technologies are developed entirely in-house by one company.

China’s Popular RoboTaxi Goes Global, Secures Two Cities in Half a Year

At WAIC, Baidu's Radish Run was selected for the "China AI Industry Innovation Achievements Exhibition" as a representative of the "national team." Additionally, Baidu and other tech giants received autonomous vehicle operation licenses in Shanghai, officially launching RoboTaxi services in Pudong.

Unmanned taxis have become commonplace on Chinese streets. Online discussions praise their convenience and safety.

In just half a year, Radish Run's testing scope expanded from airports to core urban areas like Tung Chung, with increasingly complex traffic scenarios.

Radish Run has begun testing in Hong Kong.

On July 15, Radish Run announced a strategic partnership with Uber to expand autonomous ride-hailing services globally beyond the US and China, deploying over 1,000 vehicles in Dubai and Abu Dhabi.

With mature technology and supportive policies, Baidu's autonomous driving business is further validated, indicating China's rapid ability to replicate and implement autonomous vehicle solutions.

Soon, our travel methods may be revolutionized by RoboTaxi.

AI Luo Luo Boosts Sales to 55 Million, Baidu Launches New Moves

Besides Radish Run, Baidu is also heavily investing in digital humans.

During the recent 618 shopping festival, Baidu's Luo Yonghao digital human went viral on social media.

In the live broadcast, AI Luo Luo and AI Zhu Xiao Mu joked and interacted for over six hours, attracting over 13 million viewers and breaking 55 million yuan GMV, setting a new record for digital human live commerce.

Remarkably, Baidu's digital human technology was upgraded just a month later.

This time at WAIC, Baidu announced the new generation digital human tech NOVA, expected to launch in October, bringing mass production of top-tier live hosts. This means ordinary users can also become professional live streamers.

NOVA's key feature is the upgraded script mode.

Previously, digital hosts could only read scripts with stiff expressions and movements. Now, NOVA leverages Wenxin 4.5 Turbo's advanced script mode to dynamically adjust tone, expressions, and actions based on product info, knowledge, and character settings.

Recently, I saw a video of Feng Tang, and the movements looked repetitive—nodding, waving, a bit stiff. Honestly, the main reason digital humans look fake is because of stiff movements.

NOVA's motion generation has improved significantly, capable of producing complex, contextually appropriate actions, even enabling two digital humans to interact naturally during live streams, with seamless interruptions.

Its voice cloning technology also reached new heights, perfectly replicating the host's voice and tone, even capturing catchphrases accurately. For example, Luo Yonghao's digital version can spontaneously say his classic phrases like "Did you understand?" or "Got it?" with authentic tone and style.

In dual-host scenarios, the two digital humans can coordinate smoothly, respond to each other, and even dance or raise their voices when excited, making it nearly indistinguishable from real hosts.

All this high-fidelity replication only requires 10 minutes of real human samples. Upload 10 minutes of your video, and NOVA can generate a hyper-realistic digital human matching your voice, actions, and atmosphere.

The second major breakthrough is the upgrade of the AI brain, enabling the digital human to reach top-tier hosting levels—receiving real-time data, making autonomous decisions, managing multi-agent interactions, and generating stylized content based on trending topics.

During Q&A, the AI brain responds quickly and accurately, guiding audience interaction, asking questions, and influencing purchase intent through multi-turn conversations.

It also coordinates multiple intelligent agents—such as hosts, operators, and scene controllers—to execute interactive strategies like lucky draws, price adjustments, and inventory management, boosting sales conversions.

For viewers and streamers, digital human tech offers truly practical experiences.

Behind Baidu's Full-Stack Self-Research: Building "Millions of Super Apps"

Whether accelerating overseas expansion with intelligent driving or creating viral digital humans, Baidu's AI applications are supported by a comprehensive self-developed tech ecosystem. Notably, at WAIC, Baidu's AI computing cluster and PaddlePaddle deep learning platform, along with Radish Run, were showcased at the "China AI Industry Innovation Achievements Exhibition," affirming Baidu's full-stack AI capabilities.

From the start of AI's rise, Baidu recognized that the real value lies in application, not just technology.

Based on this strategic mindset, Baidu has built a vast AI application matrix.

Beyond Radish Run and Nova, Baidu has implemented AI across search, cloud storage, and smart office products.

For example, the newly launched "Seconds Da" platform in March uses a "one-sentence application + multi-agent collaboration + multi-tool invocation" approach, enabling users to generate complete code via natural language, greatly improving development efficiency.

Meanwhile, upgrades to Baidu Wenku and Baidu Cloud enhance content management, knowledge creation, and sharing, deepening AI integration into traditional products.

Baidu's ability to push AI applications forward relies on its robust full-stack AI ecosystem, from computing power, frameworks, to models, creating a comprehensive AI environment.

The core of Baidu's AI full-stack architecture is the Kunlun P800 chip and the Baidu AI heterogeneous computing platform, which support large-scale training of billion-parameter models and fine-tuning for thousands of clients, reducing costs and ensuring domestic chip supply.

In the framework layer, PaddlePaddle, China’s first self-developed deep learning platform, supports large-scale training and inference, integrating deeply with Baidu's other tech components to improve efficiency. It supports multiple hardware platforms and offers user-friendly interfaces, lowering AI application barriers.

Additionally, PaddlePaddle is the only domestic deep learning framework capable of competing with global giants like TensorFlow and PyTorch, representing a fully controllable AI operating system in China.

At the model level, Baidu has built a multi-layered model system centered on Wenxin models, continuously pushing generative AI into industry applications. Wenxin models feature multimodal, strong reasoning, and low-cost capabilities, widely used in search, recommendation, and intelligent assistants, empowering manufacturing, finance, energy, and urban governance sectors.

In April 2025, Baidu released Wenxin 4.5 Turbo and X1 Turbo, with significantly faster inference and 80-50% lower costs, achieving SOTA in multiple evaluations. Recently, Baidu open-sourced the Wenxin 4.5 series, covering 47B and 3B parameter models, including MoE and dense versions, making it one of the few fully open-source general large models in China. These advances demonstrate Baidu's leading position in model technology and support AI's widespread industry adoption.

Unlike traditional "tech show-offs," Baidu's full-stack AI architecture focuses on practical applications, integrating technology to serve real needs and expanding AI into more markets. As CEO Robin Li states, "We aim not just to launch a super app but to create millions of super-useful applications."

More evidence shows Baidu's "AI super usefulness" approach is correct and expanding. With continuous evolution of its full-stack architecture, Baidu's goal is to make AI a common part of daily life, driving innovation in real-world scenarios.