Many Futures of AI: Reflections on WAIC
Compared with the concentrated enterprise focus of the AWS Summit, the World Artificial Intelligence Conference (WAIC) in Shanghai felt more like a panorama of competing technological futures. There was no single theme that dominated the entire event. Instead, companies of dramatically different sizes—from small start-ups to major Chinese and international corporations—presented their own interpretations of where AI is heading.
The exhibits were equally diverse. Some companies emphasized foundation models and their expanding capabilities, while others showcased consumer products and creative applications. Devices such as AI-powered glasses suggested new forms of human–computer interaction, while systems for generating short dramas demonstrated how AI may reshape media production. Walking through the exhibition, I had the impression not of one clearly defined trajectory, but of many experiments unfolding simultaneously. The field seemed to be searching for the interfaces, products, and business models through which increasingly general computational capabilities might enter everyday life.
The robotics floor was particularly thought-provoking. It was visually impressive: humanoid robots danced, robotic arms performed industrial operations, and machines interacted with enthusiastic crowds. Yet the demonstrations also revealed a gap between the public image of embodied intelligence and its current reality. Some performances appeared pre-programmed or remotely controlled, while many industrial demonstrations involved structured tasks such as repetitive grasping on production lines. These are useful achievements, but they remain far from the adaptive, general-purpose embodied agents often imagined in discussions of the future.
This gap is understandable. Intelligence in the physical world is fundamentally difficult. A language model operates through digital representations; a robot must deal with friction, uncertainty, imperfect perception, hardware limitations, and potentially irreversible mistakes. A small error in a chatbot may produce an awkward sentence. A small error in a robot may damage an object—or harm a person. Embodied AI therefore requires not only better models, but also reliable perception, control, safety mechanisms, and the ability to learn under real-world constraints.
WAIC left me both excited and cautious. The diversity of ideas showed that AI innovation is flourishing, and that no single company has defined its final form. At the same time, spectacular demonstrations can easily make technological progress appear more mature than it is.
Perhaps the most important lesson was that we should evaluate AI not only by what it can display under controlled conditions, but by how robustly it can understand, adapt, and act when the world refuses to follow a script.