美国AI三巨头重构制药产业：从卖工具到要分成，它们想拿走更大利益

Traditional AI assistants adhere to 'answering every question', but two new studies from Tsinghua University attempt to reverse this paradigm. According to reports from Machine Heart, researchers aim to teach AI to 'read people' and grasp the timing of interaction—anticipating needs before users speak, or judging whether to intervene during silence. This is not merely response optimization, but a redistribution of interaction initiative.

The key of the research lies in understanding user state and context. For example, when a user frequently checks the time, the AI might proactively remind them of the schedule; when a user is deep in thought, the AI chooses to remain silent. This capability requires the model to integrate multimodal cues (gaze, actions, historical behavior) and reason in real time. Currently, this work is still in the lab stage, but it points to a clear direction: more important than giving the right answer is 'when to answer' and 'whether to answer'.

Another path parallel to virtual assistants is robot intelligence in the physical world. Leju Robot, a newly listed company on Hong Kong Stock Exchange, recently announced its bet on 'physical AI infrastructure'. The company's core assertion is 'perception is power, data is a barrier'. Reports from QuantumBit point out that its strategy focuses on enabling robots to continuously acquire data from the environment, rather than just executing preset instructions.

The core challenge of physical AI lies in the unstructured and dynamically changing nature of the real world. Robots need to understand their own position, object properties, human activities, etc., in real time, and adjust their behavior accordingly. Leju's investment direction is to build a high-density perception network and data loop, enabling robots to learn from every interaction. This strategy regards data accumulation—rather than algorithmic breakthroughs—as the long-term moat.

On the surface, Tsinghua's 'timely interaction' research and Leju's 'physical AI' belong to different fields, but the underlying logic is the same: AI must shift from passive execution to active understanding. Tsinghua's research emphasizes interaction timing, while Leju's investment emphasizes environmental perception. Together, they point to deep modeling of 'context'. Without context, AI cannot judge when to act; without perception, robots cannot understand context.

This shift will redefine the value of AI. Most current systems rely on users to initiate requests, but future systems may act before users—for example, smart homes proactively adjusting the environment based on user habits, or robotic warehouses predicting and moving goods before orders are placed. This capability places extremely high demands on data quality and real-time performance, making 'perception' and 'timing' scarcer resources than 'algorithms'.