Recently, I’ve had quite a few conversations with investors who are active in the primary market. Compared to the secondary market, where stories of AI-driven cost reduction and efficiency gains are still being dug out from financial reports, the primary market has shown a completely different vibe since the second half of last year—somewhat akin to a "crazy break of consensus." The trigger for this wave is clear: AI is starting to move out of screens on a large scale and enter the real physical world. The projects everyone is investing in are all focused on seizing this edge-side opportunity.

How crazy will AI in the physical world get?

I largely agree with the industry consensus that the most competitive robot forms in the future are likely to be three types: humanoid robots, autonomous driving, and drones. These three directions indeed represent pursuits of ultimate efficiency under industrial logic. However, after observing recent developments, I’ve realized that AI’s invasion into the physical world is already at an earlier, more fragmented, and broader stage.

You can feel the difference through two real cases:

The first is AI smart glasses designed for birdwatching enthusiasts. The difficulty lies in the "unpredictability" of the physical environment. In the wild, migratory birds don’t stay still like machine parts. AI must, within milliseconds, extract key morphological information from chaotic tree shadows, sudden changes in lighting, and the rapid flapping of bird wings—dozens of times per second—and then match this data in real-time against thousands of species in the database. This tests not only computational power but also the AI’s ability to accurately capture dynamic targets at the highest level.

The second is an AI robotic arm at a gas station. It faces a completely open and highly risky physical scenario. The system needs to quickly identify the fuel tank cap positions of thousands of vehicle models and then operate precisely. Every step is a tough test of the AI’s environmental adaptability.

These projects may seem niche, but they expose the same core issue: AI must complete tasks in extremely complex and highly uncertain real-world environments. This is far more difficult than running models in the digital world—it's a challenge on an entirely different level.