He Xiaopeng, Chairman of XPeng Motors, posted on his WeChat Moments about topics such as robots and the AI bubble. He pointed out that humanoid robots will become a competition among giants in the future, while different specialized robots will have numerous players from various fields, offering many opportunities for success. He believes that there is currently no AI bubble, and the future AI market holds tremendous opportunities.
Below is the full text:
During this trip to the US, I chatted with nearly thirty friends, both external and internal, in the AI field. Here are some impressions:
About New Startups and Robotics in the US
It seems that startups in the US are abundant in AI, biotechnology, and finance. Regarding AI startups, looking from Silicon Valley, there are particularly many in the SaaS and physical AI robotics sectors, with very high valuations. Half of the people I spoke with this time were friends involved in robotics startups, showing a high concentration. Many Chinese robotics companies start with joints and control systems, while many US robotics companies start with models. I believe that humanoid robots will become a competition among giants in the future, while different specialized robots will have numerous players from various fields, offering many opportunities for success.
About Physical AI
Wittgenstein once said, "Language is the world." Language is a summary of knowledge and a means of information transmission, but to some extent, it also constrains our imagination of the world. Many things are difficult to imagine and learn through language alone, such as how children learn to walk or swim, which primarily rely on observing, imitating, and reinforcing through interaction with the world. Therefore, everyone expects to move from large language models to multimodal and even world models. I believe that in the next three years, the most likely area for major breakthroughs may not be in the digital world (it is clearly visible that although OpenAI's core focus is on AGI, it is fully committed to business implementation. Of course, there is anticipation for the next Transformer or a significant decrease in training costs). Instead, major changes are more likely to occur in the field of physical AI. For example, autonomous driving will directly reach near-L4 or full L4; humanoid robots will achieve a rapid leap from L1, similar to autonomous driving, to early-stage L4. These developments will produce significant breakthroughs. The development speed of physical AI will be slower than that of digital AI, but its impact on changing our lives will be greater.
About the AI Bubble
I believe that every technological era has bubbles at certain stages or in specific areas, whether it was the internet era or the new energy vehicle era. However, this is an inevitable competitive process for the market to move from chaos to order. Overall, I believe AI will inevitably drive huge changes throughout society; we are still at the very beginning stage, from 0 to 0.1. If we must talk about a bubble, I feel that valuations in China are relatively reasonable, while valuations in the US are somewhat high; China focuses more on market applications, while the US focuses more on frontier research.
Overall, I believe there is currently no AI bubble, and the future AI market holds tremendous opportunities.
About the Arrival of AGI
I think today's AI is mainly based on human-like imitation learning (such as learning knowledge from textbooks, learning how others perform tasks) plus reinforcement learning (such as repeatedly practicing problems, continuously improving the effectiveness and efficiency of a task). Similar to autonomous driving, because it can quickly learn from the driving habits of millions of people, we see that it can soon drive better and safer than most people, and exhibit entirely new emergent capabilities. However, these are not true creativity. True AGI still requires many capabilities, such as moving from multimodal to world models (similar to our understanding of the 'why' behind things, or an upgraded version of the self-play learning seen in autonomous driving world models), continuous learning, long-term insight and planning capabilities, etc., which have not yet fully arrived. This may still take several years and await further improvements in some underlying capabilities.