Summary of Key Points
After its listing, Yushu Technology's market value evaporated by 200 billion yuan. However, what's more crucial is its partnership with the AI giant DeepSeek. Yushu aims to use DeepSeek's technology to address the challenge of robot "self-evolution." Nevertheless, there are significant differences in their approaches to achieving Artificial General Intelligence (AGI), posing the risk that once partners become rivals (similar to the split between Figure AI and OpenAI).
I. Yushu's Own AI Models Are Insufficient, and They Urgently Need External Help
Over the past few years, Yushu has developed two robot AI models, both of which have significant shortcomings:
1. VLA Model (Vision-Language-Action): Relying on open-source Qwen2.5 and training data, this model has limited generalization capabilities. The robot can only perform actions based on what it sees, without considering the consequences. It requires retraining for different scenarios. Many robot companies have abandoned this model.
2. World Model (WVLA2.0): Due to insufficient computing power, this model struggles to handle complex tasks. Although it's more suitable for robots, Yushu's WVLA2.0, running on a low-power chip (100 TOPS), can only perform 10 inferences per second, making it incapable of handling multi-step tasks like "picking up a file, signing it, and putting it back in a drawer."
Therefore, Yushu needs to collaborate with a top AI company to overcome these limitations, and DeepSeek seems to be the ideal partner.
II. How DeepSeek Can Help Yushu Overcome These Challenges
DeepSeek can provide several benefits:
1. Enhanced Inference Abilities: By integrating DeepSeek's technology, Yushu's robots will be able to handle complex tasks, such as organizing a desk and throwing away trash. This will expand their functionality beyond simple actions.
2. Support for Physical AI Self-Evolution: Yushu hopes to enable AI to undertake part of the robot development process—searching for papers, writing code, and testing in simulation environments, followed by controlling the physical robot for optimization. DeepSeek's advanced models can help establish a self-evolutionary loop.
III. Divergent Views on the AGI Path: What Are the Differences Between Wang Xingxing and Liang Wenfeng?
AGI refers to the ability of robots to perform any task without specific training. However, Wang Xingxing (from Yushu) and Liang Wenfeng (from DeepSeek) have vastly different perspectives on how to achieve this:
- Wang Xingxing: He believes that a world model is at the core of AGI, as it enables robots to understand the physical world. He argues that aligning video-generated actions with robot movements can solve data-related issues and lead to true AGI, suggesting that language models may not be essential.
- Liang Wenfeng: He emphasizes the importance of continuous learning, where the AI learns to find data, fix bugs, and iterate on its own. Only once the AI can evolve independently can it develop embodied intelligence. In his view, the world model is a secondary aspect of AGI, not its core.
These differences in approach could create potential issues for their collaboration.
IV. Concerns About the Potential for a Similar Outcome as with Figure AI and OpenAI
The partnership between Figure AI and OpenAI ended poorly when OpenAI decided to develop its own robot team, turning against Figure. Could the same happen with Yushu and DeepSeek?
- Short-Term Security: DeepSeek has made a strategic investment in Yushu (141 million yuan, with a three-year lock-up period) and has pledged to prioritize purchasing Yushu's products for embodied intelligence research. For now, their interests align.
- Long-Term Risks: If DeepSeek develops its own robots and sees them as a necessary step towards AGI, it could become a competitor to Yushu, leveraging its AI technology and hardware to compete directly in the market.
In Conclusion
While the partnership between Yushu and DeepSeek is currently mutually beneficial, the underlying differences in their approaches and potential future competition make it a partnership with an uncertain outcome. Whether they can succeed depends on their ability to balance short-term interests with long-term strategic goals.