The Manus model has made breakthrough progress, sparking controversy over the path of AI development.

Recently, the Manus model has achieved remarkable results in the GAIA benchmark test, surpassing other large language models of the same tier. Manus has demonstrated the ability to independently complete complex tasks, such as handling multinational business negotiations, which involve contract clause analysis, strategy formulation, and proposal generation among various stages. Compared to traditional systems, Manus excels in its dynamic goal decomposition, cross-modal reasoning, and memory-enhanced learning capabilities. It can break down complex tasks into hundreds of executable subtasks while processing multiple types of data and continuously improving decision-making efficiency and reducing error rates through reinforcement learning.

The success of Manus has sparked discussions within the industry about the development path of AI: will the future move towards a unified model of Artificial General Intelligence (AGI), or a collaborative model of Multi-Agent Systems (MAS)? This controversy stems from the design philosophy of Manus, which suggests two possible development directions:

1. AGI Path: By continuously enhancing the capabilities of a single intelligent system, it gradually approaches the comprehensive decision-making abilities of humans.

2. MAS Path: Use Manus as a super coordinator to direct thousands of specialized intelligent agents to work together.

This discussion actually reflects a core contradiction in the development of AI: how to balance efficiency and safety. As individual intelligent systems get closer to AGI, the risks associated with the opacity of their decision-making processes also increase. While multi-agent collaboration can disperse risks, it may miss critical decision-making opportunities due to communication delays.

The progress of Manus also highlights potential risks in AI development, such as data privacy breaches, algorithmic bias, and adversarial attacks. For example, in medical scenarios, the system needs access to sensitive genomic data of patients; in financial negotiations, it may involve undisclosed corporate financial information. Additionally, during the recruitment process, the system may exhibit bias against specific groups; in legal contract reviews, there may be a high rate of misjudgment regarding terms in emerging industries. More seriously, hackers could disrupt the system's judgment in negotiations by implanting specific audio signals.

These challenges highlight a worrying fact: the smarter AI systems become, the broader their potential attack surface.

To address these security challenges, the industry has proposed various solutions, among which Fully Homomorphic Encryption (FHE) technology is regarded as a promising approach. FHE allows computations to be performed on data while it is encrypted, which is crucial for protecting sensitive information in AI systems.

Specifically, FHE can enhance the security of AI systems in the following aspects:

1. Data layer: All information input by users (including biometric features, voice, etc.) is processed in an encrypted state, and even the AI system itself cannot decrypt the original data.

2. Algorithmic Level: Achieve "encrypted model training" through FHE, so that even developers cannot directly observe the decision-making process of AI.

3. Collaborative Level: The communication between multiple intelligent agents uses threshold encryption, so even if a single node is compromised, it will not lead to global data leakage.

With the rapid development of AI technology, security issues have become increasingly important. Advanced encryption technologies such as FHE not only address the challenges currently faced but also lay a secure foundation for more powerful AI systems in the future. On the road to AGI, these security technologies are no longer optional but are essential conditions for ensuring the reliable operation of AI systems.