The AI model Manus has made groundbreaking progress in the GAIA benchmark test.

Recently, the AI model Manus achieved state-of-the-art results in the GAIA benchmark, outperforming other large language models of its class. This means that Manus is capable of independently handling complex tasks, such as multinational business negotiations, including contract clause analysis, strategy formulation, and proposal generation.

The advantages of Manus are mainly reflected in three aspects: dynamic goal decomposition, cross-modal reasoning, and memory-enhanced learning. It can break down complex tasks into hundreds of executable subtasks, handle various types of data simultaneously, and continuously improve decision-making efficiency and reduce error rates through reinforcement learning.

The emergence of Manus has once again sparked discussions within the industry about the development path of AI: should it progress towards General Artificial Intelligence (AGI), or should Multi-Agent Systems (MAS) take the lead in collaboration? Both paths have their own advantages and disadvantages:

1. AGI Path: Continuously enhancing the capabilities of a single agent to approach human-level comprehensive decision-making.

2. MAS Path: As a super coordinator, directing numerous specialized agents to work collaboratively.

This discussion actually reflects a core issue in the development of AI: how to strike a balance between efficiency and safety. As single-agent intelligence approaches AGI, the risks associated with the opacity of its decision-making processes also increase. While multi-agent collaboration can disperse risks, it may miss critical decision-making moments due to communication delays.

The progress of Manus also highlights the potential risks in the development of AI:

1. Data privacy issues: In fields such as healthcare and finance, AI may need to access sensitive personal or business information.

2. Algorithmic Bias: In scenarios such as recruitment, AI may make unfair judgments against specific groups.

3. Security vulnerabilities: Hackers may interfere with AI's judgment through special methods, such as misleading its understanding of offers during negotiations.

These issues clearly indicate that the more advanced intelligent systems are, the broader their potential attack surface.

To address these challenges, security technologies in the Web3 domain may provide solutions:

1. Zero Trust Security Model: Emphasizes strict authentication and authorization for every access request.

2. Decentralized Identity (DID): Provides a verifiable identity identification method that does not require centralized registration.

3. Fully Homomorphic Encryption (FHE): Allows computations to be performed on data while it is encrypted, protecting data privacy.

Among them, FHE technology shows great potential in addressing security issues in the AI era. It can provide protection at the following levels:

• Data layer: All information input by users is processed in an encrypted state, and even the AI system itself cannot decrypt the original data.

• Algorithm level: By training the encryption model, it ensures that even developers cannot directly observe the decision-making process of AI.

• Collaborative Level: Communication between multiple agents uses threshold encryption to prevent single point failures from causing global data leakage.

As AI technology approaches human intelligence levels, establishing a robust security defense system becomes increasingly important. Advanced encryption technologies such as FHE not only address current challenges but also lay a secure foundation for a more powerful AI era in the future. On the path to AGI, these security technologies will become an indispensable guarantee.