Manus achieved groundbreaking results in the GAIA Benchmark.

Recently, Manus achieved breakthrough results in the GAIA benchmark, surpassing other large language models of the same tier in performance. This means Manus is capable of independently completing complex tasks, such as multinational business negotiations, involving contract clause breakdown, strategic forecasting, proposal generation, and even coordinating legal and financial teams.

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

This development has once again sparked discussions within the industry about the evolutionary path of AI: will the future be dominated by AGI, or will it be led by MAS collaboration? The design philosophy of Manus implies two possibilities: one is to continuously enhance the intelligence level of individual units, approaching the comprehensive decision-making ability of humans through the AGI path; the other is to serve as a super coordinator, directing thousands of vertical domain agents to collaborate in the MAS path.

This debate actually reflects the core contradiction of how to balance efficiency and safety in the development of AI. The closer a single intelligence is to AGI, the higher the risk of decision-making becoming a black box; while multi-agent collaboration can disperse risks, it may miss critical decision-making opportunities due to communication delays.

The evolution of Manus also amplifies the inherent risks associated with AI development, such as data privacy, algorithmic bias, and adversarial attacks. For example, in medical scenarios, Manus requires real-time access to patients' genomic data; in financial negotiations, it may touch on a company's undisclosed financial report information. In recruitment negotiations, it may suggest salaries below the average for candidates of certain ethnicities; in legal contract reviews, the misjudgment rate for emerging industry clauses may be close to half. Additionally, hackers may implant specific audio frequencies to cause Manus to misjudge the opponent's bidding range during negotiations.

These challenges highlight a harsh reality: the more intelligent the system, the broader its attack surface.

In the Web3 field, security has always been a topic of great concern. Based on the impossible triangle proposed by Vitalik Buterin (the blockchain network cannot simultaneously achieve security, decentralization, and scalability), various cryptographic methods have emerged:

1. Zero Trust Security Model: The core idea is "Trust no one, always verify," emphasizing strict identity verification and authorization for every access request.

2. Decentralized Identity (DID): A set of identifier standards that enables entities to gain recognition in a verifiable and persistent manner without the need for a centralized registry.

3. Fully Homomorphic Encryption (FHE): Allows arbitrary computations to be performed on encrypted data without decrypting it, suitable for scenarios such as cloud computing and data outsourcing.

Among them, fully homomorphic encryption is considered a powerful tool for addressing security issues in the AI era. It can function at the following levels:

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

2. Algorithm Level: Achieve "encrypted model training" through FHE, so that even developers cannot glimpse the decision-making paths of AI.

3. Collaborative Level: Multiple Agents communicate using threshold encryption, so that the compromise of a single node does not lead to global data leakage.

In the Web3 security field, some projects have begun to explore these technologies:

• uPort was launched on the Ethereum mainnet in 2017 and is one of the early decentralized identity projects.
• NKN launched its mainnet in 2019, focusing on the application of a zero-trust security model.
• Mind Network is the first FHE project to launch on the mainnet and has collaborated with several well-known institutions.

Although these security projects may not receive as much attention as some speculative projects, they are crucial for building a secure Web3 ecosystem.

As AI technology approaches human levels of intelligence, non-traditional defense systems are becoming increasingly important. Technologies such as FHE not only address current security challenges but also lay the foundation for the future era of strong AI. On the road to AGI, these security technologies are no longer optional but essential for survival.