Manus ignites AGI's virtual fire. Can DeFAI follow suit?

P 三春清梦 V 半日浮生 P

Manus is not as amazing as DeepSeek V3/R1, but more of a technical hype that integrates MCP and Operator.

After Deepseek released open source 5+1 days ago, Manus took up the banner of the world's road to AGI, right?

After carefully observing the product details, you may have remembered the date of Manus wrongly. It is just right to define October 22 last year as the release date. That day was the day when Anthropic Claude released its computer use. In other words, LLM jumped out of ChatBot and became the birth day of Agent wandering and exploring in cyberspace, but OpenAI's Operator will not be born until January 2025.

There are many concepts. Let's disassemble them step by step and use CoT (chain of thought) to get a glimpse of what Manus is.

AI Awakening: The Road to Freedom

The road beyond the dialog box is paved with authorization.

OpenAI's greatness does not lie in GPT. The Transfomer paradigm was invented by Google. The real innovation lies in using Chat as the first entrance to human-computer interaction. We can understand it as an intelligent database that can generally answer any of your questions, but it emphasizes more on "solving doubts" rather than "solving doubts for you." For example, you can ask ChatGPT how to treat a cold. GPT can list answers according to different situations, but it cannot make a specific diagnosis or place an order for medicine.

In this sense, the value of DeepSeek lies in making the model smarter (DeepSeek V3) and enhancing the diagnostic capabilities (DeepSeek R1), which can determine whether it is a viral cold or a cold caused by the cold weather.

But AI still can't help you buy medicine. At this time, the complete body of GPT is sealed in the dialog box, and we hope to release it.

Computer Use came into being. In terms of path design, it is similar to the simplest external forms such as Keyboard and Mouse Wizard, Apple Shortcuts and Apple Script, that is, they all replace the operations of human hands + keyboard, mouse (or screen clicks), but they are different in nature. You don't need to customize script rules, you only need to use dialogue to command Claude to perform corresponding operations.

At this point, AI can help you open the browser, enter the Meituan address, and search for cold medicine, but new problems will arise. AI needs your Meituan account to locate the pharmacy closest to you.

We need to give AI more permissions at the bottom layer.

This is also a necessary step for Anthropic to release MCP (Model Context Protocol), the model context protocol, and for OpenAI to launch Operator. The optimization within LLM has reached the local optimum. Now we need to make AI/LLM move, LLM and LLM need to call each other, LLM and external APIs need to be integrated with each other, and LLM and humans also need to collaborate further.

Let's talk briefly about MCP first, and then an article will be published to explain it in detail.

The value of MCP lies in the hope of building a universal API/SDK framework in the LLM era. MCP hopes to standardize the communication format between AI models and other applications. For example, Claude/OpenAI/DeepSeek all use the same format to call code completion or create Meituan's rules for buying medicine. In this way, no matter what model the user uses, Meituan only needs to configure the same interface.

This does not mean that OpenAI/DeepSeek or Meituan must abide by Anthropic's specific rules, but they can be used as a reference for design. Just like ONNX (Open Neural Network Exchange), the proliferation of models naturally requires corresponding collaborative standards.

However, no matter who you use, you need to tell your Meituan account password, authorize Alipay, and take over the call system to complete the process of positioning, placing orders, and answering and calling couriers. In the end, you need to go downstairs to the express cabinet to get the medicine. For the time being, AI cannot run errands for you. It will take time for embodied intelligent robots.

The significance of DeepSeek is that LLM becomes smarter under the premise of extremely low cost, and its Chinese reasoning ability far exceeds that of its peers. This is its great significance in technology and products, not to mention that the open source model makes AI more down-to-earth.

This is the trick of Manus. Manus is not an Operator of OpenAI, or follows the MCP rules of Anthropic, which is equivalent to reinventing the wheel.

Of course, the Chinese also need to make achievements in model standards and cannot follow the old path of operating systems and chips, but this has little to do with the so-called AGI, because we have not seen what the base model of Manus is so far. If it is a self-developed, more intelligent large model, it is indeed a cause for celebration.

DeFAI and AI Agent are still in progress

The opponent of the cross-chain bridge is not the chain abstraction, but the CEX; the enemy of AI Agent is not the intelligent body, but the wallet.

After Manus dominated the media with its internal beta code and coin of the same name, Web3 AI agent also tried its hand at refuting rumors. Virtuals announced the integration of Enso Shortcuts, which facilitates one-click interaction for users and currently supports 200 protocols.

The happy side is that Web3 AI Agent has begun to move beyond the model dispute and move towards real user needs. However, it is clear that the old problem of Web 2 will still exist. Which protocol standard should be supported?

Take cross-chain bridge as an example. LayerZero has basically become the de facto industry standard protocol after years of hard work, but it still cannot connect all scenarios. There is no other reason. CEX, especially Binance, is the most convenient cross-chain bridge for assets, and inter-chain message communication is not the current pain point.

The most important attempt of Web3 AI Agent is to establish the connection between users, itself and Uniswap/Hyperliquid, that is, AI Agnet must become the de facto middleman, private key holder or custodian. Otherwise, the user experience cannot be comparable to the wallet + DEX experience coupled with the existing infrastructure, let alone compete with CEX for the market.

To say this is not to deny the prospects of DeFAI, but to point out its real obstacles - not the degree of intelligence, but how to gain user trust. Manus needs to compete with MCP and Operator for the right to define standards, so the DeFAI project party also needs to have such awareness.

All AI Agent projects must adhere to long-termism, and constantly iterate and try and fail before they can wait for their initial users. In fact, DeFAI's opponent is the wallet product form, not other intelligent entities.

Just as there are two paradigms in the industry, namely, custodial wallets and non-custodial wallets, the biggest problem of AI Agent now is the lack of strategy and fund security. Fund security is as mentioned above, and the strategy lies in the user's authorization. Even if the user dares to authorize the Agent, he still needs to face the problem of strategy setting. In a word, is it reliable for AI to help users manage their finances?

The current model and framework competition of Web3 AI Agent has not yet been decided. For further strategy optimization, no project has been truly put into practical use. The Robotaxi that Musk once imagined is still on the way. When will the AI ​​financial management master enter every currency wallet?

Conclusion

It must be emphasized that this article is not a denial of Manus. After all, Workflow + Claude + Cursor are already useful enough, and a little more is fine. If you don’t eat the AI ​​bubble, others will.

This article is not suitable for denying Web3 AI Agent. After all, staying up late to watch the market + watching private keys + Safe is safe enough to make no mistakes. Letting DeFAI play PVP for others can also save the youth of staying up late.

Just one thing, don’t fake it. If you fake it, your nose will grow longer.