Picture source: Generated by Unbounded AI

If we look at the titles of AI self-media, no one in the world of AI should have survived yesterday because they were "exploded" again.

This time the bomb was thrown by a product called "Manus". It is an AI Agent product developed by a Chinese team that only released through an all-English promotional video and cannot be used without an invitation code. It is officially called "the world's first universal Agent".

It seems that OpenAI's Operator and Anthropic's Claude Coder were blown away overnight.

After the release, it spread rapidly in a small circle, "exploding" constantly, and quickly moved towards another familiar smell:

Another DeepSeek in China is here.

China certainly needs more than one DeepSeek, but what kind of existence it should be is a question.

1. It seems that Open AI and Anthropic do not exist

Let's look at these titles for now, let's see what product Manus is.

According to its videos and information on the official website, it is an agent that can "solve various complex and changeable tasks, and can flexibly call various tools in your own virtual environment to directly deliver complete task results, rather than just providing suggestions or answers."

Today, all the AI ​​agents who are difficult to distinguish will basically introduce themselves in this way.

It is not currently available for public use and can only be used through invitation codes, while its release is mainly through a demonstration video.

In the demonstration video, it shows three scenarios: uploading a compressed package containing multiple resumes. According to your prompt words, Manus will automatically decompress and analyze your resume, and then give suggestions; putting forward your real estate needs, Manus will automatically help you organize housing prices and information; generating an analysis report for Nvidia and Tesla's stock prices.

You should know that OpenAI's Operator and Zhipu AI's Phone Use have already shown similar functions in real-time. The Agent that Google Gemini showed last year was already multimodal, based on 10 minutes of memory to play games with you through the camera.

If you are familiar with the mainstream model capability iteration direction and the current situation of AI products, you will soon realize that these scenarios and function displays have been implemented in various Agent concept products, and the interface of Manus is basically Devin andWhat Artifact looks like.

The official does not have any detailed reports on the technology. Someone has disassembled the method behind it. It is basically a combination of Claude's basic capabilities and the post-training planning capabilities of the open source model. At the same time, various agents are built into it to complete the task through some set workflow.

What's interesting about it is that unlike most Agent products before, Manus "encapsulates" these capabilities into a cloud virtual environment, and it is also a new product in the "asynchronous" interaction trend.

However, these innovations in development engineering are a bit far from the "general" approach that the industry is already pursuing today.

Anthropic's recent definition of Agent in a sharing session is clearer than describing Agent: Agents are models using tools in a loop—it must first be a powerful model. As the model with the strongest Agent capability at present, Claude implements the "generic" method of the powerful model itself, plus a general protocol (MCP).

This also points to a problem faced by products like Manus, which is that its abilities are essentially possible for any model with stronger Agent capabilities. And Manus did not give a solution.

2. Want to be side by side DeepSeek cannot rely on self-entertainment and invitation codes

But these do not prevent Manus from becoming a product that "spreads" in a small circle.

From the design of this product, it aims at the "ordinary user" Devin, turning Devin into a part of the product process that users do not need to perceive, and then providing users with more direct uses. Such products are actually proven to have a market after Perplexity became popular.

Its approach is also very perplexity. Marketing is an important part. It will design the sharing mechanism. For example, users can share their operation recordings on Manus on social media, thus creating communication.

What is more classic is the invitation code mechanism.

Invitation code is a promotional design that always appears in a small circle's self-entertainment stage. It is also the basic operation of a product when it comes to the mobile Internet era, a kind of muscle memory from the previous era.

Manus's communication is a classic template reproduction of the mobile Internet: invitation code hunger marketing ("It's only through my connections") + KOL screaming evaluation ("Humans are defeated"" and "the body is numb") + Take advantage of other celebrity products(This time it's the DeepSeek binding). Yet when people open overseas tech forums and social media, the discussion about Manus is almost vacuum.

Does it be that bullies you don’t understand English? Where is the imagined “Sleepless Silicon Valley”? Even if someone doesn’t fall asleep, maybe the QwQ open source version of Tongyi Qianwen may be discussed.

Obviously what Manus wants to pursue is the popularity of DeepSeek. Based on the dissemination and expectations it creates, it also wants to be comparable to DeepSeek. But the way seems to be wrong. Deepseek has never created an invitation code. Before it became popular, it has been a product that has been deeply rooted in the developer community for a long time, and has formed a reputation through sincere open source.

But these early work became unimportant to followers who wanted to be comparable to DeepSeek. Later, the "domestic light" style of emergence became a technical operation rather than a continuous work.

Apart from the random search today, the title comes out:

"China AI Double Star: DeepSeek Technology Breakthrough, Manus Commercial Implementation".

Manus himself is also emphasizing its achievements at GAIA and says it beat OpenAI.

But why does an AI application compete on model manufacturers like OpenAI to compete on the list? Moreover, GAIA is a real-life task test that requires calling tool chains. Manus, as a system designed for tool calls, is like taking a calculator to participate in a primary school mathematics competition. Compared with OpenAI, a model without a specific purpose, it seems to be a wrong object.

Yes, the Manus team has a good background. It has produced an AI product that is somewhat innovative in engineering and front-end interaction, and then it has become a social asset in a small circle. With a round of "explosion" spread, taking advantage of people's deep imprint and emotions about the DeepSeek phenomenon, creating an emotion that does not actually exist "another Chinese team makes Silicon Valley sleepless" is indeed quite effective, but it is not serious.

China needs the next DeepSeek too much, but the "next DeepSeek" cannot be self-proclaimed and self-interested. The situation is fine now. If many people in Silicon Valley really noticed Manus, were lucky enough to get the invitation code, and then were disappointed and commented after the test, it would really consume the positive value that DeepSeek brings to China's AI, which should be something that everyone would not want to see anyway.

I hope this won't be the case.