“The next DeepSeek”? Manus hasn’t woken up Silicon Valley yet

The emergence of Manus has caused a stir in the domestic AI circle, and it is even expected to be the "next DeepSeek". However, has this product, which is promoted as the "world's first universal AI agent", really "woke up" Silicon Valley as the outside world claims? This article will analyze the true strength of Manus from the perspective of the product itself, market reaction, technical limitations, and the team background behind it.

When I woke up this morning, my screen was flooded with news about “Manus flooding the screen”.

Thinking it was a bombshell, I opened my computer with trembling hands to search, but found that the English media was as quiet as a chicken, with no reports on this "world's first universal agent".

Isn’t it said that the “entire technology circle” is losing sleep and Silicon Valley is in tears?

There are indeed a lot of discussions on social media such as X, but most of them are Chinese self-media. Before the release of this article, there were indeed discussions in English, with many compliments, and foreigners also asking for invitation codes, but they also mentioned the names of some other AI proxy products, and some people pointed out the limitations of Manus.

On the other hand, on a domestic second-hand trading platform, some people have been selling Manus invitation codes at high prices, ranging from a few thousand yuan to 160,000 yuan. This has caused Manus officials to "clearly declare that the invitation codes released by XX and other platforms are false information" and remind users not to believe them.

Judging from the official demonstration and trial videos released by a few third-party users, Manus is indeed a very powerful AI agent product, and the experience may even be better than the products currently launched by some large model companies themselves.

The team behind it had previously launched another product that has withstood the test of the market, Monica.AI. The founder, Xiao Hong, was born in the 1990s, graduated from Huazhong University of Science and Technology, and is a serial entrepreneur.

AI agents are already a hotly contested field among AI vendors. OpenAI also launched its first AI agent product Operator a month ago, which is now available in some countries and provided to Pro subscribers.

The latest news is that The Information said OpenAI is planning to launch different levels of AI agents, among which PhD-level agents may charge $20,000 per month.

The time when this report was released was exactly when Manus said it "blew up the AI ​​circle."

An association naturally comes to mind:

At the beginning of this year, DeepSeek took the AI ​​circle by storm. The low-cost inference model made the entire Silicon Valley re-evaluate the path of large models. Liang Wenfeng, a post-80s from China, posed a difficult question to Sam Altman of OpenAI.

So now, has there been another person born in the 1990s from China, Xiao Hong, who has snatched away Ultraman’s “AI agent dream”?

It's understandable to be excited, but don't get too excited yet.

Let the bullets fly for a while.

01

Let’s start with the Manus product itself.

Official introduction: "Manus is a general AI agent that connects thoughts and actions: it not only thinks, but also delivers results. Manus is good at handling a variety of tasks at work and in life, and can get everything done while you rest."

Everyone is familiar with AI agents. Unlike ordinary chatbots, they can really "take action" to help you do many things.

As for Manus, in the official example, you can throw a compressed package containing 10 resumes to it (yes, no need to decompress it), and Manus can decompress the file by itself, browse the file page by page and sort out the key points, write Python scripts to generate corresponding tables, and finally complete the analysis of the job applicant's resume.

During the whole process, users can see Manus's entire thinking and execution process on one side of the interface, and all operations in "Manus's computer" (running in the cloud) will be displayed on the other side. From the demonstration video, the overall process is very smooth. Whether it is decompression, Python programming or direct generation of Excel tables, they are all very impressive.

In addition, the official also gave other examples, such as recommending New York real estate within the budget and conditions given by the user (low crime rate, high-quality educational resources, etc.). In this example, you can see that Manus first disassembled the task and made a plan.

It is reported that Manus tried to use the "Multiple Agent" architecture to run a variety of different agents in the virtual machine to divide the work and collaborate, such as planning agents, execution agents, verification agents, etc.

Currently, Manus is not open to everyone. To use it, you need an invitation code. Invitation codes are not easy to get, so many people "ask for a code" on social media. At around 1:00 p.m. on March 6, Beijing time, Manus could no longer open the login interface, presumably because the influx of users exceeded the load and caused a freeze.

The team that made Manus is really talented.

One of the founders, Xiao Hong, was born in the 1990s and graduated from the School of Software of Huazhong University of Science and Technology. Ji Yichao, who appeared in the Manus official video, is also the company's "co-founder and chief scientist." He was also born in the 1990s and became famous at a young age. The Mammoth browser he developed attracted investment from Zhen Fund. At the age of 19, Ji Yichao was on the cover of the Chinese version of Forbes.

The team's previous masterpiece is the AI ​​browser plug-in Monica, which focuses on "AI assistant". As of July 2024, it has 4 million users.

02

It is impossible to determine how the inspiring descriptions such as Manus "blowing up the AI ​​circle" and the "GPT moment" that opened up AI agents first appeared.

But what is certain is that Manus is really popular in China.

Putting aside the news, let's take the most intuitive "invitation code" as an example. At around 3 pm, Zhimian AI (ID: faceaibang) saw someone selling the "Manus invitation code" on the second-hand trading platform, and the price was increased by 30,000 to 160,000 yuan.

When we asked as ordinary buyers, “Can’t it be any cheaper?”, the seller with the price tag of 160,000 yuan replied, “We are out of stock. There are a lot of people asking about it.

It is conceivable that this phenomenon is based on the news that Manus blew up the AI ​​circle and even made Silicon Valley "sleepless". Many people have even called it "the next DeepSeek".

It is not yet known whether Manus will become the "next DeepSeek", but what is certain is that, at least for now, Silicon Valley is eating, drinking and sleeping as usual, and has not been "exploded" to the point of being dizzy.

Until three o'clock in the afternoon, we searched "Manus" on Google again, and there were still only five related news reports, and no major media such as Bloomberg, The Information, Business Insider, and Reuters followed up.

On social media X, there have been some non-Chinese messages discussing Manus, but not to the point of being overwhelming.

For example, @AshutoshShrivastava, who has 40,000 followers, also posted a video of Manus 12 hours ago, and then said that he got the invitation code, thanked Manus officials, and said, "I tested it last night, it was very good, but I was too sleepy to record anything." He then posted several more messages, but none of them were related to Manus.

Interestingly, after the Manus-related news, he immediately shared Ali QwQ-32B and said: "China will one day open source AGI for all mankind."

Again, Manus is good, but it doesn’t seem to have taken the world by storm yet.

HashMatrix co-founder Pang Shunxin also shared his joy of trying Manus, but also said: "The demo is cherry-picked, it's not that amazing in practice, but the horizontal comparison is really great."

Zengyi Qin also shared his views. He is a PhD student at MIT and also a startup, and is the AI ​​R&D director of MySell (the company raised $11 million last year, and investors include the author of Transformer).

He also affirmed the product itself: "Manus is an excellent product. This is obvious and I don't need to say more."

But he believes that Manus "is not a technological breakthrough" for three reasons:

a. It mainly performs tasks according to a preset process. This preset process will significantly limit the upper limit performance of the system, although it does make the system run better in the short term.

b. It works in a relatively restricted environment, somewhere between browser-only and OS-level. It cannot function outside of predefined software tools. For example, it cannot control PowerPoint to make a slideshow for you. In fact, a lot of software is outside of Manus' toolbox.

c. If Manus is not a technological breakthrough, then what is? The answer is: a system that can run in a completely open operating system-level environment, perform most tasks, and operate software like a human would be a technological breakthrough.

03

The AI ​​agent track is very lively, and large model manufacturers have already taken action.

For all to C users, OpenAI has Operator, and Anthropic has launched Artifacts and Computer Use functions. In addition, Google, Microsoft, and domestic large model manufacturers have also taken actions on AI agents.

OpenAI's Operator performs tasks in the cloud according to user instructions, such as ordering food, making plans, shopping, etc.

Anthropic's Artifacts allows users to generate code snippets, text documents, or website designs during a conversation, and display these results in a dedicated window next to the conversation with one click. Computer Use allows AI to "use computers like humans", watching the screen, moving the cursor accurately, clicking buttons where needed, and using a virtual keyboard to enter text.

Further action is on the way.

The Information reported that OpenAI executives told some investors that the company plans to launch agents with different prices for different users: $2,000 per month for "high-income knowledge workers"; $10,000 per month for software development; and $20,000 per month for doctoral-level research agents.

People familiar with the matter also revealed that in the long run, OpenAI expects 20% to 25% of the company's future revenue to come from agent products.

Many people talk about Manus and mention another product called Devin. This is a vertical product similar to Manus, launched in 2024. It also became very popular as soon as it debuted, and was even called the "first AI software engineer."

But this year, Devin's shine has dimmed and he has been criticized for "poorly completing most tasks." For example, people found that a job that they could complete in half an hour would take Devin six hours.

Many companies that develop AI applications will act as AI agents around various vertical tracks. Manus is not a large model manufacturer, but it does not act as a vertical agent. This is commendable, and if it still performs well after being fully launched on the market, it will be even more gratifying. But it is also this point, coupled with the extreme praise from the outside world, that makes people sweat.

AI agents for all users will inevitably continue to be the target of large model manufacturers. Will Manus's living space be squeezed?

This seems to have returned to the dispute between Zhu Xiaohu and Fu Sheng two years ago: Can starting a business around GPT succeed?

Now this question can be further supplemented: Is it possible to start an AI application-layer business in a field where large model manufacturers are likely to enter the market in full force?

AI applications based on mainstream large models continue to emerge, and there are countless good products, but tragic cases are also common.

Jasper, a SaaS company targeting advertising marketers, self-media bloggers and other groups, was once valued at US$1.5 billion and became one of the first unicorns to take off with the big model trend. However, the year after ChatGPT was released, Jasper AI's space was constantly squeezed and its glory was gone.

Not long ago, Xiao Hong, the founder of the team behind Manus, was interviewed by the media person Zhang Xiaojun. When asked: "Will the original manufacturer or the application company do the Agent?"

Xiao Hong replied: "I think the original factory is worth doing, but it seems that the original factory is not doing well enough."

There is no need to criticize Manus. Pang Shunxin, who said that Manus "is not that amazing", also spoke out on X, opposing the criticism that Manus "does not have deep technical barriers", saying that products are useful technologies, and the fig leaf of useless products is called "technology", "so don't look down on engineering and products."

I hope that in the future this Manus product can leverage the flexibility of its application layer and find differentiated advantages in competing with “original” products.

But at this stage, let’s not praise it too much.

Written by: Bi Andi Edited by: Zhao Jinjie