After thinking about it, I think many readers would like to read my review of Manus, so I decided to share my true feelings after 8 hours of live broadcast.　

First give an overall evaluation (I really hope you can see the final summary) :　

Overall, Manus has done a great job in engineering and productization. It is easy to use and has a clear sense of value.　

So at the beginning of the article, I have to give a thumbs up to the Manus team . It is really not easy to "think of it", "do it", and "make it to the point where people are willing to use it".　

This article does not discuss too deep technical principles, but just wants to explain to readers who have used or not used  Manus in a simple and easy-to-understand way what it can do, what its shortcomings are, and what it means .　

First, let me tell you what Manus is.

• The current Manus ≈  AI operates a Linux virtual machine & browser without a graphical interface , perceives the computer environment, and performs various operations .
• So it can run various Linux commands, libraries, programs (cd, ls commands, python...), and can also access various web pages and obtain data from some API interfaces.
• But because there is no graphical interface, it cannot run graphical programs. For example, I asked it to run "Pokemon Go", but it terminated during operation.
• When accessing a web page, various factors that hinder human use will also disturb Manus. For example: mandatory login, recharge pop-up window when the balance is insufficient
• Manus has no web account and no money. So in order to facilitate the user to intervene with the keyboard and mouse, Manus provides two options: command line window, browser, and vscode, which are visible to users  , so that they can view running instructions, take over web pages, and modify files. This also reversely explains the current boundaries of Manus's activities.
• As for Manus AI, it can smoothly interact with web pages, read data, click elements, play games, and even switch from a non-graphical interface to a user-accessible graphical interface at any time. This may be the traditional skills of Peak and Red. One is a browser giant, and the other is a browser plug-in giant. They are good at browser-related engineering development.
• By the way, you can still upload files to Manus, and I think it will be able to connect to private APIs in the future, so there is room for imagination.

Okay, let’s go back to its operating logic.

• Manus's main intelligence is Claude AI and a specially trained reasoning model. When it faces a human instruction, it will generate a todolist document according to the idea of ​​"splitting the sub-steps required to complete the task according to the human instruction", disassemble the task and reduce the difficulty of implementation.
• Based on this todolist, tick a ✅ after completing each subtask to indicate the execution requirements of each step in the multi-step task to the AI ​​(because the current AI's context memory length is limited, a memo is also needed)
• Then, for each to-do in the list, call relevant tools (linux commands, python, browser, etc.) as needed, advance the sub-steps one by one, and evaluate the completion quality. These steps can be installing software environment, creating documents, browsing the web, writing programs & web pages, etc.
• The results of multiple sub-steps can be nested and used to produce the output of the original instruction. For example: research data document summarized from web search and web reading + web page style + ... = generate xx market research analysis report web page
• Of course, Manus is not limited to making web pages and reports; he can also play 2048 games in HTML, write game guides, help me install Pokémon games, install NDS emulators, and log in to the AI ​​input prompt on my behalf to make AIGC videos. >>> You need to use your imagination here
• So it is indeed a universal Agent

How is the case effect?

1. 1.  It really created a video, using opencv to draw it frame by frame, but it's a bit ugly, and it can't add sound by itself. But after all, it made a video (some friends have also tested it with BGM)
2. 2.  It can help me draft emails, but it does not support sending emails through email services. There may be official restrictions.
3. 3.  I really played it, and it exceeded my expectations! The highest score was 192 points, which is much lower than that of humans (theoretically, it is possible to get more than 100,000 points). Each step was also very slow, but after all, I simulated keyboard and mouse input to play. It seems that Manus can play all non-real-time games. Since the context length of a single case is now limited, I am applying to the official to lift the limit to see how many points Manus can get without restrictions.
4. 4.  This case is amazing. I don’t understand how well it works. I hope everyone can try to analyze it :
1. It seems that it has found a way to make the game run by actively transforming its environment? To be honest, I don’t understand how far it has gone. Is it really done or is it an illusion?
1. It researched the requirements for running the game and installed the emulator and its dependencies.
2. But the game image resources are really hard to find, and the network disk download also requires logging into a human account and downloading the network disk client. So I gave it some special treatment and directly sent the package to it.
3. Then the installation went smoothly. You know, simulator configuration is not that convenient. This means that we don’t need to teach our elders how to use the software remotely in the future. Agent will be the best partner for each of us.
4. When running the game, Manus' sandbox did not support graphics services, so I gave it some "encouragement" and it actually started trying to improve its own virtual display alternatives in a headless environment.
5. 5.  I independently designed the script and text prompts for the entire AI video. I took over and logged into Keling AI, and was able to successfully adjust the management console parameters to try to generate it. However, because there was no money in the account, a charging pop-up window popped up, and I couldn't continue if I didn't want to recharge.
6. 6.  Most of it was limited by product engineering, but in the end, some design content that could not be found on the official website was output, and it was officially confirmed that most of it was correct. Manus can really perceive the environment and reflect on the impact of the environment on it. It is simply the AI ​​Agent version of "The Truman Show" .

So what is my assessment of Manus?

Can you think of it as a very hardworking intern?

• As mentioned earlier, it uses the LLM model to break down the original task into multiple subtasks, and then gradually generates action steps/content/plans for each subtask...
• Therefore, the ability of each of its subtasks is the current intelligence level of LLM - a college intern who lacks practical experience, lacks spirituality (long-term self-improvement and growth space), but can work at a very high throughput and very fast speed 24 hours a day.
• The final product is basically at the level of an intern (depending on the AI ​​level of the Agent core and the quality of the data it can access).
• PS: During the live broadcast, I also asked a professional financial analyst friend to comment on the level of the analysis report in the Tesla stock analysis case published on the official website. The evaluation he got was: "We don't look at fancy dashboards. The level of the report obtained by actual analysis is probably at the level of a freshman intern, which is not usable in actual combat. Maybe providing private data warehouse APIs and the analysis report templates and requirements we use in actual combat can provide more practical and valuable results."
• However, I would like to suggest again that this part is too subjective, and it is recommended to experience it yourself after the public beta!!!
• It doesn’t mean that Manus is not good and can’t solve problems. It means that the performance of this type of product will be closely related to the base model AI and will progress together (“shell” products can have unlimited growth space?)
• The official said that the token cost for each case is basically 2 US dollars, and as the model develops, this price will drop infinitely. This is also much lower than the cost of human interns, and the future organizational structure will inevitably be very different from the current one.

But Manus has some problems, or the problems that all AI agents at this stage will have.

• In the process of human-machine collaboration, it is a very bad experience for humans and AI to repeatedly pull and modify the intermediate products of the task process. Once a good result cannot be produced in one go, it is easy to fall into the dilemma of endless repeated prompts for modification.
• It's like letting an intern do the work: if the intern doesn't do a good job, he or she will easily make changes over and over again, and it's basically difficult for him or her to make a good decision based on his or her own "blind guessing". You must be more patient and give more background information, sub-step guidance, and more informative feedback.
• After all, Manus is not our personal assistant, and it does not support cross-task dialogue memory, account login status, file interoperability, and text language is a lossy compression of real information , so it is difficult for it to grasp the complete background information of what you want it to do. Many things require you to explain them honestly. It can't read your mind, nor can it read the atmosphere of tension you feel because it didn't do a good job.
• At present, Manus can indeed help you save a lot of effort in some tasks with clearly defined processes/low delivery expectations. However, in more scenarios, it may not be able to bear more expectations due to reasons such as data quality, limited background information, and inability to continue to grow.

So I am also asking myself a question: At this stage of manus, can I find the main scenario for using it in the long term?　

I haven’t figured out the answer yet. My intuition is that it’s my own problem. We should all try more and think about the best human-machine collaboration model between ourselves and AI .　

The release of Manus has made more people see the future clearly again

Although many people say that the technologies used by Manus have no barriers, and some parts of them are projects that were verified half a year ago.　

But this does not prevent Manus from being the only product that combines AI, engineering and productization so well.　

It is not easy to think of a product, make it and make it so that people are willing to use it.　

It opens up almost all links, allowing humans to complete complex tasks in a one-stop manner with extremely low experience cost (available in China, free trial, with invitation code restrictions, but being gradually released at a rate of thousands of people per day).　

Under the stimulation of Manus's physical experience, many practitioners and general users have begun to think further: What should our products look like in the AI ​​era? What will the future human-machine collaboration look like? How should humans and AI collaborate and coexist?　

Summarize

At this point, as the industry's first popular universal agent (and still in the alpha testing stage), Manus and his team have already received a lot of attention and public pressure.　

When I was thinking about the title of this article, I actually hesitated for a long time, wondering whether to exaggerate it to such an extent.　

Although there is no conflict of interest, we ultimately chose the version you see now.　

Because they are worth it:　

—— With its extremely high product completion, Manus has once again blown away some of the fog that separates the future of humans and agents, revealing more light.