China’s new general Manus AI agent experienced a system accident and server overload, but it is very comfortable and shows the real promise of AI assistants in the future.
Since General AI Agent Manas was launched last week, it has spread online like a fire. And not just in China, where the Wuhan-based start-up butterfly effect developed it. This entered the global conversation and praised the performance, including Twitter cofounder Jack Daisi and hugs Face Product Lead Victor Mustar, with impressive technology voices. Some have also called it “the other deepseek,” compared to the AI model earlier, which surprised the industry to its origin with its unexpected abilities.
Manas claims to be the world’s first general AI agent who uses multiple AI models (such as Cloud 3.5 Sonnets of Anthropic and Open Source Covens of Alibaba) and various independent fixed assets that use different AI models that use different independent fixed assets. (It makes it different from AI chatbots, including Deepsek, which is based on a single Big Language Model family and is mainly designed for conversation conversation.)
Despite all the publicity, very few people have a chance to use it. So far, 1% of users have received an invitation code in the waiting list. (It is not clear how many are on this list, but for how much interest has Manus’s Discord Channel more than 186,000 members.)
AIGuts technology reviews were able to gain access to the script, and when I gave it a test run, I found that the use it seems to work with a very intelligent and skilled trainee: While it is sometimes asked to do what is asked to do, misconceptions or guides cut the corner, which is compensating and optimizing. Ultimately, it is promising, but not correct.
Like the previous product from the original company, an AI assistant called Monica, published in 2023, the script is aimed at global audiences. English is set as a standard language, and the design is clean and minimal.
To enter, a user must enter a valid invitation code. The system then leads users to a destination page similar to Chat or Deepsek, which appears in the left column with previous sessions and a chat input box in the middle. The landing site also includes trial tasks hardened by the company, from business strategy development to interactive learning custom-sound meditation sessions.
Other logic-based agent AI tools, such as ChatGPT depressions, can autonomously network for the necessary information to break the psyche tasks and complete them. Whatever separates, the window is a “script,” “which not only lets users see what the agent does but also interferes at any time.
To keep it in the test, I gave three tasks to Manas: (1) Compare a list of remarkable reporters covering China Tech; (2) discover of two-bedroom property list in New York City; and (3) 35, potential candidates for innovators under (3) 35, a list made of reviews each year.
That’s how it was done:
Exercise 1: The first list of reporters like the script gave me only five names; there were five “respectable mentions” under them. I noticed that it was listing the remarkable work of some journalists, but did not do it for others. I asked the script why. The reason for this was that it was easy to be sure: it was lazy. The agent said to me, “This was partly due to a lack of time when I tried to speed up the research process.” When I emphasized continuity and completeness, the script responded with a comprehensive list of 30 journalists, remembered their current sales, and listed their remarkable work. (I’m glad to see that I cut with many of my beautiful colleagues.)
I was impressed that I was able to make top-level suggestions for changes, as much as someone would be with a real trainee or help, and it answered properly. When it originally ignored the changes in the positions of the employers of some journalists, when I asked it to see some results again, it corrected them quickly. Another good feature: The output was downloadable as a Word or Excel file, making it easier to edit or share with others.
However, Manas hit a garden when he reached the news articles of journalists behind the sidewalks; It often met CAPTCHA blocks. Since I was able to follow step by step, I could easily cope with them to complete them, although many media sites were still blocking the equipment concerning suspected activity. I am looking for the potential for major reforms here – and it will be useful if a future version of the psyche can constantly ask for help when facing this type of limitation.
Exercise 2: For the discovery of apartments, I gave Manas a complex set of criteria including a budget and several parameters: a large kitchen, an external location, access to the city of Manhattan, and a large railway station within a seven-minute walk away. Manas originally explained vague requirements such as “some types of outdoor spaces”, which literally, except a private roof or access to the balcony. However, after more guidance and clarification, it was able to collect a wider and more useful list, which was used to provide recommendations in the levels and clean bullet points.
The final output felt directly from Wirecator, which includes subtitles such as “Best Totally,” “Best Value” and “Luxury Options”. This task (including the front) took less than half an hour to collect the list of journalists who were much shorter (which took a little more than an hour), probably because property entries are more openly accessible and well-structured online.
Exercise 3: This was the largest in the scope: I asked the script to register 50 people for this year’s innovators under the 35 list. Producing this list is a very big effort, and we usually receive hundreds of nominations each year. So I was eager to see how well the psyche could do. This included work in the stages, including a review of previous lists to understand the selection criteria, identify candidates, gather names, and create a search strategy to ensure a diverse selection of candidates from around the world.
Related History
A fortune hangs on a crystal ball with the Deepsek logo for Taylor.
How to become a lucky Taylor for China’s youth
The AI-operated game analysis has become a new Ingal for disillusionment-seeking answers.
Developing a search strategy was the most time-consuming script part. Although this was not clearly emphasized, the script’s computer revealed the Window agent to browse quickly through the statements from prestigious research universities, technical prices and announcements in news articles. However, it again met obstacles while trying to reach academic articles and popular media content.
After three hours on the Internet—the script wisely asked me many times if I could limit the search – it was just able to give me three candidates with full background profiles. When I pushed it again to give a complete list of 50 names, it eventually arose, but some educational institutions and fields were heavily acquired, reflecting an incomplete research process. When I pointed out the problem and asked for five Chinese candidates to find it, it managed to compile a solid list of five names, although the results were diagonal against Chinese media. Finally, I had to leave after a warning warned by the system that script performance could fall if I continued to enter a lot of text.
My assessment: Overall, I found a very spontaneous tool suitable for the psyche with or without users. In two of the three tasks, it produced better results than ChatGPT Deep research, although it took longer to complete them. Manas appears to be best suited for analytical functions, which require extensive research on the open Internet but has a limited scope. In other words, it is best to stick to the talented human trainee thing during the work day.
Still, it’s not all smooth sailing. The psyche may suffer from constant crash and system volatility, and it can struggle when asked to treat part of the text. The message “Due to the current high service burden, it cannot be created. Try again in a few minutes. “When I tried to start a new request, I flashed on screen for some time, and sometimes the script freezes the computer on a particular page for a long time.
According to Manus’s chief researcher, Peak G, the team has a problem, which addresses the team. The Chinese media outlet 36 kroner reports that the copy of the cost per work is about $ 2, which is only a tenth of the cost of depressing, “so it. If the Manas team strengthens the server infrastructure, I can become a favorite alternative to the tools of individual users, especially white-collars, independent developers and small teams.
Finally, I think it is really valuable that the script’s work process seems relatively transparent and cooperative. It actively questions along the way and maintains important instructions such as “knowledge” in the memory for future use, which provides an easily adaptable agent experience. It is also very good that each session is repetitive and divided.
I hope that I will use the psyche for all kinds of tasks in both my personal and professional life. Although I am not sure that the comparison of Deepsek is quite correct, it acts as further evidence that Chinese AI companies are only running in the footsteps of their Western colleagues. Instead of innovating only on the basic model, they actively shape the autonomous AI agents to use their way.
