GuShiio.com Wen| Digital Force Field, Author| She Zongming
“In the first 23 hours of the day, almost all the history of human technology was blank, and all major developments were concentrated in the last seven minutes of the day. rdquo;
Do you still remember these words from Schram, the father of communication, which I used when ChatGPT was first released?
Sorry, now I have to move it out again because Manus, the world’s first general-purpose AI Agent product, has come out.
Just as people outside the circle are still confused about AI Agent-style terms and the most explosive AI application words of the year, insiders have already asked for invitation codes.
The reality is there: Manus has become another AI product with the imprint of China that has been screened in foreign technology forums. The last one is DeepSeek, which is praised by some as a national sport-level scientific and technological achievement.
Even I, who was immune to big words such as generative AI’s iPhone moment and Agent’s GPT moment, watched the demonstration video where Manus independently analyzed 15 resumes, and finally generated an Excel ranking table faster than drinking a cup of American coffee. I still couldn’t help but feel a little shaken.
Unfortunately, I have no culture and can only say that I am powerful.
01
Here comes the question: What is Manus so powerful?
Scenario devouring power: DeepSeek is good at handling single-threaded tasks such as contract review, while Manus can complete complex links to crawl financial reporting Python deployment websites in parallel.
Evolutionary acceleration: DeepSeek achieved a 10-fold improvement in reasoning efficiency in 3 months, while Manus refined the task disassembly granularity to 0.1-second decision-making in the GAIA benchmark.
Ecological ambition: While DeepSeek is focusing on the model layer, Manus has built a multi-agent collaboration sandbox and announced that it will open source some models by the end of the year. This is similar to Android’s early strategy of using an open ecosystem to fight iOS.
You don’t quite understand it, do you?
You don’t know much, that’s what DeepSeek said.
As 0.5 small white, I want to say something that is within my understanding.
Many people know that in the past two years, the word Agent has become very popular. There are rumors that Agent, the next block-level AI application Agent, and the next generation of entry-level opportunity Agent, will replace App.
The most vivid expression is: the big model is the AI brain, and the Agent is the AI assistant Jarvis.
At the end of last year, He Baohong, director of the Institute of Cloud Computing and Big Data of China Academy of Information and Technology, pointed out that Agents will become the focus in 2025, and the seven-year itch for large models is emerging. Next, we need to move from the big model to Agents, which are goal-oriented, in contrast to the knowledge-compression properties of the big model. rdquo;
In his view, one of the signs of the second half of the AI model is the shift of content generation to the Agent Framework (AutoGPT), which supports tool invocation (API), task planning and dynamic interactions, and the rise of Agents.
So, why Agent?
The formula previously given by Weng Li, former vice president of OpenAI, is widely recognized in the industry: Agent= large model + memory + proactive planning + tool use.
Let’s just say that AI needs to move from imagination to productivity, and Agents are an excellent ladder.
In the past two years, the concept of Agent has become popular, but in reality, in most scenarios, migrant workers still have to manually define and choreograph the workflow themselves.
AI tools are useful, but their usefulness is limited.
At this time, Manus pointed to the world’s first universal Agent logo on his body and said: Why don’t you try mine?
02
“Universal, what does it mean?
Technicians may throw out a lot of terms and terms: full-link delivery (it can be executed directly to the output of the result); cloud asynchronous (it can automatically work on the Cloud Virtual Machine and notify users after doing a good job); reliable data (it will automatically invoke authoritative APIs instead of using unknown data sources casually); good at code invocation (it will write code yourself to call different tools to complete data visualization).
This is almost a summary of the Manus demonstration video.
Some people have already made a concrete summary of Manus’s characteristics:
1. Strong tool call capabilities
Manus can not only understand your needs, but also directly call various tools, such as browsers, code editors, and data analysis tools to directly help you complete tasks and directly deliver them to the finished product.
Manus is a migrant worker in the cloud and has its own independent computing environment. You don’t need to stare at it to work, which is extremely worry-free.
3. Collaboration experience similar to human colleagues
Even if you adjust the direction of the task at any time and change your needs halfway, Manus can respond flexibly and not get stuck at all. Not only that, it can also remember your preferences and follow your preferences directly the next time, and the more you use it, the more you can use it.
4. Ability to handle tasks in multiple fields
Manus is the all-powerful king. Whether it’s education, finance, travel, programming or data analysis, it can easily handle it. It can help you do in-depth research, document sorting, visual analysis, and even generate personalized content based on your needs, such as travel manuals, research reports, codes, etc.
5. Continuous optimization and learning ability
Manus will also continue to learn and optimize. You can add your own requirements through its knowledge system, or let it remember a certain working method and use it directly next time.
Manus drew the picture of mens et manus(the unity of knowledge and action), which is also the source of its name.
To be honest, I have been hurt by Teacher Jia’s PPT, and I still have to be cautious and accept this.
But as long as Manus can deliver half of the demo video, it will be powerful enough.
03
Perhaps to some people, Manus is not a disruptive innovation at all.
On the Internet, there is a popular saying: To the extreme, it is both TPF and PMF, and ultimately leads to user value.
The implication is that Manus is ultimately a shell.
This is not unfair to Manus: Monica, who is behind Manus, started out with a leading product in the AI plug-in field. Whether it completed the cold start through the independent developer product ChatGPT for Google, or helped users access the latest SOTA (the most technologically advanced) models when GPT-4o and Claude 3.5 were launched last year, or DIY Bot, Artifacts writing Mini programs, memory and other functions, it all shows that Monica is a super stitching monster.
But this does not mean that Manus, as a breakthrough at the application level, has no technological innovation.
In “The Essence of Technology”, Brian Arthur once said: New technologies are not invented out of thin air. Technologies are created from previously existing technologies, including: a, built;b, aggregated;c, integrated.
He believes that the evolution of technology is self-creation, and the emergence of new technologies stems from the combined evolution of other technologies.
Even if Manus is not an innovation at the model level, it cannot deny the breakthrough of its technology.
In fact, even DeepSeek, which represents model innovation, was previously thought to have only built a $30 iPhone.
But everyone saw the result. Driving the revaluation of China’s asset values is a stamp confirmation of innovations such as Long Potential Attention (MLA) and Group Relative Strategy Optimization (GRPO) technology and Sparse Activation Network (MoE) architecture.
So does Manus.
Meituan and Didi are both application innovations, and application innovation is also innovation.
04
Judging from the industry’s reaction, many people have predicted the value Manus will bring to the technical side.
As far as I am concerned, I would like to talk about its more far-reaching value from a more general perspective.
That is: it evenes the difference between silicon-based and carbon-based species much more, and cuts off a large part of the barrier between the two.
After the advent of ChatGPT, many people said that generative AI technology has strong disruptive and strong embedded characteristics, and can be applied to thousands of industries at low cost.
This makes some people look forward to its efficiency improvement, while others are afraid of its substitution effect.
After all, Nassim Taleb once said that in any profession, 90% of people are ignorant and work through situational imitation, narrow imitation and semi-conscious role-playing, except social sciences and journalism, which are 99% and 100% respectively.
But two years have passed, and whether it is efficiency improvement or replacement, it is not as fierce as many people think.
The difference between AI and people is not just the ability to understand who the fish head refers to.
Science and technology observer Yuan Zheng believes that people are not special, what is really special is computing power. Watt, Maxwell, Einstein = computing power singularity of biological computing power = human intelligence supercomputing;ChatGPT= computing power singularity of biological + machine mixed computing power dominated by machine computing power, machine intelligence supercomputing.
In the past, what drove social progress was human intelligence supercomputing. The reason why machine intelligence and supercomputing cannot do many of the tasks that humans do is because they are not intelligent enough.
But Wang Xingxing said with action: Really? Then he threw out the humanoid robot.
Personalization makes AI more like people.
But Goldman Sachs is obviously still a little dissatisfied.
In the past few days, the Goldman Sachs analyst team has investigated Yushu Technology and believes that Yushu Technology’s most powerful humanoid robot, H1, has only 19 degrees of freedom (DoF). This means that it is not yet able to handle complex and delicate tasks.
It believes that it will be difficult for humanoid robots to achieve the same work efficiency as human workers in the next 2-3 years, and meaningful applications may take a time span of 5-10 years.
So, what if Ushu H1+Manus?
What is certain is that we are at least one step closer to reaching the Turing Singularity.
If ChatGPT is regarded as the first-order product of the biological + machine hybrid computing singularity, then DeepSeek, humanoid intelligent robot, and Manus are constantly pushing it further.
05
From DeepSeek to Manus, people can’t help but think of a saying when the stars of technology shone.
Such signs also seem to indicate that 2025 is the year of the explosion of technology.
In “When the Stars of Humanity Shine”, Zweig wrote: Dramatic and fateful moments are rare in an individual’s life and in the course of history; such moments often occur only on a certain day, an hour or even a certain minute, but their decisive impact spans time.
If there are any laws, such as DeepSeek and Manus, we should do the same.
