Manus is great, but it still has flaws

Manus has become popular these days. This AI tool that demonstrates powerful agent capabilities is indeed impressive, but behind this amazingness, are there some fundamental issues that we need to think about? When facing general agents represented by Manus, should we be optimistic or pessimistic?

The following discusses several practical challenges facing general artificial intelligence based on large models and attempts to propose some possible solutions.

The gap between the Internet and the real world

As cool as Manus is, it can’t get past one basic fact: it relies entirely on the Internet and Internet data to perform its tasks. But our world is much more than the digital part.

There is a huge gap between the Internet world and the real world, and many implicit information at the language level cannot be captured. Taking the stock selection function in Manus's demonstration as an example, can AI truly understand a company or industry structure based solely on public information on the Internet? Obviously not. Real investment decisions often require on-site inspections, communication with management, and other information channels that the Internet cannot provide.

Case focus : In Manus's demonstration, it can quickly analyze financial report data and market comments on the Internet and recommend stocks to users. However, when asked to evaluate the investment value of a startup, Manus cannot obtain key information such as the founding team's execution ability and the company's internal culture, which are often the most important considerations in investment decisions.

The limitations of Internet data itself

Even if we focus only on Internet-wide tasks, we still face many obstacles:

1. Paid content barriers : More and more high-value content is protected by paywalls
2. Database silos : A large number of professional databases cannot be freely accessed by AI
3. AI-generated content pollution : AI-generated content is flooding the Internet in a way that "beats magic with magic"

Imagine a simple scenario: "I want to order a good Hunan takeaway from Ele.me." Can an agent like Manus fully handle this task? It does not understand Ele.me's sorting rules, the possible rating manipulation of the store, or your personal taste preference for Hunan cuisine.

Case focus : When Manus searches for professional information in a specific field, he cites AI-generated content as an authoritative source more than 40% of the time, without realizing that the content itself may lack factual basis. This cycle of "AI citing AI" may lead to information bubbles and error amplification.

Solving real problems or meaningless automation?

When we further narrow the scope of AI use to processing local data and automating workflows, it can indeed solve some problems, but are these really the most urgent problems that need to be solved?

In the video demonstration, AI can quickly screen resumes and make summary tables, which looks cool. But in the AI ​​era, job seekers are increasingly providing portfolios or GitHub links instead of standard resumes; corporate leaders are also beginning to advocate reducing PPT reports. In this trend, how much significance does it have to use AI to optimize old workflows?

The control dilemma

From a more macro perspective, we face a fundamental choice: should we treat large language models as a new link in the existing controllable system, or should we give full control to these inherently uncontrollable models?

Manus obviously chose the latter, but this choice is worth pondering. How much work and decision-making processes are we really ready to let AI take over?

The Essence of Computing Power and Intelligence

Some artificial intelligence scholars have long been looking forward to an AI that can achieve rapid generalization and ability transfer with minimal resources, just like humans. This kind of intelligence is closer to the essence of human thinking - efficient and precise.

However, the current large-scale model development route led by OpenAI and others is more like a "miracle through great effort" approach to demonstrate imperfect but practical skills. Does this route deviate from our expectations for true artificial intelligence?

Future direction: Bridging the gap between AI and reality

In the face of these challenges, future AI development may require breakthroughs in the following directions:

1. Hybrid Intelligent Architecture

Future AI systems may need to combine symbolic logic and neural networks to create hybrid systems that can handle both explicit rules and fuzzy cognition. Such systems can provide more reliable reasoning capabilities while maintaining flexibility.

2. Enhance embodied perception

Providing AI with the ability to interact with the physical world, such as combining robotics and IoT devices, enables AI to obtain information outside the Internet. Research institutions such as MIT and Stanford have begun to explore how to enable AI to learn physical laws and common sense through physical interaction.

3. Personalization and continuous learning

Develop AI systems that can continuously adjust to individual usage habits and preferences. Rather than pursuing general AGI, it is more practical to build specialized AI assistants that can deeply understand specific user needs over time.

4. Human-machine collaboration rather than replacement

Position AI as an extension of human capabilities, not a replacement. Design human-centric interfaces and interaction models so that AI complements human intuition and expertise, not competes with it.

Case focus : For example, combining Manus with wearable devices enables AI to obtain real-time contextual information about users, such as location, activity status, and physiological data. This approach helps AI consider the user's current location, mood, and health status when recommending restaurants, rather than relying solely on reviews on the Internet.

Large model with missing "limbs"

A large model without "limbs" is like a hungry ghost looking at a pot of hot chicken soup without a spoon to scoop it out. Even if the model understands the world, without the ability to interact with reality, its intelligence will hardly work.

Manus is undoubtedly an impressive technical achievement and represents the forefront of current AI capabilities. But it also reveals to us the fundamental challenges in the current development path of AI. As technology advances, we need to think more deeply about the relationship between AI and humans, AI and the real world, and how to design intelligent systems that can truly address core human needs.

Future breakthroughs may come not only from bigger models and more computing power, but from a deeper understanding of the nature of human intelligence and a closer connection between AI and the physical world.