Manus has swiftly garnered widespread attention, with invitation codes fetching astonishing prices of up to 50,000 or even 100,000 yuan on Xianyu. During a Friday conference, senior officials delved deeply into discussions about Manus. Coincidentally, within the AI chat group Xingkong Jun frequents, numerous friends have secured invitation codes and shared their testing insights.

Recently, the internet has been abuzz with Alibaba's qwq, dominating major model websites and surpassing DeepSeek in evaluation results. Xingkong Jun has informed numerous software industry peers that qwq is a genuine productivity tool.

Manus's popularity in China is underpinned by several factors.

However, based on the demonstrated effects and the hands-on experiences shared by friends who have obtained invitation codes, Manus's capabilities are truly impressive.

Manus's greatest feat lies in successfully implementing an approach akin to ChatGPT's path, achieving remarkable results through brute force.

Fundamentally, Manus comprises several key modules:

1. Virtual Machine: A Linux-based virtual machine equipped with:

• Chrome browser for web access
• Python runtime environment for script execution, data analysis, and web runtime environment launching

2. Task Planner: Breaks down user-input tasks into a ToDo List. It's speculated to utilize the Claude model, given this step's criticality and the model's strong reasoning abilities. Claude 3.7 Sonnet appears to be a highly cost-effective choice.

3. Task Execution Scheduler: Executes tasks sequentially based on the ToDo List, selecting the most suitable Agent for each. Given the focus on Agent selection, a highly sophisticated model isn't necessary. Slightly fine-tuned open-source models like Qwen suffice.

4. Various Agents: Manus features numerous built-in Agents for diverse tasks, including complex ones akin to OpenAI Operator's web browsing Agent or Agents retrieving specific data via APIs. Each Agent logs task results back to the virtual machine.

5. Task Summary Generator: Upon subtask completion, the Task Execution Scheduler notifies the Task Summary Generator, which reads the ToDo List and subtask results from the virtual machine, summarizes them, and generates the final output. Depending on the task, this could be a research report or web application. Given its heavy reliance on reasoning and language abilities, a powerful model like Claude 3.7 Sonnet is essential.

Before large models achieve seamless human interaction, Agents represent the best—perhaps the only—viable approach, as large models alone may not suffice for true seamless communication.

Historically, Agents relied heavily on prompts and knowledge bases, addressing efficiency needs but lacking a true 'productivity' revolution. Manus's independent problem-solving ability marks a significant breakthrough.

Nevertheless, I believe Manus holds a slight lead, but its competitive advantage isn't substantial. This approach has sparked numerous inspirations, enabling followers to quickly catch up.

The Manus team previously developed Yiban Assistant, a formatting plugin for public accounts, demonstrating their proficiency in managing traffic.

Prior to Manus's rise, Xingkong Jun collaborated with the core team to create an AI Agent+RPA demo, which was a resounding success. Our internal discussions closely mirrored Manus's concept. If we could develop an RPA based on our demo, it would essentially become another Manus.

Barely had the words been spoken when, the following day, a group of young developers from MetaGPT implemented Manus's core functions within two hours after dinner and open-sourced it as OpenManus.

Currently, a truly universal AI Agent doesn't exist, though it may become feasible with significant computational breakthroughs.

Manus operates using virtual machine-like technology, demanding substantial computational power. Its primary applications are internet-based, with booking tickets and creating PPTs posing the greatest challenges.

What about industrial applications? Can Manus optimize my production line processes or generate an ERP system? At this stage, it's unrealistic.

The day a large model can instantaneously generate an ERP/MES to run business operations will mark the true arrival of a universal large model. When that day comes, the world's computational demand will skyrocket, delighting Huangpiyi.