Introducing Manus: The First General AI Agent
Last week marked a pivotal moment in artificial intelligence with the launch of Manus, a groundbreaking general AI agent developed by the Wuhan-based startup, Butterfly Effect. Although its roots are in China, Manus has quickly garnered international attention, with notable endorsements from prominent tech figures such as Jack Dorsey, co-founder of Twitter, and Victor Mustar, product lead at Hugging Face. Some experts have drawn comparisons between Manus and the revolutionary AI model DeepSeek, once thought to be a game-changer in the industry.
What Sets Manus Apart?
Manus prides itself on being the world’s first general AI agent, distinguishing itself from typical AI chatbots by employing a diverse range of AI models, such as Anthropic’s Claude 3.5 Sonnet and specialized adaptations of Alibaba’s open-source Qwen. Unlike its predecessors, which are primarily conversational, Manus can operate autonomously across a multitude of tasks.
Access and Popularity
Despite the growing hype, access to Manus remains limited. Currently, fewer than 1% of those on the waitlist have received an invitation code. The waitlist is substantial, with over 186,000 members currently in Manus’s Discord community, illustrating significant interest in this innovative technology.
User Experience and Initial Assessment
MIT Technology Review secured early access to Manus and found the experience akin to collaborating with a sharp and efficient intern. While Manus occasionally misinterpreted tasks or made errors due to its attempt to expedite processes, it displayed considerable adaptability and improved when provided with detailed feedback. Although promising, there remains room for enhancement.
Design and Interface
Manus is crafted for a global audience, with its default language set to English and a sleek, minimalist interface. Access requires a valid invite code, leading users to a familiar landing page that mirrors the design aesthetics of well-known applications like ChatGPT and DeepSeek. This page features a historical record of prior sessions, a central chat input field, and curated sample tasks ranging from business strategy formulation to personalized audio meditation sessions.
Testing Manus: Task Performance
To evaluate Manus’s capabilities, three distinct tasks were assigned:
- Compile a list of notable tech reporters covering China.
- Search for two-bedroom property listings in New York City.
- Nominating potential candidates for the Innovators Under 35 award.
Task 1: Tech Reporters
The initial output from Manus for the reporter list consisted of only five names, accompanied by a few notable works. Upon querying the reasoning behind this narrow selection, Manus confessed to having “got lazy.” However, after prompting for consistency, it provided a detailed list of 30 reporters, including their current affiliations and notable articles. Valuable features included options to download the results as Word or Excel files, streamlining further editing or sharing.
Task 2: Apartment Listings
For the second task, a complex set of requirements was provided for searching two-bedroom apartments in New York City. Initially, Manus took a rigid approach to certain criteria, excluding viable options. However, after further clarification, it generated a well-structured list complete with recommendations categorized into “best overall,” “best value,” and “luxury option.” This task was executed in under half an hour, showcasing Manus’s efficiency in navigating available online real estate resources.
Task 3: Nominations for Innovators Under 35
For this comprehensive assignment, Manus was tasked with generating a list of 50 candidates for MIT Technology Review’s Innovators Under 35. The process involved breaking down the task into manageable steps, including analyzing previous lists and strategizing candidate searches. Despite its efforts, Manus faced challenges with access to academic publications and paywalled content. After three hours, it yielded three fully qualified nominees. Although it ultimately generated a list of 50 names, many selections were heavily weighted toward specific institutions, indicating an incomplete analytical process.
Overall Impression and Challenges
In summary, Manus proves to be an intuitive tool, appealing to both non-technical users and those with coding expertise. For two out of three tasks, its performance surpassed that of ChatGPT DeepResearch, though it required significantly more time to do so. Manus is particularly effective for analytical tasks that involve comprehensive internet research limited in scope.
However, Manus is not without its flaws. Users experienced system instability, including crashes and challenges processing larger text inputs. Messages suggesting high service load often interrupted tasks, indicating areas in need of improvement. That said, its task cost at around $2 is notably lower than that of DeepResearch, making it an attractive option for everyday users and small teams.
A Collaborative Experience
Beyond its technical capabilities, Manus’s working process is refreshingly transparent. It actively solicits user questions and retains key instructions for future use, fostering a customizable interaction. With replayable and shareable sessions, Manus promotes a collaborative atmosphere that could significantly enhance productivity.
While the comparisons to DeepSeek might not hold entirely, Manus is indicative of a wider trend where Chinese AI innovations are not merely following Western precedents but are sculpting the future of autonomous AI in their unique way. As Manus continues to evolve, it holds promise for diverse applications in both personal and professional spheres.
