Manus AI pushes the DeepSeek moment further

By Catherine Thorbecke

After the global hype generated by Manus AI over the past week, obtaining the invite-only access code to try it out felt like scoring free front-row tickets to a Beyoncé concert.

The artificial intelligence agent from Chinese startup Butterfly Effect went mega-viral since its release of a demo video that seemed almost too good to be true. Twitter co-founder Jack Dorsey called it “excellent.” The head of product at leading AI startup Hugging Face said it was “the most impressive AI tool I’ve ever tried.” Its exclusivity only compounded the hype. Just 1% of those seeking access to the beta version have gotten in, and invitation codes are being listed on reselling websites for upwards of $1,000.

After spending a couple hours playing around with the agent, I found it imperfect and glitchy — but still incredibly impressive. Manus doesn’t represent a second DeepSeek shock, but something else entirely: It reveals that Chinese startups can compete with comparable US companies building AI products.

DeepSeek released its own impressive AI models, and at a fraction of the cost of US firms. Manus was built on outside models that were fine-tuned to create a practical service beyond just a chatbot: It can complete a range of tasks on its own. Manus might not be the next leap in artificial general intelligence, but it is a reminder of how this emerging technology can be useful to human beings now.

While the DeepSeek moment was about the technical breakthroughs, Manus is about product innovation, showing Silicon Valley that it should pay more attention to practical consumer applications. Most people don’t need or want AI to take over creative endeavors, they want the technology to help with more tedious grunt work.

I asked Manus to build me a personal website, and within roughly 30 minutes it fully deployed a temporary URL with a simple design and everything I had asked for. It drafted a bio page based on publicly available information, included a portfolio tab with links to my latest columns, and a contact page with my email, LinkedIn and X accounts. The copy could use a light edit, but it was a remarkable first draft. It even offered instructions for deploying the website to a permanent hosting service, or gave me the option of downloading the project files. This has been on my to-do list for years. Without the agent I would have either had to hire a coder, pay a software service, or spend days building the site myself.

There were still noticeable glitches. It crashed while conducting some real estate market research for me and had trouble accessing some websites due to human-verification restrictions. It made me a highly detailed itinerary for a weekend train trip, but stopped short of being able to book tickets. Yet it was remarkably user-friendly, outlining each step of the process so I could follow along and intervene if needed.

Manus has defended some of these issues by saying that the current version is far from the final product, and the invite-only status is due to “genuinely limited server capacity.” But even the beta version offered some impressive capabilities.

DeepSeek, meanwhile, published its research online and let this speak for itself. Most technical details about Manus have been shared via social media posts that reveal it was built by fine-tuning Anthropic’s Claude and Alibaba Group Holding’s Qwen models. But Manus has never claimed to be breaking the frontiers of AI model research — it was always a product.

Leading US AI companies have long said that so-called agents, or tools that can operate autonomously for users by browsing the web and completing tasks, are the next big leap in this technology. This intensified when OpenAI unveiled its own similar offerings including Operator and Deep Research earlier this year. Manus claims that its service outperforms Deep Research on industry benchmarks, and is widely speculated to do so at a fraction of the cost.

The likes of OpenAI have raised billions of dollars as they chase AGI, or building models that are as smart or smarter than humans. But what Manus brings to the table is much more grounded in real-world applications, in a way that should put Silicon Valley on notice. It further commodifies the underlying technologies being built by large labs and adds new urgency to the race to create products on top of them.

There are myriad valid reasons the global industry has been cautious about giving AI agents too much autonomy. Back in the US, they have shown the potential to go rogue — in one viral example, OpenAI’s Operator spent $32, on its own, to buy a tech columnist a dozen eggs. One can only imagine how much worse the consequences could be as these tools are given more power. There is also the mountain of geopolitical concerns that will rise from the unleashing of Chinese-originated AI agents to consider, along with the labor market impacts. Policymakers should keep pace with the sector, and urgently develop frameworks to address the new safety and liability issues ushered in by this new era.

Manus also showed that China knows how to play Silicon Valley’s game. The launch put on quite a show, even if it still seems to be working out some kinks. It’s clearly targeting international users; the interface was entirely in English and the well-curated demo video looked like a typical US tech industry product unveiling. And like so many of those, time will tell if the final version lives up to the splashy reveal.

The excitement Manus has created around AI agents will be good for the US companies that have been prophesying its future potential. But it’s a reminder that businesses and consumers are now hungry for tools that will help make their everyday lives easier, not far-off promises that this technology will be smarter than them in the future.