Why Manus Matters

The American AI Diffusion Challenge

I appeared on ChinaTalk to discuss the topic of this week’s essay (and related topics) with the enduringly excellent Rohit Krishnan, swyx, and of course, Jordan Schneider.

Introduction

Last Thursday, the Chinese AI startup Monica released an agent called Manus. In a demo video, co-founder Yichao “Peak” Ji described the system as “the first general AI agent,” capable of doing a wide range of tasks using a computer like a human would. While numerous startups, as well as OpenAI and Anthropic, have released general computer-using agents, the company itself claims superior performance to those products. Many reports from social media seem to agree, though there are notable exceptions.

Manus is not available to the general public as of this writing. Monica has given access to a select group of users—seemingly focused on high-profile influencers. I was not offered an access code, but I was able to use the system for a couple of prompts. My tentative conclusion from that experience—as well as the uses I have seen from others—is that Manus is the best general-purpose computer use agent I have ever tried, though it still suffers from glitchiness, unpredictability, and other problems.

Some have speculated that Manus represents another “DeepSeek moment,” where a Chinese startup is surprisingly competitive with top-tier American offerings. I suspect this analogy confuses more than it clarifies. DeepSeek is a genuine frontier AI lab. They are on a quest to build AGI in the near term, have a deep philosophical conviction about the power of deep learning, and are staffed with a team of what Anthropic Co-Founder Jack Clark has called “unfathomable geniuses.”

While I have respect for what Monica has done with Manus, I do not think any of these things are true for them. They are, according to CEO and co-founder Xiao Hong, a product focused company, not a research lab. Unlike DeepSeek, Monica has not been releasing intriguing papers for the past year. Nearly every AI policy researcher I know had heard of DeepSeek well before that company’s r1 and v3 models went viral; almost no one I know had heard of Monica or Manus before last week.

On the other hand, DeepSeek’s v3 and r1 models at best matched the capabilities of competing American models (in my opinion, o1 is usually superior to r1, and o1 pro is unambiguously superior). Manus is legitimately better than competing American computer use agents I have seen (though like all such products, it is deeply flawed). That, too, is a discontinuity with DeepSeek.

Monica did not train their own models to create Manus, nor do they appear to have made any deep technical breakthroughs. Instead, Manus appears to be a multi-agent system composed of Claude 3.6 (née 3.5.1/3.5 new) Sonnet and Alibaba’s Qwen line of models. They are using open-source scaffolding, prompt engineering, and other tricks common in AI agent development.

And yet, for the two tasks I gave the system (a limited sample size, to be sure), Manus outperformed Anthropic’s Computer Use feature and OpenAI’s Operator. Given that they are using an old version of Claude, it seems quite likely that American companies either do have or easily could have similar capabilities behind closed doors. The interesting question, then, is simply “why did an unheard-of Chinese startup ship a better product than any American company I am aware of?”

The answer to this question gets to the heart of many issues I write about on Hyperdimensional: the interaction of AI with the existing American legal system, the looming prospect of new AI regulations, and the intellectual culture surrounding AI in the West. These factors combine to make dynamism in American AI less likely than it could otherwise be. Some of these factors are driven by good underlying reasons; others are more questionable. Some of these factors can be alleviated; others are probably unsolvable.

One note: while Manus is made by Monica as far as I can tell, nothing on Manus’ branding page refers back to Monica. Monica makes a host of separate AI products that I have not used. I am unsure if they are undergoing a rebrand or restructuring, or if they separated the brands for some other reason. I will follow ChinaTalk in referring to the company as Monica and the system as Manus.

With that said, let’s dive in.

What is Manus?

Manus is an agent designed to perform arbitrary tasks on the internet using a computer the way a human would. Like OpenAI’s Operator, it instantiates a virtual computer with a browser, but unlike Operator, it adds a coding environment and other tools. Then, it employs LLMs such as Anthropic’s Claude and Alibaba’s Qwen to use those tools to accomplish practical tasks (the use of Claude might in fact be just as significant a piece of news as the product release, since Anthropic disallows and ostensibly monitors for usage of Claude in China—but that won’t be my focus here).

It is not clear to me precisely what models did what tasks in the version of Manus I used. One of the things that seems obvious, however, is that none of the underlying LLMs it currently uses are smart enough to rival OpenAI’s o3 and their much-loved agent product Deep Research. I would pick Deep Research over Manus any day for the kinds of agentic research assistant tasks the latter system is designed to handle.

But Deep Research cannot book travel for me, or shop for a new pair of shoes. Tasks like that are within scope for OpenAI’s Operator product, however. Operator is based on a specialized version of GPT-4o, and in general I have found it to be unreliable. For example, the system consistently struggled to book travel on Amtrak’s website. Manus was able to do this on its first try, though I should note that I did not complete the transaction for security reasons (I have no reason to trust Monica, and under very few circumstances would I route my payment information through an agent I do not trust).

Manus still suffers from glitches, like every other general computer use agent I have tried. I personally saw it lose track on one task, and stumble into an infinite loop in another (extremely common agent problems). I cannot tell if these are glitches caused by the models themselves, in the agentic scaffolding, or simply due to compute constraints on Monica’s part, but I would guess it’s all three. I’ve never seen a general-purpose computer use agent I would call “good,” and Manus does not change that. But it is the best one I have used.

There is no magic here, no deep technical insight or feat—at least not one I can see. Manus took existing approaches to agent development, combined them with off-the-shelf (though fine-tuned, in the case of Qwen) LLMs, and shipped a product with a well-executed marketing campaign. Many in the AI community seem to turn their noses up at such things, but I struggle to understand why. The snobbery misses a fundamental point: existing LLMs were good enough to achieve Manus’ capabilities, but until last week no one I’m aware of had, in fact, done it publicly.

Manus is not a technology innovation story. It is a technology diffusion story.

Is there an Emerging Chinese Edge in AI Diffusion?

In his excellent paper and book, George Washington University scholar Jeffrey Ding makes a subtle but compelling argument for how technology leads to prosperity. He argues that innovation is essential, but not enough. As I have written many times, the existence of an innovation is only the beginning; technology changes the world by being integrated into products and services that improve productivity or otherwise perform useful work. With novel technologies—especially general-purpose technologies—the work of integration is unglamorous, nontrivial, and utterly essential for realizing the benefits of innovation.

Ding describes how, in the 19^th century, Europe led the world in scientific innovation. The best research universities in the world were there. Aspiring researchers from all over the world—especially America—would travel to Europe to learn in their top-tier research institutions. But Europe did not use their innovations to their maximum advantage; instead, that honor went to America—the hungry, less sophisticated upstart. Europe led in innovation, but America led in diffusion (and eventually innovation also, but this took time). America thus reaped just as many, if not more, of the economic benefits of new inventions—even if those inventions were not primarily made on American soil.

This pattern has repeated itself with the United States and China today—except this time, America is in Europe’s position. The United States led much of the innovation on lithium-ion batteries, solar panels, and electric cars, yet it is China that has exploited those technologies to their fullest advantage. Manus is evidence that something similar could happen in AI.

It's not the only evidence either. Compared to the US, China has focused much more heavily on open-source frontier AI, which I and others have argued benefits technology diffusion. Manus itself relies heavily on open-source tools and models (with the exception of Claude). Polls indicate that Chinese consumers are much more excited about AI than Americans, who in general seem much warier of the technology. The Chinese government has made accelerating practical, industrial applications of AI a major priority.

Why does the US risk lagging in AI diffusion?

The American Diffusion Challenge

To understand this, it is best to start with what Manus does not appear to have much of: guardrails. I am aware of no public comments by either founder on matters of AI safety. The company has not released any public documentation about what, if any, safety mitigations they have performed (for example, robustness tests against prompt injections, which could cause the system to be tricked by a malicious actor into divulging sensitive user data). You can rest assured there is no safety and security framework to be found.

On balance, that is probably a good thing. There are legitimate security issues with agentic systems like this, but we are unlikely to make progress purely by thinking about them. We must take the risk of real-world deployments to understand the actual, as opposed to theoretical, security vulnerabilities, harms, etc. these systems can create. In the spirit of OpenAI’s philosophy of iterative deployment, I applaud the Manus team for having the chutzpah to simply ship.

Of course, OpenAI originated the iterative deployment approach in frontier AI. But Monica is a heretofore unknown startup with nothing to lose; their risk calculus is undoubtedly different. OpenAI almost surely is more risk averse, not just to protect their own reputation but because they understand themselves, rightfully, as a kind of ambassador for AI. If they err on a major product release, the public’s perception of AI as a whole might be irreversibly harmed. I am sure at least some of their staff feel that burden each day, and for what it’s worth, OpenAI and similarly positioned frontier labs seem to be walking the tightrope reasonably well.

What about US startups? Here the picture is somewhat more complicated. Early agentic product efforts (pejoratively referred to as “wrappers,” since they often “wrap” frontier labs’ foundation models in software scaffolding) tended to perform poorly. The ambitious efforts to create generalist agents (such as AutoGPT) failed because the underlying models were not capable enough (this is still arguably true, though less so today). More narrow efforts, such as PDF analysis tools, were eventually subsumed by frontier labs adding native capabilities to their systems, or simply by the models improving with scale. Sam Altman famously said that such efforts would be “steamroll[ed]” by OpenAI.

It is fair to wonder whether the West’s obsession with “AGI” has distracted the AI industry—investors, entrepreneurs, and the frontier labs—from doing everything they can to extract the maximum value from existing LLMs. Why bother putting all this work into making today’s LLMs do something when next year or the year after we’ll have systems that can do everything?

Maybe this is the right intuition to have. I think it is very likely that products like Manus will be steamrolled by future agents from frontier labs. Perhaps a product like Manus is great for getting short-term social media hype, but a poor business strategy. Nonetheless, if one is trying to understand why Manus came from a Chinese rather than an American startup, this seems like one plausible and partial explanation.

The fact that the default reaction among some prominent AI observers has been to dismiss Manus as a mere “Claude wrapper” with a fancy marketing campaign speaks volumes. Rather than wonder how this low-hanging fruit had not been plucked earlier by Americans, too many are inclined to discount Manus simply because it does not fit into their existing assumptions about what “AI progress” looks like—and that progress tends to be about innovation rather than diffusion. The AGI obsession might be correct, but it is probably negative, at the margin, for near-term AI diffusion.

While one can imagine some ways to make progress on these barriers, they exist in many ways for good, or at least unavoidable, reasons. As China’s own sophistication and societal awareness of advanced AI grows, I expect them to encounter similar challenges. These are problems that are not entirely unique to the West, though I suspect we worsen them in various ways.

Finally, though, there is public policy. Each of the use cases Manus shows in their demo video is in heavily regulated sectors (employment, real estate, and finance). Under the algorithmic discrimination laws I have covered extensively, all these use cases shown are plausibly regulated, meaning they would require the user of the system to write an algorithmic impact assessment and a risk management plan and be subject to potential liability. Furthermore, under these laws, Monica itself would be considered both a developer and a deployer of high-risk artificial intelligence systems, facing extensive transparency and paperwork requirements (as well as liability). None of these laws are in effect yet, but even having them on the table in more than 15 states tells you quite a bit about the posture of the American policymaking apparatus with respect to AI.

It's not just potential laws, either. AI has been the subject of extensive regulatory interest at the state and federal level. State attorneys general issue “guidance” documents regularly to remind economic actors that they are paying close attention to AI. They have brought major investigations against AI companies, some of which have resulted in large settlements. The same is true of federal regulators.

Over the past decade, the technology industry has faced an unprecedented onslaught of regulation and threats of regulation on a much broader range of issues: social media algorithm “design” laws, child safety laws, app store laws, data privacy laws. After all these laws, and all the heated, bipartisan rhetoric, committee hearings, guidance memos, investigations, enforcement actions, should anyone be surprised that the technology industry as a whole seems more risk averse than it used to? Wasn’t the death of “move fast and break things” the implicit, and sometimes explicit, goal of this whole crusade?

And you will note I have not mentioned the European Union.

Whatever you think of the merits of any particular policy issue I’ve raised above, the directional thrust of the past decade of Western technology policy has been overt and relentless hostility to technology firms of all sizes. The onslaught has been especially harsh in consumer technology, which is likelier to be tangible and visible (through media) to policymakers. When you encourage society to scapegoat technology for all of its problems, do not be surprised when technologists respond by putting forth fewer solutions.

Remember that the biggest new consumer technology hit of the decade was ChatGPT—a system released as a research preview whose consumer market success was an accident. An accident that led to numerous government investigations, new proposed laws, countless intellectual property lawsuits, and, incidentally, my own career shift—to AI policy.

If you attempt, or even stumble into, having a day-to-day impact on the average American’s life, our policymaking apparatus has made one thing clear: they will fight you tooth and nail.

Many factors have driven technology to the less risky posture of preferring business-to-business, and business-to-government, sales. But it should come as no shock that all this policy action adds up to having a real-world impact.

How big of a real-world impact? I cannot say. It seems implausible, though, to argue that the effect is zero. It has always been fiendishly hard to make consumer products, so causal factors that are small in absolute size can still have large effects.

Of all the potential causes I have discussed above, this is the one that we are best-positioned to alleviate, if not fix. This is an own-goal, and we should shift to a more friendly posture toward technology. It is worth noting that China, too, has been engaging in a crackdown on the software industry, though it is also worth noting that they seem to be reversing it.

There is cause to be optimistic here, but America has a long way to go. Real action, rather than mere words, will be required. A good starting place might be a proactive, innovation-friendly, but realistic approach to AI liability—but there are many more options one can imagine.

Conclusion

Manus is not, in the grand scheme of things, that big of a deal. A Chinese startup marginally advanced the agentic capabilities of existing LLMs through clever product engineering. That’s it. I suspect we’ll have all forgotten it before too long. It probably will not have the staying power of DeepSeek. Deep learning wizards pushing back the veil of ignorance on the nature of intelligence is just sexier than good product engineering and craftsmanship, particularly in the still-immature category of computer use agents.

But the history of technology teaches us that the unsexy work of diffusion should not be mistaken as unimportant. I noticed a year ago that I was not seeing a greater diversity of compelling consumer AI use cases out of Silicon Valley—but diffusion takes time, I figured. A year later, I am starting to feel more concern. It seems to me that we have barely scratched the surface of interesting, unexpected uses of LLMs and other contemporary AI. And I find myself wondering: where is the Cambrian explosion I expected? Where is the American innovation ecosystem? Why is the AI transformation thus far so colorless compared to previous technology transitions? Where is the joy, the Steve Jobsian creativity? Have we somehow squashed it?

Maybe it will all come in due time. Or maybe this time, the diffusion will come from a hungrier, less sophisticated upstart.