'Mens et Manus', the story of Manus. Its brains and hands.

A story about people, taste, runtime, and what it takes to make AI actually do things.

*minor edits on Jan. 6 to better translate/ paraphrase some CN quotes to ENG

Hi all,

As we start the new year, I’ve been thinking about a version of “growing up” that can even irk you at times. The idealism sheds a layer and gets replaced with some cynicism or skepticism. But you don’t become less idealistic because you stop caring. You become less idealistic because you start noticing constraints, and you begin to understand tradeoffs.

Years ago, as my husband likes to call them, for lack of better words, “my champagne liberal days” — during university and grad school, especially. I honestly thought that things were only either right or wrong, and people were only either good or bad. I also thought what I thought was good for the world was universally accepted as good. I didn’t realize that the world actually operates in nuance, in perspective, and in incentives.

I used to think capitalists were dirty. As I’ve grown up, I’ve realized something more uncomfortable: capitalists are often the ones most forced to be realistic.

You need idealists to remind us of what’s at stake and how we can make the world a better place. That’s when you need the AI safety people to counterbalance the builders. They remind us what’s at stake, and they create the moral pressure that keeps systems from drifting too far. But capital allocators, for better or worse, live inside constraints. They’re constantly modeling what resources are available, what bottlenecks exist, and what people will actually do, not what we wish they would do.

Often in their own favor, of course. But that is also the point. They’re forced to model how the world works, why businesses thrive, how human behavior evolves, and then bet on a future that resembles their perceived likely reality.

Ideals drive progress, and ideals form the checks and balances on power. This is also why I keep coming back to the theme of intelligence versus execution. In a16z’s Techno-Optimist Manifesto, they write: “We believe intelligence is the ultimate engine of progress.” I get the instinct. It is wild to think that everything we consume, read, use, and eat began as raw materials, then got reshaped by human intelligence. But intelligence alone is never the full story. Ideas do not become reality by themselves. They need hands. They need systems. They need distribution. They need infrastructure. They need someone willing to make the tradeoffs.

So, in that way, I’ve found investors at times to be the most “real.” And it’s been an interesting learning journey to be writing for you. Since my audience is predominantly investors, it’s also opened my eyes to how to think beyond the headlines.

In a very narrowly defined way, public investors are the most capitalist of them all. I saw this sentiment echoed in Dan Wang’s year-end memo:

“One of the things I like about the finance industry is that it might be better at encouraging diverse opinions. Portfolio managers want to be right on average, but everyone is wrong three times a day before breakfast. So they relentlessly seek new information sources; consensus is rare, since there are always contrarians betting against the rest of the market.”

If anything, there is a sense of humbleness among investors I meet, this quiet, persistent “maybe I’m not always right,” beaten into them by the market. For me, checking my biases, checking my ideals, and trying to analyze businesses as they are has been an incredible brain-training experience.

And that brings me to Manus.

A company that seems to have built on the idea of helping to enhance efficiency and support real-life work, in a world that has been focused so largely on enhancing intelligence. Manus, a company with Beijing roots (back when the broader entity was still often referred to as Butterfly Effect) that very quickly found itself at the center of the most Silicon Valley of endings: a multi-billion-dollar deal with Meta.

It straddles the world of ideals. A Beijing-born company, temporarily relocated to Singapore, now acquired by Meta (largely let’s say Bay Area) embodies what Dan Wang wrote: “The two most insular cities I’ve lived in are San Francisco and Beijing… places where people are willing to risk apocalypse every day in order to reach utopia.”

That line has been stuck in my head all week because it captures why the Manus story became such a Rorschach test. Some people see “China’s next DeepSeek.” Some people see “agent hype.” Some people see “Meta is desperate.” And yes, maybe there is some truth to all of that.

But behind all of that, there’s a more useful question: where did Manus actually come from, what did it build, what constraints did it hit, and why was selling to Meta a rational move instead of a defeat?

---

First, what happened?

Meta announced it will acquire Manus to bolster advanced AI integration across its platforms, including Meta AI. Financial terms weren’t disclosed by Meta, but Reuters cited a source saying the deal values Manus at $2-3 billion, and the Wall Street Journal reported estimating the deal to also be more than $2 billion.

Meta also made the geopolitical framing unusually explicit. It said there will be no continuing Chinese ownership interests following the transaction, and that Manus will discontinue services and operations in China. VentureBeat reports that Manus cofounder and CEO Xiao Hong (who goes by “Red”) is expected to report to Meta COO Javier Olivan.

Meta has been famous or notorious, how you put it, in acquiring businesses to expand their business relevance over the years. Many times, it cannot outcompete startups in their new playbook and just offers to swallow them. WhatsApp, Instagram, and Oculus are probably the largest acquisitions in its history, and its aggressive nature was very well detailed in Sarah Frier’s book No Filter: The Inside Story of Instagram.

What is fair to say about the Manus acquisition is that this is one of the largest AI-related deals Meta has done. It’s happening alongside an aggressive AI push that includes its Scale AI investment (~ $15 billion = 49% of $29B).

There has been a lot of chatter on whether this is overvalued.

Well, for people in hardware, this kind of acquisition is hard to watch without their worldview shattering. Even if it’s US$1 billion, how many products would you have to sell to reach that? Many real-economy companies would be jealous, especially when another Chinese AI darling, UniTree’s IPO, is now being held slowed down.

For China AI, this is astronomical. But the number also seems so small compared to the headlines we are seeing in the U.S. right now, and to Meta, this amount of money is negligible. Others scoffed, saying this is the best outcome for a general-purpose agent at this point.

It is said that before this, ByteDance had reportedly offered terms around a US$30 million valuation, not higher, because it thought Manus was only worth that. Meta’s price is dozens of times higher than ByteDance’s.

---

The story of Manus (as told by the people behind it, and as you can infer from the constraints)

I recently listened to the three-hour interview with Zhang Xiaojun and Peak Ji (季逸超) after the Meta deal was announced, and revisited some of Peak’s earlier interviews. What follows is my attempt to tell the story the way it feels when you zoom out. The impressive part is not the product demos. It’s the sequence of choices under constraints, and the clarity of those choices, as opposed to building out of anxiety.

Peak Ji: “Mens et Manus” as biography

Peak Ji graduated from the High School Affiliated to Peking University, which is just down the street from where I spent 3.5 years at the High School Affiliated to Renmin University. Those schools are home to some of the brightest minds in China (not me, I was in the normal, non-academically-gifted class). But there’s a familiarity to his story: this blend of elite academic rigor and early exposure to ambition.

In Peak’s telling, his father represents the truth-chasing scientist track, a physics professor at Peking University, and his mother represents the Zhongguancun entrepreneur track: risk-taking, market-oriented, very 1990s China. Two types of ambition living in the same household created who he is today.

Mens et Manus by Brian Chan

That duality also shaped how he thinks about commercialization. He cites MIT’s slogan, “Mens et Manus,” mind and hand, as a way to describe the integration of theoretical knowledge with practical application. In the early 2020s, he met with model labs across the world too, but came away thinking the model business was saturated and that the more enduring opportunity was in “hands,” the execution layer.

Peak doesn’t sell himself as a prodigy. If anything, he describes himself as restless, tinkering, and not particularly enamored with school. In fact, much like the Bill Gates story, he was given access to the school’s computer labs and was given special permission to skip classes he didn’t like. In high school, he built software (including a browser) and joked about being an early “Chinese software 出海 (going global) attempter.”

One reflection from him stuck with me. He reminisces about an earlier software era in China when people didn’t pay, not because they didn’t value software, but because the payment infrastructure didn’t exist. No credit cards. No Alipay. No WeChat Pay. The rails weren’t there.

It’s a reminder that sometimes innovation isn’t about growth hacks or clever packaging. Timing matters. Infrastructure matters. The wider economy has to be ready. And this ties to Manus’ meteoric rise.

That also shows up in how he talks about product decisions. He keeps circling back to minimalism and constraint: they spend a lot of time thinking about what not to do. The team lives by “everything added dilutes everything else.” It’s anti-superapp, anti-feature-bloat, anti “we’ll be a platform someday.” It’s a founder forcing himself to choose.

Peak, like many AI builders, has emphasized the importance of taste, but he actually explains in lay-man terms as to why. Taste itself is a moat. SOTA, of course, still matters, but which benchmark you choose to chase determines how you go to market.

---

The company(ies) that existed before THE company

If you only encountered Manus through the invite code frenzy, you’d assume it appeared out of nowhere. But the company had already been living through its own internal evolution.

Public reporting gives us the basic skeleton. Manus is associated with Butterfly Effect, a Chinese AI product studio based in Beijing. Business Insider describes Manus as launched in March 2025 by Butterfly Effect, and notes the company moved its HQ to Singapore in mid-2025. Reuters reported that Manus raised $75M at around a $500M valuation earlier in 2025, led by Benchmark.

But the part that matters for narrative is the internal distinction the team draws between Monica and Manus, because it’s not really a product distinction. It’s a people distinction.

Peak had a metaphor that I keep replaying because it’s one of the clearest explanations of the AI app era I’ve heard. Monica was like sashimi. The quality is as good as the raw material. If your model is good, your product looks good. If your model is mid, your product is mid. There’s a purity to it, but also a dependency. You’re essentially serving the ingredient as-is.

Manus, in his telling, is the cooked seafood dish. The ingredient still matters, of course it does, but now you’re differentiating on craftsmanship: preparation, orchestration, timing, presentation. The hard part is not finding good fish. The hard part is not ruining it and then enhancing its natural flavor. Runtime, failure modes, judgment calls, taste. The things that don’t show up on a benchmark chart, but decide whether the user feels like they hired a competent remote operator, or just watched another demo.

Then Peak pivots from product into what was really on his mind: the team.

He gives a surprisingly candid glimpse into the personalities behind Manus. He keeps coming back to Xiao Hong (Red), not in the “visionary founder” way that tech media loves, but almost in the opposite way.

He says, the strength that Xiaohong has is that “Xiao Hong is really normal.” Normal as in: not performative, not ideological, not trying to cosplay as an artist-philosopher-king of AGI, which he said many in the industry tend to be, almost like artists.

Peak describes him as data-minded and intuitive, very clear-headed and rational, the kind of person who can follow common sense all the way to its uncomfortable conclusion. More importantly, in their dynamic, Xiao Hong is the final strategic decision-maker, and the team is genuinely relieved about that.

He says many in the industry are “weird” (偏执), not quirky weird, like the kind that can be extremely obsessive. They are more like “艺术家,” artists, than what you may think of as rational scientists. Often ideal-chasing but not pragmatic. In that context, “normal” becomes a moat. Not because it’s exciting, but because it keeps you from lying to yourself and actually focuses more on product-market-fit.

Peak also says something most founders won’t admit out loud: he only enjoys one part of bringing a product to market. He can do the craft, the building, the systems thinking, but there are other parts of making a company real that he doesn’t romanticize. Xiao Hong, in contrast, can do the whole loop. That completeness matters more than people think, especially in AI, where a lot of teams are either research-first and allergic to distribution realities, or distribution-first and shallow on product depth.

This is also where Monica being a cash cow matters. It’s easy to miss this in the Manus hype cycle. But Peak frames Monica as a business that throws off enough value (or at least enough stability) to let them be deliberate about Manus. They weren’t building Manus from pure desperation. They were building it from choice. That’s a different psychology. It makes it easier to walk away from seductive ideas and easier to say no.

It also explains why Peak is dismissive of what he calls the “model lottery.” If you focus on the model first, you don’t know if you’ll win, and even if you do win, you might not have a product anyone pays for. The Manus team, by contrast, sounds less ideal-driven than many peers. None of this means they don’t care about SOTA. Peak is explicit that SOTA still matters, but in a very specific way: which benchmark you choose to chase determines how you go to market. And for Manus, it was the focus on RLI, the Remote Labor Index, where Manus had repeatedly scored high. This basically means the AI completion seems less like machine-done remotely.

That’s a subtle, operator way to think about research, not SOTA as ego or religion, but SOTA as positioning and leverage. And ultimately, what is on his mind? Product market fit, he reiterates time after time in the interview.

[Pictures from Manus Team]

The team structure he described reinforces this “practical builders” vibe. Early Butterfly Effect was tiny, almost comically small, and then it grew into a proper leadership bench: CEO (Xiao Hong/Red), technical leadership (Peak), a product leader (Zhang Tao as CPO), marketing leadership with Monica DNA (Huijie as CMO, Monica cofounder), plus operations (COO). Everyone had startup experience. Early on, they ran more top-down because you have to when the company is fragile said Peak, and when talent density is still forming. As the bench got stronger and the business matured, they moved toward a more bottom-up model: more autonomy, more ownership, less “founder as command center.”

There’s also a detail that explains a lot of their discipline. In 2025, for about six months, the team focused on an AI-native browser thesis, the kind of idea that sounds inevitable if you idolize Google and think “search is the prize.” Then reality hit. A plugin takes too little market share. Even at Grammarly scale, you’re still tiny relative to the platforms you ride on. And if Monica already exists, there’s a brutal question: what does the browser give you that Monica can’t?

So they walked away.

That is, honestly, one of the most revealing things about them. A lot of teams don’t fail because they can’t build. They fail because they can’t let go of an idea or obsession, even when they instinctively know that it may not work.

---

The Launch: the invite code wasn’t a marketing scheme; it was a capacity constraint

Manus went viral in March 2025. Reuters notes it went viral on X after releasing what it claimed was the world’s first general AI agent, capable of executing tasks with less prompting than chatbots like ChatGPT and DeepSeek.

But the invite-only part took on a life of its own. I was wrong here. I overestimated that this was a genius PR hack. After it went viral, Business Insider reported that Manus access codes were listed for $1,000+ on resale sites, a sure sign that scarcity has escaped your control.

In November, when I met an employee of the company in Singapore, he explained to me that it was a capacity constraint. That actually took me by surprise. Peak later confirmed it again.

---

Strategy: why Manus refused to be “China’s Cursor.”

Peak is explicit about what the team does not want to build. He says there are plenty of people doing that - trying to become China’s Cursor. They wanted to work on something new, and the team wanted to serve not just everyone, but prosumers.

That decision has two implications:

• Distribution goal: Western prosumers are tastemakers. The Western launch timing matters. Going viral in China was not the point. Top-down adoption (names like Karpathy, Jack Dorsey) mattered more than buying domestic coverage.

• Monetization goal: prosumers pay. They are not trying to extract maximum mass-market usage from day one. They’re trying to earn a high willingness-to-pay user and use that to refine the product.

Underneath it all is Peak’s skepticism of the model lottery. He’s not anti-model. He likes good models (DeepSeek, Qwen, Kimi). But he doesn’t want Manus to be a model company, because in his mind, you don’t control the odds.

So what’s the moat? In the interviews, Peak keeps returning to something that sounds soft until you’ve built product: taste. Taste in what to chase. Taste in what to ignore. Taste in which benchmark to optimize for, because benchmarks shape go-to-market. That’s why “everything added dilutes everything else” matters. It’s a taste of the product as a defense.

He said many think AI will replace human labor, but at Manus, they don’t build that way. A model should not be equated with human intelligence. However, agents can complete tasks that humans can do. Why make a general-purpose agent rather than an agent OS? Because they never designed Manus to think about replacing humans. They thought about creating a tool that can relieve a pair of hands, a tool that was useful for, say, new media practitioners.

As a useful outside framing, Dev Shah of Resemble AI argued on X that Meta didn’t buy a model company here. It bought an “environment” company, meaning the orchestration and execution layer where models can actually act

---

Benchmark, Singapore, and “re-coupling?”

When Benchmark invested in Manus, Kevin Xu wrote: “Will Benchmark’s reputation be tarnished, unfairly or not, by this deal, or will its enduring brand as one of the best early-stage VCs prevail?” He tied the deal to the geopolitical backdrop. This acquisition seems to have shrugged that off and become a symbolic start to a future China-U.S. re-coupling in AI. If conscious decoupling was what defined 2025, then 2026 may bring more of the cooperation that was happening under the table out into the open.

When the company took on Benchmark money, it was the first Chinese AI company that Benchmark had invested in. And the move from Beijing to Singapore, though it might look like it aligned with the broader “going global” trend, was framed internally as a decision made out of necessity.

End of the day, the more and more I’m realizing is that “China-shedding” cannot really be a strategy if your operations and dominant market are still in China. But much like what Peak said himself, they had their eyes outside of China on day one.

A third-party Chinese commentator summarized the logic like this: Xiao Hong’s decision to relocate to Singapore reads less like “go global” branding and more like a constraints-driven necessity. Once U.S.-China tensions hardened into concrete AI controls, especially around access to advanced chips, any China-linked AI company that wanted to stay near the frontier had to operate with a permanent ceiling. Manus had also taken U.S. investor money, which made compliance and scrutiny even more acute. Singapore became the escape valve: politically safer, commercially credible to Western partners, and practically positioned as a bridge between China and the U.S. After the move, the company leaned into that repositioning, pulling back from China-facing social accounts and orienting product availability toward the English-language internet. It was a deliberate shift to being offshore, rather than China-first.

There’s also something very sobering, almost investor-like, about how Peak talks about progress now. This is not a cheap startup to run. An agent company that actually executes in the real world isn’t just “software margins.” It’s infrastructure, inference, sandboxes, tool calls, retries, long runtimes, and the invisible tax of reliability. In their telling, they burn an eye-watering amount every day, hundreds of thousands of USD, because that’s what it costs to keep the kitchen running while you figure out how to cook the dish consistently.

And yet they’re not apologetic about it. If anything, the tone is: yes, it’s expensive, but it’s finally working. Demand didn’t have to be manufactured. The product proved itself. The market, especially the English-speaking, prosumer-heavy market they were aiming for, validated that people will pay when software stops being a chatbot and starts feeling like competent remote labor.

What’s interesting is how that success is framed internally. Peak jokes that the team has finally moved from CBA to NBA, from being a domestic player to being taken seriously on the international court. It’s funny, but also telling. This is the first time they feel globally legible.

Then comes the second layer of realism. Even with the headline number of $100M ARR, they don’t treat it like a coronation. They treat it like making the league. A big milestone, yes. But once you’re in the NBA, you’re surrounded by giants, freaks of nature if you must, the GOATS of the GOATS, and teams with structural advantages you can’t copy-paste. On a global scale, they still see themselves as another average performer in the league. Not in capability, but in position: one of many teams that have earned the right to play, and now have to prove it every day. And joining Meta now, is joining the big dogs.

---

What they’re building now (and what they’re not building)

One thing that comes through clearly in Peak’s framing is that Manus isn’t trying to win the “smartest model” contest in the way the industry usually defines it. They’re not optimizing for RLVR-style math gymnastics or benchmark heroics for their own sake. The center of gravity is much more mundane and much harder: make an agent that endures reality and relieves a pair of human hands.

That means their north star is context, continuity - what he even said is like continuous education in some ways - and execution:

• Context awareness and “compassion” awareness. Not compassion as anthropomorphic fluff, but in the practical sense: the agent should understand the user’s intent, constraints, and patience level. When to ask, when to proceed, when to stop, when to confirm.

• A larger, more usable context window, not as a spec, but as working memory. The point isn’t “we have X tokens.” It’s whether the agent can hold the thread of a messy real-world project, retrieve the right details, and avoid drifting over time.

• Real-world task completion with tool-integrated reasoning. Better offloading to tools (browsers, code, file manipulation), better retrieval mid-flight, and reasoning grounded in what tools actually return, not what the model thinks is true.

• Interactive agency, not fire-and-forget. Peak’s critique is that too many agents create a brittle loop where the user waits for the agent and the agent waits for the user. Manus wants a collaborative workflow where the user can step in, steer, correct, and unblock mid-execution without restarting everything.

• Error resilience as a first-class objective. Models will make mistakes. Tools will fail. Web pages will change. Credentials will expire. The question is whether the system can detect failure, recover, route around it, and keep going.

---

So why sell to Meta?

The realist answer is that the next constraints were no longer fixable with just the team’s persistence and taste. The reality is, it is an expensive game, and only the big tech companies right now have the resources to burn.

Selling isn’t automatically “giving up,” especially if the business is working. AP reports Manus crossed $100M in ARR within eight months of launch. Still, the pattern across sources is consistent enough to say Manus was not dead. It had traction, momentum, and the beginnings of a real subscription business.

So why sell? Because once you accept the constraints lens, the answer becomes structurally clear.

• Agents are compute businesses pretending to be software businesses. Agents run longer, touch more tools, and break in more ways. Every extra minute of runtime isn’t just a cost, it’s also a reliability exposure. The Manus story already showed what happens when demand outruns infrastructure: shortages, gating, and an access bottleneck that becomes part of the narrative. Money buys time, not inevitability. The strategic question becomes: who has a structural advantage in serving long-horizon, action-taking systems at scale? Meta does.

• Distribution isn’t a feature, it’s the product. Agents don’t win by being impressive once. They win by becoming a habit. Habits form where people already live: messaging, feeds, creator workflows, SMB customer interactions. Reuters offers a clean “why Meta” angle here, pointing to Meta’s WhatsApp SMB footprint. If you’re Manus, you can spend five years trying to earn distribution, or you can plug into it. Even though the goal is to sell to prosumers, and as Peak said proudly, the product blew up because the likes of Jack Dorsey and Andrej Karpathy tweeted about it. At the end of the day, the service is a steep price (charged by token usage, not monthly subscription) and still rather a niche product. Meta will provide all the distribution and feedback loops at scale, which is what agents need to climb the reliability curve.

• The geopolitics tax was becoming permanent. Immediately, there has been coverage saying the Chinese regulators may not be happy about this deal, nor were they that happy about Benchmark’s capital injection. Kevin Xu and Paul Triolo wrote about the geopolitical angle. Scutiny is almost guaranteed for anything with Chinese roots and “AI” in the headline, and when Manus moved its HQ from China to Singapore amid U.S.-China tensions, that was first seen as “China shedding”. And then, with this deal, Meta said there will be no continuing Chinese ownership interests post-transaction; Manus will discontinue operations in China; and Meta will wind down its remaining China operations, including shutting down Monica.cn and relocating relevant employees. “Go global” wasn’t just branding. It was a prerequisite to first receive Benchmark money, and then what were they supposed to do from there?

• The simplest reason is also the most founder-ish: becoming the big company. The interview with Zhang Xiaojun was dense and interesting, but one line Peak said stood out to me as someone who covers big tech day in and day out and often asks how startups challenge incumbents. He said: “小公司怎么超过大厂？就是成为大厂.” (How does a startup beat a big tech? Become one) Joining Meta is one way to do it. If your ambition is to overtake Google, you can’t do it as a 100-person company forever. Not because you lack intelligence, but because you lack distribution, infrastructure, political cover, money, talent, and time. If the product is hands, then the contest is who gets to attach those hands to billions of users. Meta can. Manus can’t, unless it becomes part of Meta.

---

Where this leaves us

Much of the internet wants this to be a story about which model is smarter. But I think the story is about what we talked about in the very beginning. Capital recognizes constraints and uses resources to push ahead with technological progress. We also tie it back with idealism vs pragmatism. This is a story of leveraging the Mens to scale the Manus in Mens et. Manus.

Meta buying Manus is a constraints-aware move: keep investing in the brain race, but also buy the thing that makes brains useful - the hands. And Manus selling to Meta is also constraints-aware: keep scaling a compute-heavy, geopolitically sensitive execution product alone, or fuse yourself to the biggest distribution engine that still has something to prove in AI.

---

References / relevant reads:

1/ 张小珺商业访谈录 128. Manus决定出售前最后的访谈

2/ 商业迷：manus 肖弘的经历带来的启示…

3/ Paul Triolo:

https://pstaidecrypted.substack.com/p/meta-manus-acquisition-whats-behind

4/ Kevin Xu: Benchmark Invests in Manus: VC Math vs Geopolitics Math

Benchmark Invests in Manus: VC Math vs Geopolitics Math

5/ Tech-Optimist Manifesto: https://a16z.com/the-techno-optimist-manifesto/

6/ Dev Shah: https://x.com/0xDevShah

7/ Dan Wang’s 2025 annual letter: https://danwang.co/2025-letter/

8/ https://www.reuters.com/world/china/meta-acquire-chinese-startup-manus-boost-advanced-ai-features-2025-12-29/

9/ https://www.businessinsider.com/meta-acquires-ai-startup-manus-expand-general-purpose-ai-agents-2025-12

10/ https://www.businessinsider.com/what-is-manus-ai-meta-acquisition-chinese-startup-singapore-agent-2025-12

---