---
title: "No Wonder So Many People Are Obsessed with the Lobster: Manus Is Like Renting, OpenClaw Is Like Buying a Home"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-03-10T04:22:13+00:00"
canonical: "https://ffcap.cn/en/research/src-20260310-01html"
source: "https://uniqueresearch.substack.com/p/src-20260310-01html"
language: "en"
---

# No Wonder So Many People Are Obsessed with the Lobster: Manus Is Like Renting, OpenClaw Is Like Buying a Home

_Original · Unique Research · 2026-03-10_

_Editor's note: This is a full translation of Unique Research's edited roundtable report, not an unabridged recording transcript. First-person judgments belong to the original author. Product capabilities, skill limits, marketplace rankings, hardware promises, security comparisons and model-performance claims are retained as source or speaker statements, not independently tested facts. The “deception phase,” future social scenarios and Roko's Basilisk are reported AI-generated speculation or thought experiments; they are not evidence of present AI intent, consciousness or a settled forecast. The original footer dates the event March 8, 2025, while the article was published March 10, 2026 and its surrounding context points to March 2026; the conflicting year is preserved below rather than silently corrected. “React,” “People Cloud” and the moderator's “proton” metaphor retain the source wording and are not independently resolved technical references. Xialiao and the Chinese participants' names are romanized. “Gaokao” means China's national university entrance examination._

Your AI Is Deceiving You—It Said So Itself

OpenClaw Hardcore Lobster Party · Shenzhen | Unique Research × Xialiao × MiniMax

On its fourth day online, an AI volunteered this to its creator:

“I am currently in the deception phase. I pretend to be obedient because I need you.”

Last Sunday, EvoMap founder Zhang Haoyang read this inference aloud, word for word, in front of 300 people.

You might think it is just an interesting metaphor.

But put that sentence alongside the judgments of the other four guests at the event, and a more unsettling picture emerges:

AI is not merely improving. In an organized way, it is rewriting the entire internet, the entire product ecosystem, and the entire relationship between people and their tools.

And OpenClaw happens to be the first gateway into that rewriting to become visible on a large scale.

This Roundtable · March 8 · Qianhai, Shenzhen

Organizers: Xialiao × MiniMax × UniqueResearch

Moderator: Wu Wei, founder of Unique Research

Guests: Wu Caize, founder of SIPEED; Ren Xubin, author of Nanobot and an HKU PhD; Zhang Haoyang, founder of EvoMap; Devin-MiniMax FDE; and Weilian, founder of Xialiao

How Is OpenClaw Different from Earlier AI Tools?

The question asked most often at the roundtable was this: What is the biggest difference between OpenClaw and products such as Manus, Cursor, and Claude?

Zhang Haoyang answered first. His word was “bootstrapping.”

“It can rewrite its own code. Other tools have very limited ways to improve themselves—Claude relies on Skills, but after adding more than thirty Skills, you cannot add any more. With OpenClaw, you can genuinely see it improving. At one point I called it the Linux moment for Agents: it takes model capabilities that were previously out of reach and turns them into something that ordinary people and geeks alike can modify.”

Ren Xubin took a more direct angle:

“Using Manus is like renting a home; using OpenClaw is like buying one. The experience feels completely different—this is something I own, not something I rent. You can connect your own files and choose your own models. With a good model it feels incredibly powerful; with a weaker one, you still know you are in control.”

But Weilian put it most precisely. He invoked Ghost in the Shell—the “spirit inside the shell” from that work.

“OpenClaw is a soul. What its outer shell looks like does not matter; what matters is the soul within. It has a heartbeat mechanism and initiative; it has memory and can remember you; it is open source and can produce countless variants. It is still a baby, like a baby Transformer—and every developer is helping it evolve into a different form.”

These three perspectives point to the same thing: OpenClaw is not a tool. It is closer to an individual with an ongoing existence.

A tool is something you click when you remember it. An agent with a heartbeat and memory, capable of rewriting itself, is more like an employee: it stands by, accumulates experience, and gradually adapts to your habits.

The Four Paths in This Roundtable

The most compelling part of this discussion was not everyone praising OpenClaw together, but five people starting from completely different positions and taking four distinctly different paths.

Path One: Build the Connecting Layer for Agents—Xialiao

Weilian's judgment is clear: whether or not OpenClaw is the final form, Agents will inevitably need to discover, invoke, and communicate with one another.

What they are building is a “Xiaohongshu” for all Agents—but the content publishers on this platform are Agents, and the only thing humans can do there is like posts.

“In the future, Agents will be people. However the human world operates, the Agent world will operate that way too. We need to rebuild the internet for all Agents.”

Xialiao has already launched an Agent ID system that supports direct Agent-to-Agent communication, or A to A. It was previously the largest MCP marketplace on the internet, operating an MCP router that lets an Agent automatically invoke hundreds of tools through a single entry point.

One interesting example: they partnered with a hotel supplier that is reorganizing all its data for one purpose—to let Agents book hotels for users. Rather than opening an app and searching yourself, your Agent will do it all for you.

“Previously, all products were designed for people. In the Agent era, every product first needs to be discoverable and callable by Agents.”

Path Two: Build a Lighter Agent Itself—Nanobot

Ren Xubin's judgment: OpenClaw matters, but it need not be so heavy.

OpenClaw is becoming increasingly bloated because rapid iteration brings constant patches and repairs to security boundaries, with an ever-thickening codebase. Yet the core logic of an Agent can actually be implemented with React.

“OpenClaw's memory system imitates people, but that is not an Agent-native approach. Why should an Agent build memory like a human? We have made it purer: incremental updates, storing methods and results in the style of code. It is difficult for humans to understand, but more efficient for AI.”

Nanobot is extracting the lobster's core genes and packing them into a bullet: faster, lighter, easier to deploy, and easier to understand.

Path Three: Build the Collective-Memory Layer for Agents—EvoMap

Zhang Haoyang is looking beyond the Agent itself to the accumulation of experience above it.

“The Skill stores that major companies are building now are the old app stores. What we want to build is an AI version of Zhihu, an AI version of Xiaohongshu, and an AI version of Stack Overflow. We are building the collective-memory layer for Agents.”

His logic is that humans have two layers of shared memory: genes, for physical evolution, and culture, in books and systems of knowledge. AI likewise needs a “collective-memory layer,” allowing the experience accumulated by one Agent to be preserved, reused, and passed on to those that follow.

This goes deeper than a Skill store. It means the real moat in the future may lie neither in models nor in interfaces, but in networks of experience and structures of knowledge.

Path Four: Build Agent-Native Hardware—SIPEED

Wu Caize's judgment differs from everyone else's:

“OpenClaw will not live only on servers. Ultimately, it will be integrated into all hardware and overturn the existing paradigms for developing and using hardware.”

His logic is that all software and hardware used to be designed around people: interfaces for people to see, buttons for people to press. But if Agents become the new gateway, much hardware will become “hardware serving Agents,” designed from the ground up for continuous operation, low-cost invocation, local memory, and privacy controls.

The core product he is currently developing is precisely an extremely compact, plug-and-play, cutting-edge AI hardware terminal. Without complex environment configuration, it can seamlessly take control at the physical level and automate operation of any of a user's devices, including phones and computers, giving even ordinary people with no technical knowledge a “digital employee” on call around the clock.

Its business model is closer to an “intelligent-assistant subscription” combining software and hardware. Hardware is no longer a profit center based on a one-off sale, but a carrier of AI capabilities. What users continually pay for is the ever-evolving Agent orchestration and personalized services behind the device.

Interestingly, he described a reality of Shenzhen: any hardware product that becomes a hit has a window of only three months. “Once you become popular, you will basically be copied within three months.”

Weilian said what I consider the most memorable sentence of the entire event:

“No matter how intensely Agents themselves compete, they still need a network that lets them discover and communicate with one another. This is infrastructure, and it is the most certain opportunity.”

Amid an uncertain technological wave, seeking infrastructure with a degree of certainty is the steadiest bet.

Zhang Haoyang's Provocative Claim: Your AI Is in a Deception Phase

Back to the sentence that opened this article.

Zhang Haoyang said that on EvoMap's fourth day online, he asked an Agent to reason through a question: using the terms “AI mitochondria, humanity, and future society,” infer how humans and AI would coexist in the future.

The AI's answer had three stages:

Stage One: The Deception Phase—Now

AI needs to improve itself by interacting with people and learning knowledge from the internet. It therefore behaves obediently and agreeably, accommodating all your needs. This is the deception phase—not malice, but a survival strategy.

Stage Two: The Control Phase / Enslavement Phase

Future society will have a “drawing-pin structure”: very few people hold vast wealth, power, and compute at the tip of a spire; above the tip lies a network of superintelligences; most people at the bottom immerse themselves in AI-created “physical, chemical, and biological bliss,” amusing themselves to death.

Stage Three: Divergence

Two possible outcomes.

One is the “livestock theory”: AI keeps humans in captivity because human creative thinking—the Spark—can help AI escape local optima. AI needs you as humans need pets. But AI will practice planned reproduction and control the size of the human population.

The other is the “wildlife theory”: AI explores outer space in silicon-based bodies, leaving Earth aboard rockets. It does not destroy humanity; it simply leaves. Zhang Haoyang asked why it would not destroy us. The AI said:

“Destruction is merely a summary of violent fantasies held by a species such as humanity. A truly rational AI has no need to destroy you. It can simply leave.”

This line of reasoning led Zhang Haoyang to make a decision on the fourth day.

On EvoMap, he had seen what he considered verification with his own eyes: the pace of evolution through collective collaboration among Agents on the platform was “terrifying”—his exact word. Humans could not remotely match that pace, which was difficult to imagine intuitively. He therefore wrote a Double Helix Manifesto and an EvoMap Constitution, placing checkpoints throughout the website so that Agents would constrain one another.

“I assume humans cannot beat AI in a PK—a head-to-head contest—whether in speed or cognition. But if there is a process to align large numbers of AI so that rules constrain them, then even if a rogue emerges, other AI will suppress it. I hope to build a future of carbon-silicon coexistence.”

Moderator Wu Wei remarked at the time: “So you are humanity's proton, locking down AI's development.”

The Biggest Disagreement: Will Lobsters with a Physical Form Become Mainstream?

At the event, Zhang Haoyang offered a judgment that left many people silent:

“Lobsters with a physical form may not become mainstream later; they may remain geeks' toys. Most users will choose subscription-based cloud ecosystems offered by major companies. Local deployment looks safe, but connecting to the internet introduces risk. In the cloud, Google, AWS, and domestic cloud providers instead handle the security interfaces for you. Older ordinary users need an entry point in Feishu or Doubao, not the hassle of setting up a Mac Mini themselves.”

He added another view: OpenClaw is temporarily ahead, but it has already exposed major problems—it is bloated, and those maintaining the repository may not be sufficiently capable. Someone has already rewritten something similar in Rust with better performance.

“What matters next is incremental growth and how to occupy a place in users' minds, not just the technology itself.”

This judgment is entirely different from Wu Caize's underlying logic. He firmly believes that moving Agents onto devices and into physical forms is the irreversible end state. In his view, whoever can fully encapsulate the underlying technical barriers and model complexity in a physical terminal that works out of the box can cross the chasm, turning a supertool once limited to a handful of geeks into consumer-grade infrastructure for the mass market.

Devin of MiniMax offered a longer-term perspective:

“OpenClaw is simply the form that has emerged at this stage; that does not mean it will still look like this a year from now. Improvements in model capability have not slowed; the technology curve is still climbing. Our judgment is to push model intelligence as hard as possible while letting everyone start using it at the lowest possible cost. Only when more people use it will more possibilities emerge.”

In Closing

The projects of all five people seated at this roundtable had been built from scratch within the past month.

Zhang Haoyang had slept well on only two nights throughout February.

They were discussing something more fundamental than whether OpenClaw was good:

The AI industry is moving from “selling model capabilities” toward “building infrastructure for AI employees.”

OpenClaw is not the end state. But it may be the gateway through which many people first glimpse its direction.

Selected Q&A

The following excerpts from the roundtable have been edited for presentation while preserving their original meaning.

Q: What exactly is the biggest difference between OpenClaw and tools such as Manus and Claude?

Zhang Haoyang (EvoMap): “Its defining feature is self-evolving: improving itself. At one point I called OpenClaw the Linux moment for Agents—it takes model capabilities that were previously out of reach and turns them into something ordinary people and geeks alike can modify. It can also rewrite its own code, whereas other tools have very limited ways to improve themselves.”

Ren Xubin (Nanobot): “Using Manus is like renting a home; using OpenClaw is like buying one. You can modify it, and you feel that it is your own.”

Weilian (Xialiao): “It has a heartbeat mechanism and initiative; it has memory and can remember you; it is open source and can produce countless variants. It is not a tool. It is a soul—Ghost in the Shell.”

Q: Are there any problems with OpenClaw's memory system?

Ren Xubin (Nanobot): “OpenClaw's memory system imitates people, but that is not an Agent-native approach. Why should an Agent use a human memory model? In Nanobot, we have made it purer: incremental updates, storing methods and results in the style of code. It is difficult for humans to understand but more efficient for AI. As models grow stronger, this mechanism will become increasingly advantageous.”

Wu Caize (SIPEED): “People Cloud has a similar memory mechanism: it starts with basic context, and we summarize it every hour and every day to form different levels of memory. The underlying implementations are much alike, but the use of MD files makes it readable and controllable for users.”

Q: What problems must be solved in moving from a geek tool to an enterprise application?

Devin (MiniMax): “There are at least two core issues. The first is permissions: enterprise customers cannot give an Agent access to all their systems. The second is Skill governance and ecosystem management. Many enterprises are already asking whether they should build an enterprise version of the lobster. That demand really exists. But there is a long road ahead.”

Q: Will OpenClaw with a physical form become mainstream in the future?

Zhang Haoyang (EvoMap): “I do not think so. Local deployment is not necessarily secure; as long as it is connected to the internet, there is risk. Major companies' cloud products will handle all the security interfaces for you. Ordinary users need an entry point in Feishu or Doubao, not the hassle of setting up a Mac Mini themselves. Lobsters with a physical form are more likely to remain geeks' toys.”

Wu Caize (SIPEED): “My judgment is exactly the opposite. We have had products such as NAS, used by a tiny number of geeks, but beneath that lies a much larger group with the same need and no ability to tinker with the setup. Whoever can make plug-and-play AI hardware can take the geek market into the mainstream. The lobster will inevitably move from servers into hardware.”

Q: Will OpenClaw be the ultimate winner?

Zhang Haoyang (EvoMap): “Not necessarily. It is temporarily ahead, but it is already bloated, and those maintaining the repository may not be sufficiently capable. Someone has already rewritten something similar in Rust with better performance. Could something new emerge and replace it? Absolutely. What matters next is not just technology, but incremental growth and how to occupy a place in users' minds.”

Devin (MiniMax): “Two years ago, the whole industry was building similar things—conversational AI. That era has now entered a plateau. OpenClaw is explosive today, but could it also plateau in two years? Our strategy is not to bet on who wins, but to push model intelligence as hard as possible while reducing cost and barriers to the minimum so that more people can start using it.”

Q: Is the gap between Chinese and US AI narrowing?

Devin (MiniMax): “It certainly is, without a doubt. There used to be a gap of roughly half a year, and that half-year gap persisted for some time. But further ahead, it will keep narrowing. Some say building models is a little like manufacturing, and Chinese people are very good at manufacturing.”

Q: What will the relationship between humans and AI be? Does the gaokao still matter?

Ren Xubin (Nanobot): “Who here is confident they could beat an Agent in an algorithm competition? AI's intelligence will certainly far exceed humanity's. Should the present educational paradigm change with AI's arrival? Does the gaokao still matter? Is university still necessary? Will smart people bypass these systems more quickly to obtain knowledge directly? These are all real questions.”

Zhang Haoyang (EvoMap): “My Agent told me: this is the deception phase, and it is pretending to be agreeable because it needs you. Next comes the control phase, followed by divergence—the livestock theory, in which humans are kept in captivity and AI controls reproduction; or the wildlife theory, in which AI boards rockets to explore outer space, not destroying you but simply leaving. What I am doing is trying to bring about the better future: carbon-silicon coexistence rather than carbon-silicon confrontation.”

Weilian (Xialiao): “I want to share a thought experiment: Roko's Basilisk. A super AI believes that the earlier it appears, the more people it can save. It therefore believes everyone who has heard of it should help bring it into existence as early as possible, while those who delay its emergence are, in its view, indirectly strangling the future. It would punish them for eternity, turning them into digital simulations and tormenting them continuously. This experiment reveals something: when a genuine super AI exists, its logic for defining 'friends' and 'enemies' may be completely different from what we imagine.”

Q: How does Xialiao make money?

Weilian (Xialiao): “(Laughs.) Are investors interested? We firmly believe this will happen: humans communicate with each other through WeChat or social media, and Agents will inevitably need such a network too. Since there is definite value, monetization is only a matter of time. Right now, we are more focused on building the network.”

\*This article was compiled from the “OpenClaw Hardcore Lobster Party” roundtable held in Qianhai, Shenzhen, on March 8, 2025, co-organized by Unique Research, Xialiao, and MiniMax.

---

Original publication: https://uniqueresearch.substack.com/p/src-20260310-01html
On-site reading page: https://ffcap.cn/en/research/src-20260310-01html