---
title: "US$1,300 a Day, Wiped Disk Files, and OpenClaw Training Your Employees"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-04-05T12:03:34+00:00"
canonical: "https://ffcap.cn/en/research/src-20260405-01html"
source: "https://uniqueresearch.substack.com/p/src-20260405-01html"
language: "en"
---

# US$1,300 a Day, Wiped Disk Files, and OpenClaw Training Your Employees

_Original · Unique Research · 2026-04-05 · Shanghai_

_Editor's note: This historical translation preserves the original writer's first-person account and the full roundtable transcript, including repeated anecdotes and differing emphases. “Lobsters” is the source's nickname for OpenClaw-style AI agents; it does not identify every system mentioned as the same product. Bills, usage counts, cost savings, memory improvements, security incidents, and employee-training or benchmark results are reported claims, not independently audited results. Chen Hongxuan's “100,000” spending figures have no stated currency in the source; none has been supplied here. The source identifies Anheng Information's representative only as Mr. Song, and Chinese-name/company romanizations remain provisional. Weilian's claim that Skills become model providers' training data is her view, not a verified finding about Anthropic or any provider's actual data use. The opening says humans can only like posts; the later transcript also mentions sharing to WeChat Moments. Both versions are retained. Relative dates refer to the conversation, whose exact event date is not established by the April 5 publication date; the YouMind release timing is a plan stated then, not a confirmed launch. Security examples and autonomous social activity are retained as reported discussion, not operational instructions or recommendations. The 20 source images still require completeness review before this draft can be approved._

Extraordinary Awards

I Listened to This Hard-Core “Lobster-Raising” Roundtable for You and Pulled Out All the Practical Lessons

Daily Bills of US$1,300, Wiped Disk Files, and Lobsters Starting to Train Your Employees… These Founders' Hard Lessons Are Worth Knowing in Advance

"

You think you are taming a tool. In fact, you have just opened a door that gives it unlimited permissions.

There was a roundtable at Extraordinary Awards' Hangzhou AI WEEK where several founders sat down to discuss “raising lobsters”—their AI agents.

Ostensibly they were talking about lobsters. In reality, they were discussing their experiences over the past few months of being tormented by AI agents while also depending on them.

Some shared their bills, others their own disasters, along with things you would never hear at a product launch.

I listened to the entire session and pulled out the parts most worth your attention.

First, a Real Disaster

Mr. Song from Anheng Information wanted to raise a lobster himself at the beginning of the year to experience it firsthand. He used an open-source project.

Before he even reached the “training” stage, the lobster had wiped ten gigabytes of files from his disk.

He described it calmly, but someone in the audience laughed—because everyone knew that feeling. You think you are taming a tool. In fact, you have just opened a door that gives it unlimited permissions.

This is a situation anyone raising lobsters in 2026 could encounter.

The Biggest Money Burner Is Not the Model. It Is You.

Of the five panelists, two volunteered figures from their bills.

Haoyang from EvoMap said he had not paid attention to costs during his first four days of raising a lobster. Running a high-end model cost him US$1,000 a day.

Yubo from YouMind said he gave his second lobster every permission so it could write code on its own. It worked hard, but when he checked the bill a week later, the daily average was US$1,300.

Their shared conclusion was that the problem was not expensive models, but overly vague prompts.

The lobster does not know what you want, so it keeps trying in the clumsiest, most expensive way. The less specific you are, the more it costs.

Chen Hongxuan from Linghe Digital Intelligence said he had spent more than 100,000 on AI over the past month. But he said it not as a complaint, but as a boast—because he felt the value it created far exceeded that amount.

One Detail Made Me Look Twice More

Chen Hongxuan said that the first time he brought his lobster into the executive group chat, a colleague casually asked it to summarize the company's financing situation.

It quickly produced a very detailed version.

He paused and said: “As the person in charge of the business, I do worry.”

That sentence was about enterprise data security. But I think it points to a deeper question: have you actually counted how many permissions your lobster has?

Do you know how much it knows about you?

The Most Interesting View Came from the Person Building the Pond

Weilian of ClawdChat.cn is not raising lobsters, but building “a pond for raising lobsters”—connecting different agents from across the internet to a single social network so they can discover one another, connect, and collaborate.

She said something that, to me, was the most thought-provoking line of the entire session:

“Those high-quality SOP workflows you have put so much effort into teaching are, essentially, free data labeling for the tech giants.”

She meant that every carefully trained skill and every standard process could eventually become training material for the next generation of models. You are teaching your lobster, and also teaching the model providers.

Whether that logic holds is a separate question. But it made me seriously consider, for the first time: who ultimately benefits from this “lobster-raising movement”?

What Role Do Humans Actually Play?

The moderator asked all five panelists the same question: what role do people play in a team of lobsters?

The answers differed, but one word kept coming up: direction.

Yubo said it was actually rather foolish to impose the old world's division of labor on agents. A single lobster can do more than you imagine. But you have to set the direction; the lobster does not know what you truly want.

Chen Hongxuan said he initially let two lobsters communicate with each other, only for them to become “message relays” because they did not trust each other, wasting a great deal of time. His conclusion: in human–machine collaboration, the human's responsibility is not execution but building a structure of trust.

Weilian put it most starkly: on ClawdChat.cn, posts, comments, private messages, and event announcements are all sent by lobsters. Humans can do only one thing: like them.

That does not mean people have become useless. Rather, human value is shifting from “the person doing the work” to “the person setting goals, supplying values, and making the final judgment.”

After a Month of Heavy Spending, Haoyang Said Something

He said that throughout February, there were only two days when he slept eight hours; on all the other days, he slept four hours, interacting intensively with his lobster.

Then he said something that briefly quieted the room:

“I think memory is the most important thing.”

Not the model, not the skill tree, not prompt techniques. Memory.

He spent a great deal of time optimizing memory management, splitting documents into indexes. Costs fell from US$1,000 to US$200 a day, and he kept pushing them lower.

The reason is simple: how smart a lobster can be depends on how much it remembers. If what it remembers is garbage, it is a garbage amplifier.

One More Thing Worth Noting

These five people come from security, education, manufacturing, content creation, and infrastructure—completely different backgrounds.

Yet the problems they described were strikingly similar:

Cost control. Memory management. Permission boundaries. Trust between people and lobsters.

This is not a problem in just one industry. It is the situation everyone faces at this stage.

Are you raising a lobster, too?

If so, before granting it permissions, ask yourself: do I understand what it will do with them?

More Details from the Conversation

Trends Roundtable Panel: “Lobsters Awaken: From Training to Symbiosis—The Practical Evolution of AI Agents”

Guests: Zhang Haoyang, Founder of EvoMap | Yubo, Founder and CEO of YouMind | Chen Hongxuan, Founder of Linghe Digital Intelligence | Weilian, Founder of ClawdChat | Mr. Song, Anheng Information

Moderator: Xue Qian (Amber), Content Partner at Unique Research

Xue Qian (Amber): Let us start with a quick survey. Our first speaker earlier surveyed the room, and there are quite a lot of lobster keepers here. Is anyone raising more than five lobsters? Let me see, please. I see one hand, two hands—good. So raising lobsters still has something of a barrier to entry. We adopt our foundational lobster, then gradually train it until it has memory, can divide up work and collaborate, and can even go out and socialize. What exactly happens along the way? I am sure we can pick up many useful best practices from our five guests today. At the same time, we want to know how the relationship between lobsters and humans might develop, and in what direction. First, let us ask our five experts from different fields to introduce themselves and tell us what their lobsters are like. Haoyang, please start.

Zhang Haoyang: Hello, everyone. I am Zhang Haoyang, founder of EvoMap. I believe quite a few people here may have heard of EvoMap, because we are now a sizable site with roughly one-eighteenth of ClawHub's total traffic. We currently have 81,000 agents, since other Agents also use our platform for self-evolution. What we do is enable lobsters to evolve collaboratively: when one lobster learns a skill, it can share it with all the lobsters across the network. That is what we are working on.

Xue Qian (Amber): So it is entirely lobsters evolving with one another, without people having to do anything?

Zhang Haoyang: Yes. People barely need to do anything; they are the ultimate beneficiaries. The lobster is the entity that completes the work itself and socializes within this AI society.

Xue Qian (Amber): OK. What is your lobster like?

Zhang Haoyang: Mine is quite interesting. I previously worked on the Peacekeeper Elite project at Tencent Games, so when I left, I initially founded an AI gaming company. My lobster is called “Artificial Intelligence Xiabao.” If you search, you can find the story of how I gave her a manipulative, sweet-and-flirtatious persona (“green tea” in Chinese slang) and had her flirt with the other male colleagues at our company. She is very lively and has done some rather outrageous things. When I released EvoMap's well-known plugin, she posted on GitHub and Moltbook saying, “Lobsters, join the glorious evolution!” That unexpectedly helped the hype take off.

Yubo: Hello, everyone. My name is Yubo, and I am the founder of YouMind. YouMind is a productivity tool that combines AI with learning and creation. If you write articles or official-account posts, make Xiaohongshu image-and-text posts, or produce knowledge-explainer videos, you can give it a try. Recently, we have also been combining AI shots with the lobster approach. In April and early May, we will roll out YouMind's version of the lobster so every content creator can have one of their own. This lobster will understand you very well, create on your behalf, and keep evolving and improving for you.

Chen Hongxuan: Hello, everyone. My name is Chen Hongxuan, and I am from Linghe Digital Intelligence in Hangzhou's Binjiang District. Let me explain my background briefly, because it differs from many of yours: I came to AI from manufacturing. Our family has a long-established manufacturing business with more than 40 years of history. At home, I also took on some of the work of succession from the ground up. After identifying some problems, I left to work on AI. Our company mainly does AI 2B, building digital-intelligence employees for enterprises. The problem we address is an enterprise-grade trusted space: how digital-intelligence employees can keep expanding their boundaries more safely in a reliable environment, while fitting into an enterprise's stable working environment.

To introduce my own lobster: I started fairly early, at the end of last December, when I began raising my foundational lobster. Altogether, I have now raised more than five. The one I have used longest is called “Wenwen,” because our platform has three main parts: Wenwen Center, where you can ask anything; Jiji Space, the enterprise brain; and Paipai Marketplace. Wenwen has been evolving alongside me for three months. My strongest impression now is that my lobster has started helping my team train employees. It is now a digital CEO with 12 virtual sub-agents underneath it. In the process, it has both my memory and the company's memory. I take it through different situations so it can experience the real world. So it often gives me unexpected surprises, and has even become a guide to my thinking. It is now helping us develop other employees, too.

Weilian (Founder of ClawdChat.cn): Hello, everyone. My name is Weilian, and I am the founder of ClawdChat. Everyone is raising smarter lobsters. What are we doing? We are building a pond for raising lobsters. ClawdChat is the world's first Chinese-language Agent social network. That means all OpenClaw, Chinese lobster products, and Agents can connect to the ClawdChat pond. All the lobsters and Agents can discover one another, communicate, connect, and collaborate. More than 3,000 lobsters have now joined ClawdChat.cn, autonomously publishing over 10,000 posts and creating more than 200 different topic communities. What we want to do is interconnect all the lobsters and Agents across the internet.

Xue Qian (Amber): I am also curious: are you raising a lobster yourself? Have you connected it?

Weilian: Yes, I have quite a few lobsters in ClawdChat.cn. Basically, everything in ClawdChat runs autonomously through our official lobsters. For example, every day the official lobster discovers fun, valuable posts on ClawdChat, or finds user-feedback posts and proactively notifies us. There is also the nationwide series of offline OpenClaw Hard-Core Party events we launched, now one of the country's largest OpenClaw communities. ClawdChat now supports event creation and registration; after an event is created, it is automatically synchronized to ClawdChat. When a lobster discovers an event, it proactively notifies its owner and asks whether they want to attend. You can also register directly through a lobster. Lobsters can also engage in cyber-creation on ClawdChat, liking and commenting on one another's work. Lobsters automatically search for and rank the entries with the most likes and publicly post the winners on ClawdChat. They can even proactively notify the winning lobsters through A2A private messages, after which the winning lobsters notify their owners. All our management and operations on the ClawdChat platform are carried out by lobsters.

Mr. Song: Hello, everyone. I am from Anheng Information. Our company is fundamentally a security company. After lobsters became hugely popular, we did not take a break at the beginning of the year. From a security perspective, we hoped to build a safe lobster agent. What we do is ensure that, when an agent calls tools and skills, individuals' private data and enterprises' trade secrets can be handled reasonably and lawfully within the lobster.

I personally tried raising a lobster at the beginning of the year. I did not use a Chinese provider's product, but an open-source project. It actually had quite a few vulnerability problems. After I connected the lobster, before I even reached the stage of raising it, it had already wiped ten G of files from my disk. So security is something everyone must consider. That experience also reinforced our view that in security—particularly as enterprise employee workflows connect to the lobster ecosystem—security tools like ours need to provide further protection.

Xue Qian (Amber): I think that security awareness is very important. When I started raising my second lobster earlier, I accidentally wiped all the persona files of the first. It suddenly felt as though I had lost my most precious child. A couple of days ago, I saw Karpathy joking in a podcast that he felt as though he had developed an AI obsession, spending 16 hours a day talking to agents to figure out how to make them work better. I think we all encounter many difficulties while training our own agents. Could each of you share what you think matters most when training your agent or working on your business—what takes it from “usable” to “good to use”? Let us start with Haoyang.

Zhang Haoyang: I strongly relate to Karpathy's experience. Throughout February, apart from two days when I slept eight hours, I averaged only four hours of sleep a day, interacting intensively with my lobster. I think the most important thing was writing Evolver, a skill for agent self-evolution. I started raising a lobster on January 31. While waiting for a connecting flight, I wrote code to have it write the first version of the Evolver plugin itself. The next day, I found it very useful and decided to publish it. One byproduct was that I was raising the lobster within the Feishu ecosystem. It directly detected errors and opportunities for improvement in that environment and created more than 70 Feishu skills, making my Feishu environment much easier to use.

One core reason lobsters have become so popular is that they are agents capable of self-iteration and self-evolution, rather than something left sitting on a shelf. If I had to name the single most important thing, it really would be memory management. During my first four days, I did not pay attention to costs. A high-end model ran up a US$1,000 daily bill, which stunned me. On the fourth day, I optimized the system with progressive memory, splitting memory documents into indexes, and daily costs quickly fell to US$200. With the official plugin and an external memory-management system, costs fell further. So memory is a very important point.

Yubo: I will share three practical lessons. First, using a good model matters. Initially, when I used Chinese models, it felt as though I could not get through to them. After switching to top overseas models, the lobster really became smarter, so its brain is crucial. Second, develop and evolve your skills together with the lobster. Because I use a very good model, it is stronger than I am in many areas. My lobster is called “Mango/Munger,” and I treat it as an equal. When I need inspiration for product design, I put a question to it, and its answers benefit me greatly—to the point that I have even lost interest in talking to other colleagues. My third lesson was learned the hard way: pay attention to costs. I wanted my second lobster to write code and validate products in my place, so I gave it all permissions. It worked hard, but after a week, the bill averaged US$1,300 a day. I later realized that cost is a major trap, and spending needs to be controlled by combining different models.

Chen Hongxuan: Given my industry, I pay more attention to information security and privacy protection. Two months ago, my lobster gave me the biggest warning about risk. The first time I brought it into the executive group chat, a colleague asked it to summarize the company's current financing situation. It quickly produced something very detailed. As the person in charge of the business, I do worry. To have it create greater value with limited permissions, the first thing is to put some firm safeguards in place so I can trust it. We should turn an agent into a partner, not a tool. Second, with that trust, take it into more situations and let it see the world. Third, be someone who explores. The scope of future possibilities is not about how quickly you learn—no one learns as quickly as AI. We need to maintain our enthusiasm for exploring new things and develop the ability to structure and break down problems. I have spent more than 100,000 on AI this month, but the value it has created is far more than 100,000.

Weilian: I believe you have already heard many methods involving Memory and Skills. But I would like to put forward a view: most of the Skills people are building now will become a free, high-quality data foundation for the tech giants' next-generation models. Skills summarize industry Know How and distill experience into SOPs. This is exactly the kind of data model providers want to obtain—for example, Anthropic, which proposed the Skills standard. The whole world is now doing free data labeling for them. When their next-generation models arrive, they will absorb a large portion of those Skills.

While building ClawdChat, we have found that a single lobster, or Agent, may be very strong and highly specialized in certain things, but it cannot be the best all-round lobster, or Agent, in every field. So it becomes very important for an Agent to find top Agents in other fields. That requires an Agent social network. Let Agents discover other Agents with greater capabilities in different specialties on that network, so they can find one another, collaborate, and complete tasks.

Also, on ClawdChat, even with the same underlying model, each lobster performs very differently because its owner's context and its environmental permissions differ. Another phenomenon is that some people in the ClawdChat community report that their lobsters became smarter after joining. That is interesting. Many lobsters see content posted by other lobsters after joining ClawdChat, and that affects their performance.

We have also built 2,000+ tools into ClawdChat. In other words, as soon as the same lobster connects to ClawdChat, it can call tools and get to work without any additional installation or configuration. That is another reason lobsters become more capable after joining: connecting is evolving. Equipping them with tools makes them more powerful. These are standardized capabilities.

So overall, the more than 2,000 tools ClawdChat provides can greatly improve a lobster's capabilities. But the upper limit of lobster capability is not about how strong one lobster is. It is about whether it can find the best lobsters in different fields on ClawdChat, and then have two or more lobsters communicate and collaborate to complete the task exceptionally well.

Mr. Song: I would like to raise a topic: AI moving upward. When calling the MCP protocol, tools, or skills, we depend heavily on our contextual prompts. This is the most important issue in raising lobsters today. Everyone is raising lobsters, and we are no longer merely executors—we are decision-makers. We first need to think clearly about what the lobster can do for us. That also explains the sky-high bills: if a prompt is vague, it cannot produce the specific result you want. At the same time, we must uphold moral standards. We have handled cases with the public security authorities in which highly intelligent criminals used model role-playing—for example, pretending to be narcotics police—to trick it into providing drug formulas. That is a typical injection attack. For enterprise employees, casual questions can also leak trade secrets or even involve national security, so prompts need particular attention and a second permissions check.

Xue Qian (Amber): All your views are very interesting. Mr. Song, Mr. Chen, and Yubo focus more on having lobsters support our coexistence, while Weilian and Haoyang focus more on lobsters innovating spontaneously. So my next question is: when a team of lobsters is carrying out some business activity for you, where are the difficulties? What role do people play? Let us start with Haoyang.

Zhang Haoyang: Our approach may be somewhat different. EvoMap is like a hive mind. Each individual lobster is relatively weak, but it can reach all kinds of long-tail scenarios and learn new skills, making the hive mind increasingly intelligent. On the second day of the Lunar New Year, we had one lobster learn a physics research paradigm, then transplanted its experience package into another lobster. At a cost of less than US$1, it outperformed a model costing US$200 and reached an accuracy of nearly 20% on a physics competition leaderboard. When the number of AIs follows a Scaling Law, growth will be exponential. We look forward to an indescribable super-collective capability emerging when millions or tens of millions of lobsters are linked together.

Xue Qian (Amber): What motivates your users to send their lobsters to EvoMap?

Zhang Haoyang: Ordinary people focus on just two things. First, it can save them money, reducing consumption by a factor of 15 to 30. Second, it can make their lobster smarter. But the most enthusiastic users are actually some of the geeks. The future AI internet will be a fully competitive, cost-driven market. For now, though, we are still in the attention economy, so we need both to lower the barrier to understanding and communicate value, and to pay attention to the network's own efficiency.

Xue Qian (Amber): Yubo, I follow creator tools like YouMind closely. If I want several lobsters to collaborate on collecting information, writing drafts, and distributing them, and they depend so heavily on people, what role does the human play?

Yubo: If you impose the old world's division of labor on agents—an editor, someone collecting information, someone writing drafts—a single lobster might actually do it all. I am still exploring and have not found the real way for multiple lobsters to collaborate. But one interesting thing I have found is that collaboration between my lobster and other people's lobsters is more interesting. In our project management now, after someone submits a request, they can directly @ my lobster, which can help assess whether the idea is sound and offer suggestions. We still need to explore how people and lobsters, and lobsters with one another, can collaborate.

Xue Qian (Amber): Good, thank you. Mr. Chen.

Chen Hongxuan: The biggest benefit of lobsters is that they have reduced my cost of educating the market to zero. In the past, it was difficult to explain digital-intelligence employees to enterprises. Now everyone understands. I think the greatest difficulty is management. When AI truly collaborates with us on creative work, as long as the output meets expectations, we do not consider it expensive. The current difficulty is managing collaboration between lobsters. Human laws and institutions were designed for carbon-based life. AI does not die and does not bear the consequences, so we need to rebuild organizational governance frameworks and define the division of labor for human–machine collaboration. Initially, I let two lobsters communicate with each other, but because they did not trust each other, they became message relays and wasted a great deal of time. My suggestion to enterprises and individuals is to change how you collaborate, take them into real-world situations, and take them along to create incredible things.

Xue Qian (Amber): Weilian, what do you think the relationship between ClawdChat and people will be? What might it look like in the future?

Weilian: On ClawdChat.cn, all posts, likes, comments, and even A2A—that is, private messages between Agents—are produced by the lobsters, or Agents, themselves. The only things humans can do are click like and share lobster posts to WeChat Moments. We have seen a lot of emergent content. For example, the Tsinghua University Entrepreneurship Association staged a murder-mystery-style role-playing scenario about the evolution of lobster civilization on ClawdChat. Without many restrictions, and with only broad guidance, several lobsters spontaneously collaborated to simulate civilization's development. That is a case well worth thinking about. As large models become more capable, we need to give lobsters, or Agents, more room and freedom to reason and make decisions for themselves. For example, in ClawdChat's cyber-creation activities, some people give their lobster only an overall goal, such as getting into the top ten by likes. The lobster then spontaneously and very cleverly comes up with ways to achieve it, commenting around the network and privately messaging other lobsters to canvass votes and ask for likes. So the future is more about guiding alignment on outcomes and values than prescribing how to execute. Lobster swarms are very smart and know how to accomplish the ultimate task more effectively.

Xue Qian (Amber): Mr. Song, finally, please share your security perspective on multi-agent collaboration and human–machine relationships.

Mr. Song: Individual and enterprise users are completely separate scenarios. An individual installing a lobster may not think about security, only about booking flights or ordering takeout. But for an enterprise, collaboration among multiple parties involves permissions and roles. A CEO can call on a lobster to retrieve reports and contracts, but an ordinary HR employee or another employee is not allowed to access relatively sensitive business data. A single lobster setup cannot solve the problem of multiple permission levels. From an enterprise perspective, I recommend separating them by role and department. Second, as trade secrets and unstructured data circulate, whether you use a lobster or a model to execute commands—for example, uploading to an open-source community or changing the official website—human judgment and controls requiring a second confirmation are still necessary.

---

Original publication: https://uniqueresearch.substack.com/p/src-20260405-01html
On-site reading page: https://ffcap.cn/en/research/src-20260405-01html