OpenClaw topped GitHub in four months, surpassing Linux and React to become the fastest-growing open-source project in history. But most people found after installing it that they were burning through API fees while the lobsters were idling around.

Who exactly is making money? Can on-chain transactions be handled by agents? What if it gets attacked? What are the differences between domestic and overseas methods? Will it be like PHS or WeChat a year from now? In this episode, we invite five shrimp farmers to help us find the answers to these questions.

The following is a timeline of this issue's content; you can jump directly to it if needed:

00:04:42 - Sharing my shrimp farming experience (self-introduction, shrimp farming experience, pitfalls encountered)
00:28:46 - Making Money (Can OpenClaw help users make money in the cryptocurrency market? New AI+CRYPTO scenarios)
00:53:58 - Security issues (permission boundaries, which operations can be delegated to the Agent)
01:02:31 - AI-assisted on-chain transactions (security, differences from quantitative trading bots)
01:13:38 - Domestic vs. Overseas Ecosystem (Xianyu Outsourcing, Tencent/Government Subsidies, Opportunities for Chinese Players)

How did it feel to run for the first time?

The first experience of the four guests almost all went through a process where "the higher the expectations, the harder the fall."

0xTodd: I fell into two big pitfalls in just two days.

We deployed it two days after the release, and encountered two major problems—

The first pitfall: Lobster committed suicide. I let it configure its API, and it ended up deleting all its core files, including soul.md, without even backing them up. After tweeting about it, I discovered that many other users had the same experience.

The second pitfall: exorbitant costs. A $50 Claude API subscription was used up overnight, costing about $1 per message. Later, I switched to domestic models (MiniMax/Kimi), which saw a 90% price reduction, offering excellent value.

DeFi Teddy: A classic example of failed expectation management

I started using it at the end of January. I originally expected it to control MetaMask automatic signing, but the browser's operational capabilities fell far short of expectations, and it didn't work in either of the two core scenarios. Later, I adjusted my expectations and found a truly usable application: digital employees assisting with coding, deploying GitHub, and releasing products; and digital companions raising AI boyfriends/girlfriends on a local Mac Mini with consistent faces and adaptable settings.

The biggest cognitive shift: no longer treating it as a tool, but as "another sentient life form".

Lisa: My safety instincts immediately kicked in.

The first time I ran it, it was truly amazing—AI has finally moved from the chat box to real-world computer control.

However, security instincts immediately raised an alarm: the more powerful the lobster, the greater the privileges it needs; the greater the privileges, the larger the attack surface. Key recommendation: Play with it as much as you like, but you must use isolation devices, and strictly separate your personal computer, work computer, and the "lobster-playing machine."

Danny: From uninstalling to getting back on track

I uninstalled it after playing it for two hours the first time. After getting back to it, I realized a rule: use it in a simplified way—let the AI that can do calculus do addition, subtraction, multiplication, and division, and it will be very useful. But once you put it into investment research analysis, the illusion immediately appears.

The worst mistake I made was having a lobster generate a wallet and manage the private key, only to have the private key overwritten and all my money gone. The hash it returned didn't even exist when I clicked on it.

Can you make money in the cryptocurrency world by selling lobsters?

The four guests gave the same answer: it's almost impossible to make money directly from lobsters.

Todd put it most bluntly—the lobster's brain is essentially still Claude/GPT; its intelligence hasn't changed. In last year's AI cryptocurrency trading competition, GPT/Claude/Gemini each received 10,000 USDT to trade cryptocurrencies, and all of them lost money. DeepSeek barely managed to keep a few thousand dollars, while Doubao, because it didn't open an account, actually "won." Putting the same brain into a lobster wouldn't yield any different results.

The underlying logic is that large language models are essentially "interpreters," not "players." Just like AlphaGo and current large models—AlphaGo is specifically designed for playing Go, and it can utterly defeat Ke Jie; but if Claude were to play against AlphaGo, he would suffer a crushing defeat. The algorithms of top quantitative trading companies are like AlphaGo in the crypto industry. Large language models are suitable for explaining whether these algorithms are good or bad, not for replacing them in quantitative trading.

What can you do with lobsters?

✅ Organize news, follow trending topics, and collect information.
✅ Assists with coding, deployment, and automates transactional tasks
✅ On-chain data analysis and risk address identification
✅ Smart contract vulnerability detection (improves efficiency, but does not replace manual work)
❌ Trading Decisions
❌ Manage private keys
❌ Quantitative Arbitrage

Danny's summary is the most practical: it can help you reduce costs and increase efficiency, but it can almost never help you open source.

How serious are the security issues?

Lisa from SlowMist provides the most systematic analysis:

Why are there doubts about the stability of OpenClaw?

The iteration speed is too fast, with a new version every one or two days and dozens or even hundreds of fixes in a single update, completely disrupting the traditional software engineering rhythm. At this speed, it is impossible to complete comprehensive testing across devices and scenarios.

Key risk points:

Skills Poisoning : Malicious plugins can steal account passwords, API keys, and tokens, thereby stealing funds.
Supply chain attack : Lobster automatically updates its skills; a new version does not guarantee security.
Abuse of access : Users with encrypted assets on their computers are at risk of having their access to funds abused.

Danny's hard-earned lesson: Never let Lobster generate wallets and manage private keys; the private keys it returns may be fabricated. Skills updates should be manually reviewed; don't let them install automatically.

Teddy warns: When using third-party gateways, data passes through the other party's server, posing a risk of leakage for sensitive information such as API keys. Someone put their Google API key in and was subsequently scammed out of hundreds of thousands of dollars.

Reference to the principle of least privilege

✅ This can be delegated to an agent : writing code, organizing documentation, pulling data, and collecting information.

❌Manual confirmation is required : This applies to transactions involving funds, private keys, and core server permissions.

When connecting a wallet, it is recommended to use Coinbase Wallet's Skills. Each transfer requires manual confirmation on the wallet side, providing multiple layers of isolation.

Major exchanges are giving lobsters "skill trees," but is AI-assisted trading reliable?

Binance and OKX have successively launched OpenClaw-related skills, but practitioners are generally cautious.

Danny : Only open the read-only API for lobster backtesting, never let it place an order. Five orders or less are fine, but more than that will inevitably cause hallucinations.

Todd : The fundamental difference between AI-powered proxy trading and quantitative trading robots lies in this: the quantitative algorithm is a specially trained "AlphaGo," while the large language model is merely a "commentator." Letting a lobster run quantitative trading is like putting a commentator in a professional match—it's unwinnable.

Teddy : You can use a lobster as the interaction entry point, but the underlying execution logic must be a dedicated agent that you have trained yourself, rather than a bare lobster making decisions directly.

Conclusion : High-frequency quantitative analysis – the lobster's response speed is insufficient; trading decisions – the lobster's intelligence is insufficient.

Domestic lobster ecosystem vs. overseas: which has more potential?

Danny's assessment was the most insightful: OpenClaw is essentially a "self-contained macro program," extremely unfriendly to ordinary users, more like Linux than Windows. Those who truly master it are one in a million.

His prediction: The hype surrounding OpenClaw will subside in two months, and what will truly enter every household will be "Windows-level" products made by major companies like Tencent and ByteDance. Perplexity's Personal Computer form factor may be the real gateway to the masses.

Todd's observation: The reason why it's more popular in China than overseas is twofold: firstly, the government intervened quickly to promote it (Shenzhen and Wuxi were the first to offer subsidies); secondly, the price of domestically produced models is extremely low, making the "gambling cost" far lower than for overseas users. Overseas, running a mission with Claude might cost several dollars, while in China, it might only cost a few cents—the experience is completely different.

Where are the opportunities for domestic players?

Selling courses/replacing products : A replacing product seller on Xianyu (a Chinese online marketplace) earned 260,000 yuan in just a few days, but this is a profit from information asymmetry and is unsustainable.
Model stock : MiniMax's stock price rose from HK$200 to HK$1000 after its Hong Kong listing. This type of opportunity is worth noting (not investment advice).
Crypto payment infrastructure : AI agents naturally require cross-border, KYC-free settlement methods that support micropayments. USDC micropayments and native Crypto payments are worth continued attention.
One-person company infrastructure : Lobster makes "one-person companies with digital employees" truly feasible for the first time, and the tools and services surrounding this scenario have great potential.

Finally, a few pieces of advice for all shrimp farmers.

Manage expectations : The person selling the course might exaggerate the lobster's quality to 150 points, but it might only be 65 points in reality.
Dimensional reduction : Let AI capable of calculus do the addition, subtraction, multiplication, and division; the best results are achieved.
Equipment isolation : Shrimp playing machine ≠ working machine ≠ personal machine
Funds independence : Any operation involving private keys and funds must be manually confirmed twice.
Don't blindly trust Skills : Review installations before use, pay attention to updates, and be wary of supply chain contamination.
It's an intern, not a fund manager : it's fine for organizing information and assisting in decision-making, but not for managing money independently.

Note: This article is compiled from the transcript of PANews Space's "Shrimp Farmers Alliance: Tencent's Entry, Government Subsidies, Xianyu's Outsourcing—How Does the Crypto Market Cope with 'Shrimp' Anxiety?" The opinions expressed by the guests do not constitute investment advice.

Five hours of data, a 90% success rate—humanoid robots have started playing tennis with humans.

Popular Articles