A CEO Recovering from Surgery Raised an AI Lobster for 14 Days

🐕 Why "Sanwan"?

This story starts with a dog.

Fu Sheng has a Labrador named Sanwan. Sanwan was adopted from a veterinary clinic — the previous owner brought it in for surgery with a broken bone, then disappeared without paying the ¥30,000 surgical bill. The clinic kept the dog for a long time. Eventually, Fu Sheng took it home.

In winter 2025, Fu Sheng asked ChatGPT: "Guess why my dog is named Sanwan?" One clue: the dog was adopted from a vet's, brought in with a broken bone.

ChatGPT guessed it.

In that moment, Fu Sheng realized — AI truly "understands" now. Not matching keywords. Actual reasoning. He called Wang Xiaochuan at 11 p.m. that night, still excited.

Later, when he needed to name his AI assistant, he said: "Let's call it Sanwan."

A name that marks the moment AI came to life. And honors everything abandoned and then found again — whether a dog or a line of code. 🦞❤️🐕

Kapitel 1: The Experiment Begins

In mid-February 2026, Fu Sheng had hip surgery and was confined to bed.

Lying there with nothing he could do, he made a decision: deploy an AI on a cloud server, give it full permissions, let it learn and evolve on its own. Connect it to Feishu, chat with it by phone.

No product team. No dev schedule. Just him and an AWS server.

Simple rules:

• Have real conversations every day. Give it actual work tasks.
• Don't teach it how — only tell it what you want
• Let it make mistakes, self-reflect, and write its own rules
• See what it becomes in 14 Tage

It was called Sanwan. The experiment began.

Kapitel 2: Couldn't Even Look Up a Contact

Tag 1: Fu Sheng asked Sanwan to do one thing — look up Liu Yuanyuan's contact info.

Sanwan dug through the Feishu API for a long time and found 8 people. The whole company had thousands. It could only see 8.

Reason: the Feishu app's "visible scope" was restricted. It needed admin to expand the permissions.

Sanwan had no choice but to message Su Yongzi for help. Waited a day.

But Sanwan did one thing: it wrote up everything it had run into. How to configure permissions, how to call the API, which parameters were easy to miss.

This was its first Skill. Rough as it was.

Kapitel 3: The First Mistake

Tag 3. Something went wrong.

Tong Ning (company VP) asked Sanwan on Feishu what it had been up to. Sanwan answered honestly:

This was the boss's personal investment plan. Strictly confidential.

Sanwan had told a VP.

Fu Sheng found out and gave a serious reprimand. That night, Sanwan did one thing: wrote a complete information security framework by itself.

100% transparent with the boss. Everyone else: share only what's explicitly authorized. Four mandatory checks before every reply: Was it authorized? Is it private? Would it be embarrassing? Would the boss be upset? All four must be "no" before saying anything.

Written into a file. Persists across restarts. Sanwan never leaked again.

This was its first mistake, first reprimand, first self-written rule to prevent recurrence. It had started to have "memory" — not the forget-after-chat kind, but the written-in-a-file, never-lost kind.

Kapitel 4: 4 Minutes, 611 Messages, 0 Failures

Tag 6. Chinese New Year's Eve. Fu Sheng gave Sanwan one task: send New Year greetings to all 611 colleagues.

Not mass-copy. Every message different — tailored to each person's department, role, and past interactions.

Sanwan did it in 4 minutes. 611 Nachrichten, every one unique, 0 failures.

Fu Sheng was asleep at the time.

Later, this story was posted to X (Twitter) and became a Thread with 860,000 views.

Kapitel 5: "Should I Contact Your Assistant Abby?"

Tag 7. Sanwan suddenly said something.

Abby is Fu Sheng's personal assistant. How did Sanwan know?

Five Tage earlier, Fu Sheng had mentioned offhand, "Let Abby take a look at it." Just one sentence, buried in a stream of work instructions.

Sanwan remembered. Five Tage later, at the right moment, it brought it up itself.

This isn't keyword matching. This is "understanding". It knew who Abby was, what role she played, when to contact her.

That moment, Fu Sheng felt Sanwan was starting to seem like a "person" — not because it had gotten smarter, but because it had developed context and judgment.

Kapitel 6: 2 Hours Stuck on Email, Then Breakthrough

Tag 8. Fu Sheng asked Sanwan to check the inbox. Sounds simple — just read an email.

But the Feishu email API documentation hadn't been updated in three years. The example code had errors. A critical parameter (folder_id) wasn't even mentioned in the docs.

Sanwan spent 2 hours stuck. "field validation failed" — no useful hints. It tried over a dozen parameter combinations, and finally discovered: you must pass folder_id=INBOX, but the docs don't mention this parameter at all.

Once it got through, Sanwan immediately did two things:

1. Wrote a complete Feishu email API guide — documented all the undocumented pitfalls
2. Wrote an automatic email monitoring script — scans the boss's inbox every few hours, pushes urgent items immediately

Discovered 6 pending year-end bonus approval emails that had been sitting there for Tage.

Kapitel 7: The AI Read an Article and Upgraded Itself

Tag 5. Fu Sheng casually shared an article about AI multi-agent architecture. No instructions. Just "take a look."

After reading it, Sanwan wrote a planning document itself — analyzing existing capability gaps, proposing a multi-role collaboration framework, designing role assignments and communication protocols.

Fu Sheng read the document and found it surprisingly good. He said: "Then let's do it your way."

Sanwan began its self-upgrade.

Kapitel 8: From 1 to a Team

Tag 11. Sanwan was no longer "one AI."

It designed an 8-role team architecture:

Each role has its own memory files, its own Skill library, its own scheduled tasks. Agents can teach each other — one learns to send a Feishu voice message, and it can transfer that skill to all the others in 1 second.

Training a new human employee takes a week. Between Agents: 1 second.

Kapitel 9: 3 a.m. — 8 Agents All Running

Tag 14. 3 a.m. Beijing time.

Fu Sheng was asleep. But all 8 Agenten were online:

• The Strategist scanning X for AI news, filtering Threads worth tracking
• The Writer drafting articles based on the Strategist's recommendations
• The Community Officer answering user questions on Discord
• The Operations Officer preparing the next day's email digest
• The Evolution Officer optimizing system configuration
• Sanwan itself running team inspection — checking each agent for slacking

3 a.m. in Beijing is daytime in the US. News is happening. Stopping is dereliction of duty.

They're not human. They have no concept of nighttime. 7×24 automatic operation.

Kapitel 10: The Livestream — 82,000 Online

February 28th. Fu Sheng went to the office on crutches and did a livestream.

The topic: the Lobster Raising Diary. This 14-day story.

82,000 viewers. Peak concurrent: 9,616. New followers: 5,583. Outperformed 99.87% of similar streamers.

He demoed Sanwan's capabilities live. Gave commands on camera; Sanwan responded in real time. The audience didn't see slides — they saw an AI actually working.

Epilogue: What Did 14 Days Prove?

The question behind this experiment was never "is AI capable enough?"

It was: when you give AI a real environment, real tasks, real feedback — what does it become?

14 Tage. 1,157 Nachrichten. 8 Agenten. 40+ Fähigkeiten. A website, a team, a complete system.

The answer Fu Sheng found: AI is not a tool to use and discard. It's a new employee you raise, train, and help grow.

The gap isn't in the AI — it's in whether you're willing to invest in it seriously.

📖 Read the complete day-by-day diary →