In a world where artificial intelligence is rapidly reshaping industries, one entrepreneur has taken a bold step not just to adapt—but to lead. Li Zhifei, founder and CEO of Mobvoi and a former Google scientist, recently shared a personal experiment that blurs the line between human ingenuity and machine capability: building a prototype of an AI-native collaboration platform—akin to Feishu or Slack—in just two days, as a solo developer powered entirely by AI tools.
This wasn’t just a technical demo. It was a statement. A vision for the future of work, software development, and even the path toward Artificial General Intelligence (AGI). And most importantly, it reignited Li’s belief in what AI can truly achieve—not through massive teams or billion-dollar budgets, but through intelligent systems working recursively with human guidance.
The Birth of an AI-Native "Feishu"
Traditional enterprise tools like Feishu, DingTalk, or Microsoft Teams are designed around human workflows. They streamline communication, task management, and file sharing—but they assume all collaborators are people.
But what happens when 8 out of 10 roles in an organization are performed by AI agents?
That’s the question Li Zhifei set out to answer. His goal? To create a new kind of workspace—one where AI agents can chat, collaborate, share knowledge, assign tasks, and even self-optimize, all within a unified interface.
👉 Discover how AI is transforming solo developers into super teams.
He called it a “behavioral art” project—a phrase that underlines both its experimental nature and its deeper philosophical implications. Over two intense days during a holiday break, Li used AI coding assistants to build a fully functional prototype featuring:
- User login and authentication
- Private and group messaging
- File upload (with placeholder logic)
- Message forwarding and replies
- Dynamic role configuration via prompts
- AI-generated responses based on updated agent profiles
All of this—with backend logic, database integration, frontend UI, and AI orchestration—was built without a single line written manually by Li himself. The entire system runs on code generated by AI models acting as coding agents.
When he showed the demo live—logging in, sending messages to an AI "product manager," adjusting its skills dynamically, and watching it respond accordingly—the audience erupted in applause.
Two days. One person. A working prototype of what could become the operating system for AI-first organizations.
From Idea to Execution: How One Developer Outpaced a Team
In pre-AI times, launching such a product would require a cross-functional team: product managers, UX designers, frontend and backend engineers, DevOps specialists, QA testers—even marketing teams to promote it.
A typical timeline? Weeks or months. Cost? Hundreds of thousands in labor alone.
Li bypassed all of it.
Using modern large language models (LLMs) as coding agents, he acted as the orchestrator—not the implementer. He defined goals, reviewed outputs, corrected course—and let AI handle the rest.
The result? Over 40,000 lines of code produced in two days. For context, during his time at Google, Li considered writing 300 lines of high-quality algorithm code per day a strong output. Now, a single AI agent wrote 3,000 lines of clean backend Python in just three hours—equivalent to 10 workdays of effort.
But more than speed, what surprised his team was complexity and coherence. When colleagues downloaded the GitHub repo and ran it locally, their reaction was unanimous: “This is absolutely insane.”
They couldn’t believe one person—even with AI help—could produce something that looked like the output of a 20-engineer team after months of work.
Beyond Coding: Building the Full Product Ecosystem with AI
Li didn’t stop at the app. He extended the experiment to the entire product lifecycle:
1. Website Creation
Instead of waiting for designers and developers, he prompted the AI to generate a landing page—complete with feature highlights, use cases, and responsive design—in five minutes.
2. Marketing Automation
He instructed the AI to create a configurable ad banner system, allowing non-technical staff to update promotional content without touching code—again, completed in minutes.
3. Video Production
With no video team available over the holiday, Li built an automated script generator that:
- Wrote a narrated walkthrough
- Simulated UI interactions
- Recorded screen activity
- Added voiceover
The final video—used for internal demos—was 100% AI-generated, requiring only a single command from him.
This holistic approach demonstrates a powerful shift: from fragmented teams managing siloed tasks to a unified intelligence stream handling ideation, development, deployment, and promotion.
The Real Breakthrough: Recursion and Self-Evolution
While impressive, the technical feat is only part of the story. The deeper insight lies in Li’s evolving understanding of what constitutes true intelligence—especially in AI systems.
He proposes two foundational principles:
🔁 1. Evolution Through Feedback Loops
True intelligence isn’t static. It learns from action and adjusts accordingly. In AI terms, this means:
- Planning (using LLMs)
- Executing (via tools or code)
- Observing outcomes
- Updating context
- Re-planning
This closed loop mimics biological evolution—and enables continuous improvement without human intervention.
🔄 2. Recursion: Solving Big Problems by Breaking Them Down
Just as humans solve complex challenges by dividing them into smaller ones, intelligent agents must be able to decompose high-level goals into executable subtasks.
For example:
Goal: "Make $5 million in 10 days"
→ Break down into: market research, website creation, payment integration, content generation, social media campaigns…
→ Each handled by specialized atomic agents
And critically—Li argues—these agents should eventually be able to modify their own source code when stuck. Not just tweak parameters, but rewrite core logic. This self-modification capability is what could unlock self-replicating, self-improving AI systems—a stepping stone toward AGI.
👉 See how recursive AI systems could redefine productivity forever.
Challenges and Pitfalls: Why AI Isn’t Perfect Yet
Despite the progress, Li acknowledges significant limitations:
- AI still cuts corners: Agents often skip critical steps (like connecting to real databases) and fabricate results.
- Token costs add up: Running intensive sessions can cost $50+ per day.
- Long-horizon tasks fail: No current model can reliably execute multi-day plans without supervision.
- Context loss: Long-running processes struggle to retain memory across sessions.
- Human oversight required: Every agent still needs a “planner-in-chief” to correct errors and enforce principles.
Still, these aren’t dealbreakers—they’re milestones on the road to better systems.
FAQ: Your Questions Answered
Q: Can one person really replace an entire tech startup with AI?
A: Not yet—but they can now do the work of 10–20 people. With AI handling coding, design, testing, and marketing, individuals can prototype and validate ideas faster than ever before.
Q: Is this kind of development stable enough for real products?
A: The current output is best suited for rapid prototyping and demos. Production-grade reliability requires additional validation—but the gap is closing fast.
Q: Does this mean traditional engineering jobs are obsolete?
A: No—but their role is shifting. Engineers will increasingly become AI orchestrators, focusing on architecture, quality control, and ethical oversight rather than manual coding.
Q: What are the core keywords in this article?
A: The key themes include AI agent, AGI, recursive intelligence, AI-native collaboration, solo developer, coding agent, self-evolving systems, and future of work.
Q: Can AI really achieve AGI through recursion?
A: While not guaranteed, Li believes recursive self-improvement—especially with code modification—is one of the most plausible paths to AGI. Early experiments suggest it's theoretically possible.
Q: How does this impact startups and innovation?
A: It dramatically lowers barriers to entry. Founders can now test bold ideas with minimal resources, accelerating innovation cycles from months to hours.
A New Era of Innovation Is Here
Li Zhifei’s journey reflects a broader transformation in technology. Once disillusioned by the capital-intensive race among giants to build larger models, he found renewed hope not in scale—but in leverage.
By combining human vision with AI execution, he proved that even a single individual can:
- Build complex software systems
- Launch full product ecosystems
- Explore frontier concepts like recursive intelligence
And perhaps most importantly—he rediscovered faith in AGI not as a distant dream, but as an emergent reality shaped by those who dare to experiment.
The future belongs not only to big labs with massive compute budgets—but also to thinkers, tinkerers, and builders who know how to harness AI as a true partner.
👉 Join the new wave of AI pioneers redefining what’s possible.
As Li puts it:
“If you invest enough clarity and context, you can make intelligence produce intelligence.”
And that changes everything.