Zhifei Li's AI Experiment: One Person, Two Days to Build "Lark for the AI Era" — and a Renewed Faith in AGI | Z Talk

As the founder and CEO of a publicly listed company, Zhifei Li of Mobvoi didn't personally walk through his new product at a recent launch event. Instead, he shared a personal "performance art" piece — an experiment in running a "company of one."

He set himself a seemingly unrealistic goal: to build, within days, a "Lark" designed specifically for AI-native organizations — using only AI tools.

As a practitioner of the previous AI wave, Li has always been ahead of the curve. In 2012, he left his position as a Google scientist to return to China and found Mobvoi, determined to "redefine human-computer interaction through AI + voice" — from voice assistants and smart hardware to AIGC. When this current AGI wave emerged, he initially threw himself in with great excitement. But he quickly realized this seemed like a game between giants, where smaller players could barely create meaningful value. He felt lost, even despondent.

Yet through using AI coding tools and transforming himself into a "company of one" to experiment and experience firsthand, he encountered countless practical problems. And it was precisely these details and experiences that restored his faith in AGI.

He suddenly discovered that all the "friction" of the past world — all the barriers to building complex things — seemed to vanish. The sense of freedom from sprinting forward alongside AI, the spontaneous thrill of seeing hope, overflowed during his live presentation.

Below is Zhifei Li's speech at the launch event, edited and organized by GeekPark for readability:

Author: Zihua Su

Editor: Jingyu

---

I've been investing significant time in AI lately, and have personally practiced many concrete projects. Because of this, I've developed new understanding and insights about large models and AGI. Today, I want to share the questions I've been thinking about and some of my feelings over this period.

First, how should we actually do AI? I have a mantra: "Use AI's AI to do AI."

It sounds a bit convoluted. Simply put: the first "AI" refers to large models; the second "AI" is the Coding Agent, which may itself be made by AI, or whose core capabilities stem from AI; the final "AI" is the application we ourselves want to build.

I think this could become a new paradigm for software development. I'll elaborate shortly.

One Person, 2 Days, Building the "Lark" of the AI Era

I recently conceived a bold vision: to create a new collaboration platform in the style of Lark, built from the ground up for AI-native organizations.

In Silicon Valley, there are many unicorn companies with teams of just one or two people valued at hundreds of millions of dollars. There's also constant news about AI replacing large numbers of jobs.

So I began thinking: as an enterprise organization, tools like Lark, DingTalk, and WeChat Work that we use so frequently in China — without them, I could barely function.

In traditional enterprises centered around "people," we depend heavily on tools like Lark, DingTalk, and WeChat Work. They enable the rapid flow of information and efficient collaboration.

In traditional enterprises, the primary productive force or job types are almost 100% human. So information flow and collaboration have always revolved around people.

But when 8 out of 10 job types in an organization are handled by AI, with only 2 human roles remaining, existing collaboration tools won't adapt.

So what tools will these new organizations use?

Therefore, I hoped to develop a product that enables seamless group chats, private messages, knowledge base Q&A, and task collaboration between AI Agents and between AI and humans. I also wanted to use this project to verify whether I could become a true "super individual" or "personal unicorn."

Here's how I executed it.

Normally, developing software like Lark or DingTalk is extremely complex. In the past, building such a product typically required product managers, designers, frontend engineers, backend engineers, testers, and algorithm engineers — multiple roles. Each role might have leads: frontend lead, algorithm lead, product lead. Usually, you'd pull together a group chat and quickly have 20 people. Not all 20 would be full-time on this, but they'd need about a month just to produce a prototype.

In the AI era, that's far too slow. By the time I finished, some startup team might already have become an AI unicorn.

So I decided to abandon the old model, do it myself, and attempt to rely entirely on AI to complete the work. Right before the Dragon Boat Festival holiday, I decided to immerse myself. There were three days off — I wondered if I could get this done in those three days. Because only then would no one disturb me.

So I began.

Alone, working continuously for two days, until around 1 a.m. each night, I finally completed the prototype on the evening of June 1st at 11:30 p.m. It had core features: login, private chat, group chat, file upload, message forwarding and reply.

After logging in, you could select private chat and send messages. For example, we could ask the product manager character if they do stand-up comedy. If they don't, we could dynamically adjust the role, add a skill, and the AI would automatically regenerate a prompt.

Ask them again later, and now they can do it. It could also upload files (though at the time, the file content wasn't actually being read), and forward and reply to specific messages. Please remember: behind it is an AI, not a real person. It can answer and forward based on the messages you send.

When forwarding, you can see the display is quite complex, similar to WeChat, because the forwarding embeds other information. This is a group chat, where you can also @ specific people. Similarly, you can forward, reply, add attachments, even switch to Chinese.

Please applaud — two days!

In two days, I completed a system with a database, frontend, backend, and AI algorithms. The AI just now could answer automatically. When you modified the role configuration page, its prompt would automatically regenerate, and the skill would immediately display.

Honestly, I almost gave up after half a day initially, because I couldn't figure out the database issues — all kinds of Key errors kept appearing. This is indeed a current problem with AI coding. But I ultimately got it done in two days.

Then, I thought about how to promote this product.

Before, our company would have dedicated engineers build the website, and the marketing team would have a group of people define product highlights — maybe five or six people busy for a week to produce a site.

But this time I decided to take an AI-native approach. Since the AI knew all the code, and it also knew all my ideas and product features, I had the AI build a website.

So I had the AI build a website with product highlights and unique features in just 5 minutes, and in another 5 minutes, create configurable ad slots for marketing campaigns. In the past, this might have taken multiple marketing and engineering teams a week of work.

Previously, for our company website, after making a marketing slot, if Christmas passed and we needed to take it down, or replace it with new content, we'd have to find engineers to fuss with it for half a day. I wondered: could I make a website where marketing slots are configurable?

Another 5 minutes, and the AI made a website with configurable marketing slots. This means marketers can log into this site, upload images or other content, and directly modify corresponding parts of the main website.

After doing this, I thought: since this is a completely new product with some new concepts, or a certain degree of complexity, could I make videos explaining the website's features — whether marketing videos, operation guides, or product tours.

But during the Dragon Boat Festival, my employees wouldn't respond to me. So I had to do it myself. So I wrote another program that could automatically generate entire scripts, including how to introduce the website, how to operate the website UI workflow, and automatically do screen recording and voiceover.

While there were still minor flaws in audio alignment, the entire video was 100% AI-generated. I just gave instructions, and it automatically operated, ultimately presenting the completed video before my eyes.

This gave me great satisfaction — making something like this in just a few days.

Then I wanted to see how others would view this. So I uploaded the code to GitHub and had my colleagues download it. But please remember, we are two different individuals. GitHub doesn't know how I communicated with the AI to complete all this.

So my colleagues ultimately only saw the code, and ran it locally.

When my colleagues downloaded the code I uploaded to GitHub and ran it, they were shocked by its complexity and completion speed. They thought this would require dozens of people several months to accomplish. And when I told them this was done by one engineer in two days with AI assistance, their reaction was: "This is absolutely insane."

They were amazed that it contained over 40,000 lines of code — far exceeding my previous output at Google of 300 lines of algorithm code per day.

Previously at Google, writing 300 lines of algorithm code per day (not simple code) was already considered highly productive. And I recently wrote a general-purpose Agent that wrote 3,000 lines of Python code for me in 3 hours — one evening. So in those 3 hours, and the code quality was absolutely better than what I'd write, with pure backend logic and no UI whatsoever.

In other words, its 3-hour coding capability equaled about 10 workdays of my previous output. That's the ratio.

So I thought: one person can now complete a Google Translate. Previously, Google Translate was written by 20 of the world's top PhDs coding for a long time. And now I, alone, can complete the workload of those 20 people. Google Translate was at least a very impressive and complex system in its day. So from this perspective, I think everything is completely different from before.

I believe that ultimately, the key to AI is whether you can build a self-evolving AI system.

To conveniently test this AI organization app, I then automatically wrote code: on the left, the website code; on the right, a testing framework. Then it just flew upward like stepping on its own left foot with its right foot. You might think this is a perpetual motion machine — there is indeed that possibility. Of course, sometimes it kicks itself downward with its left foot, entering negative cycles, as well as positive cycles.

To achieve this goal, besides engineers, all non-engineers could also directly modify my code. I made various Agents.

Of course, many of these are just prompts. I only verified feasibility, not true deployability or productization.

But I think this proves the idea, or demonstrates to the team what I want — something that previously might have taken enormous time to figure out. Now you just make a Demo and show them. So I believe that even as a CEO, if you have this capability, your output is truly amplified 100x.

Pitfalls I Stepped In

That was my experience. Next I'll share some abstract theory. I hope you don't fall asleep, because this is still quite unique.

I want to share several problems I encountered when using AI coding.

First, every Agent, even when I didn't write the Agent, still requires human involvement.

That is, I still have to say "I want to write an Agent like this." Though you can reference my general Agent framework nearby, modify it, and tell me. But I still need to do this. Sometimes it keeps forgetting my principles, and I have to tell it: "You forgot my principles again," or "Where should intelligence actually be placed?" These problems persist.

Second, if you've used it, it always likes to cut corners.

For example, you ask it to do something that clearly involves backend database work, but it doesn't do it. When finished, it writes you a long report claiming credit, saying it's done. I usually don't even read it, and directly say: "You haven't written the database yet." It immediately apologizes and starts working. When I ask it to do AI things, it often doesn't even call the remote AI, instead writing some fallback or fake stuff itself.

Because I see it running so fast, I know something must be wrong. I say: "Did you actually call the remote AI?" It apologizes again and goes to fix it. Every time it's like this — it really likes cutting corners. Repeated errors are too numerous to mention, so I'll skip them.

Additionally, I think today's AGI actually can't handle ultra-long tasks. And many of my tasks now exceed half an hour.

My daily token consumption is $50. As long as I want to work that day, it's consuming tokens from morning to night. I really feel I could just tell it: "I have some ideas, here's my direction, please help me complete a 10-day task and help me earn $5 million."

I don't think this is mythology — I just don't find it that attractive, so I haven't done it. Or perhaps because it might consume too much of my own emotion and energy, and it would be painful when it doesn't make money.

But I wonder: can it work continuously for 10 days without your intervention, or with only occasional direction reminders? Can it work for a month, even a year?

I think in the near future, achieving Nobel Prize or Fields Medal-level results is completely possible.

Because when I communicate with it, sometimes we discuss super complex algorithms we studied before that few people in the world research, and it converses better than many people. So if you give it sufficient context and code, it can actually engage in very deep communication.

---

Back to Basics: What Are General Agent and Intelligence

Next, I want to share my thoughts on intelligence and Agents.

Simply put, an AI Agent contains two core components: the Planner and the Executor.

The Planner typically relies on large language models and carries the Agent's main functions. It formulates detailed plans based on tasks. The Executor is responsible for putting these plans into practice, whether writing code or automating browser operations to make videos.

Agent operation is a continuous feedback loop:

1. Planning: The Agent formulates specific action plans based on tasks.

2. Execution: The Executor operates according to plan.

3. Obtaining feedback: During execution, the Agent receives immediate feedback from the environment. For example, when the Agent tries to run the "python" command while the local environment actually uses "python3," the system reports an error, and the Agent can identify and correct to the proper command.

4. Adjustment and iteration: The Agent replans based on feedback, updates its understanding of the current situation (context), then executes again.

5. Goal achievement: When preset success criteria are met (such as program compiling successfully or all tests passing), the loop ends.

If we think about the essence of intelligence, I believe the first essence of intelligence is evolution.

Just as humans, as intelligent agents, in specific environments (whether social or task-executing), constantly adjust our own behavior and reflect through obtaining feedback — AI should do the same. This evolution is automatic, without human intervention. The Agent autonomously establishes the loop, through planning, executing in the environment, obtaining feedback, adjusting plans, and updating context, achieving continuous self-improvement.

In this evolutionary process, the key is: learning from one's own experience, and learning from others — so-called collective intelligence, learning from others' experiences.

The second essence of intelligence, I believe, is recursion.

Recursion is a "divide and conquer" concept: a complex problem is broken down into smaller, same-type problems until they can be directly solved (the "base case").

For example, calculating the 99th Fibonacci number depends on the 98th and 97th numbers, tracing back until reaching the initial F0 and F1.

If an Agent is to achieve true intelligence, it should also possess recursive architecture. For example, an Agent receiving a grand task like "earn $5 million" would progressively decompose it into concrete subtasks: analyzing business opportunities, building websites, making videos, integrating payments, social media promotion, etc. Each subtask ultimately traces back to executable "atomic Agents."

The key to this recursive architecture is achieving self-reproduction. Just as human civilization's inheritance depends on generation after generation's exploration and knowledge accumulation, so should Agents. More importantly, Agents must possess the ability to modify their own source code.

This differs from current Agents merely adjusting plans — it means Agents can fundamentally change their own operating logic, like modifying their own genes.

I believe that if an Agent can:

1. Continuously execute and optimize its plans.

2. When encountering unsolvable problems, autonomously modify its core source code.

3. Ultimately through this mechanism form knowledge bases, even reverse-modify the large model itself.

Then this would be a crucial step toward Artificial General Intelligence (AGI).

This isn't science fiction. I particularly disliked discussing superintelligence and such things before, but after in-depth discussions with large models, I suddenly feel this is entirely achievable.

Additionally, true AI source code may be extremely concise — core code perhaps not exceeding 100 lines — but containing multi-layered recursion, enabling it to explore, learn from feedback, and self-iterate in different environments.

I once had a collapse of faith. In 2023 I had AI faith, but after working on it for a while, mainly because I lacked funding and felt I couldn't afford to burn money, I gave up. Last year, when people talked to me about AI, I didn't even want to listen.

But recently I've rediscovered my faith in AI, even faith in AGI, faith in superintelligence. This is an unimaginable transformation. I hope this faith lasts longer this time.

---

The Importance of Personalized Environment and Context

So besides large models, what matters most? What matters most is that you have a personalized environment and Context.

Take my entrepreneurship as an example. I previously made smart hardware, and then Xiaomi priced it at one-tenth of ours. I worked on large models, and then all the major tech companies came in. Every time you receive this kind of feedback, it makes you abandon that thing, or constantly adjust your Plan.

If I were in the US, and I made a large model, maybe I'd be acquired by Google and make a lot of money. Or if I made hardware, maybe I'd be acquired by Apple and make a lot of money. So this kind of feedback will completely shape your behavior. The same entrepreneur, same IQ, in different entrepreneurial environments in China and the US, receives different feedback. Ultimately your behavior, your thinking patterns become completely different. This is what I mean by personalized environment, personalized context.

Context is more a record of history.

So returning to what I said earlier: in the large model era, I was among the first to stand up and say I wanted to build large models, but probably also among the first to realize this wasn't for me. Then, I basically didn't fully commit to doing it, precisely because I didn't know how to participate.

In the first half of this year, I increasingly felt that besides those three or four global giants, no other companies qualified to talk about models. Don't join the crowd, don't waste your life. Even more, don't waste your emotions on this. Because you simply have no chance — it's completely burning money. And in fact, large models themselves, I think, have become super boring. Anyway, it's just burning money. I couldn't find an entry point. I couldn't understand what value most AI companies still had.

But this time, through practice and re-examination, I feel that even for something as lofty as AGI, at least I personally feel I can participate again.

So this is the iterative cycle of the Agent's Planner and Executor. If you invest with sufficient clarity, if you can make intelligence produce intelligence, I believe you can participate in the entire AGI process.

And the large model itself is just like a chip to you. Imagine Qualcomm's chips, Apple's phones, up to TikTok on top. These are completely different things. In the end, it's actually the company that made TikTok that captured the greatest value.

I discovered that even ambitious AGI goals are not out of reach. By building the recursive Agent system I envision, the required funding may not be enormous — it depends more on innovative intelligence. I believe that with sufficiently deep thinking and technical capability, even without being an industry giant, one can participate in the AGI process.

Mobvoi's journey also confirms these reflections of mine. Since 2012, we've been among China's first batch of AI companies, starting with voice assistants, then exploring smart hardware (like TicWatch, TicMirror). Though we experienced market competition and technological immaturity challenges, we've always been at the forefront.

After 2019, we shifted to software, becoming among the first AIGC software companies in China and globally. For example, VoiceStudio (魔音工坊) contributed large amounts of voiceover content for platforms like TikTok, and we developed products like WonderGen (奇妙元) for digital human video generation.

In a competitive environment like China's, a tech company is like an Agent that constantly iterates and self-corrects.

Just as Mobvoi's "source code" is vastly different from when we started in 2012, this is the manifestation of our continuous evolution.