Starting Small
Great startups always begin with an extremely narrow entry point.
A former boss recently visited our company and shared his latest entrepreneurial insights.
He observed that many successful ventures start with a tiny idea—then go deep, execute thoroughly, and build solid foundations before gradually expanding. That’s spot-on.
Take Toutiao (now Jinri Toutiao), for example. Before launching its flagship app, the team experimented with smaller applications—testing core assumptions and refining their algorithmic recommendation engine. Only after validating that logic did they scale up to a full-fledged product—and later spin off related apps built on the same foundational principle.
Two points stand out here. First, you must anchor your effort in a specific, well-defined scenario: What exact problem does it solve? For whom? The user base doesn’t need to be massive—a niche image community with just a few million users can be more than enough to validate your business model or operational logic. From there, thoughtful extension and iteration become possible.
Second, when choosing that entry point, you must assess its ceiling—its inherent scalability and long-term potential. If no matter how hard you push, the idea hits a hard limit—say, a market too small or a technical constraint too rigid—you’ll stall early. So picking the right starting point demands both tactical clarity and strategic foresight.
How AI Can Multiply Productivity Tenfold
Designing a full course curriculum is grueling work.
Traditionally, it involves industry research, audience positioning, cross-team alignment, and line-by-line editing of every lesson script.
For a 50-lesson series, this process routinely takes at least one month—even with experienced teams. Several friends confirmed the same timeline.
But recent hands-on experience shows AI can dramatically accelerate this—by roughly ten times.
Yet three human capabilities remain irreplaceable:
- Methodological clarity: You must define the project’s underlying logic and rules—not just what to produce, but why and how it fits together.
- Structural discipline: AI writes well only when fed clear frameworks—content outlines, logical flow, consistent voice, and quality benchmarks. Even then, outputs require iterative refinement.
- Prompt engineering fluency: Translating those frameworks into precise, context-aware prompts—and adjusting them continuously during production—is where real skill lies. It’s not intuitive. It’s learnable—but it’s nontrivial.
AI does hold enormous productivity potential. But unlocking it isn’t plug-and-play.
This week, my former online school supervisor stopped by. We exchanged practical AI use cases—and he told a telling story: while teaching AI to his child, he realized kids approach it differently than adults—and it’s harder to teach than he expected. He admitted his own learning curve was steeper than anticipated.
Another friend echoed this on a recent trip. He argued that while using AI feels effortless, mastering it is genuinely hard—far harder than learning to search on Baidu was for ordinary users two decades ago.
Understanding Fine-Tuning vs. Prompting
How do fine-tuning and prompting differ?
A friend offered a vivid analogy:
Writing a prompt is like training a bright child—you give a rough instruction (“Explain photosynthesis like I’m twelve”), and they deliver something solid—say, an 80% solution.
Fine-tuning, by contrast, is like training a slower learner. You feed them hundreds of examples, correct their mistakes repeatedly, and slowly lift their performance—from 70% to 80%, then 90%. It’s painstaking. Think of teaching Yang Guo martial arts: progress is real, but slow and labor-intensive.
It’s not a technically rigorous comparison—but for absolute beginners, it’s remarkably effective at clarifying the practical difference: prompting leverages existing intelligence with light guidance; fine-tuning reshapes the model itself through sustained, data-driven training.
Running as Social Infrastructure
Today in Tianjin, I met a friend at the Olympic Sports Center track for a morning run.
I woke at 6 a.m., grabbed my gear, rode a shared yellow bike to the stadium—and we started running together.
We ran for over an hour, talking the whole time. Afterward, we sat on the track for another hour, continuing the conversation.
I’ve grown deeply fond of this kind of interaction: movement + unstructured, high-bandwidth dialogue.
We covered MCN operations, education product design, teacher development models—and his own evolving thinking on all of it. His insights—grounded in real cases—gave me fresh perspective.
What stands out is how clear-headed and productive these conversations feel. Physical exertion seems to lower mental friction, deepen attention, and raise the quality of reflection. It’s not just exercise—it’s cognitive infrastructure. And it’s highly recommended.