9天重写100万行代码，但AI项目编译通过率仅27% (English)

Generated: 2026-06-22 05:51:05

---

Okay, let me check the facts and polish the text, removing the AI-generated feel.

---

Can AI Really Handle Complex Projects?

Let me tell you a true story.

I have a friend who's been working on enterprise digital projects for ten years. Recently, he ran a crazy experiment—alone, using only Chinese commands, without touching a single line of code, he managed to build a complete enterprise office system. Smart customer service, OA approvals—all written by AI. That system is now actually running in his company.

Guess what?

When he told me about it, my first reaction was: "You're kidding, right?"

But then I saw it. The code was there, the system was there, and it was running pretty smoothly.

---

People Think AI Fears Complexity, But It Actually Fears Something Else

Let's get one thing straight first: What exactly makes AI so powerful?

AI programming has redefined the word "complex."

In the past, when we said a project was complex, we meant a large amount of code, difficult algorithms, and tangled business logic. But AI doesn't care about any of that—the world's largest software project's codebase is just a day's work for AI. It can solve math competition-level problems; your code is child's play to it.

What really drives AI crazy is something else: things that can't be verified.

For example.

You ask AI to design an interface and say, "It's too dark." AI immediately switches to a brighter color. But if you say, "It's not pretty"—it's stumped. Why? Because "pretty" has no standard, no test, no way to verify.

See, what AI is truly good at are things that can be accurately verified.

One developer took 10 days of wedding leave and used AI to generate a self-operating "AI town," complete with maps, characters, and interaction logic, all automated. Another even more extreme example: Bun (a JavaScript runtime) used AI to rewrite 1 million lines of code in 9 days, submitting 6,755 changes with a test pass rate of 99.8%.

Sounds impressive? Don't jump to conclusions just yet.

---

AI's Most Fatal Flaw Is Also Its Most Exciting Feature

I'll be blunt—don't take offense.

AI-generated code has a core defect: it's not "clean" enough.

According to a report by research firm Snyk, the probability of known security vulnerabilities in AI-generated code is 1.5 to 2 times higher than in human-written code. A joint study by Fudan University, Tencent, and Peking University also showed that AI-generated code is becoming a potential "gray rhino" in the software security ecosystem.

Specifically, AI tends to make mistakes in business logic and API authorization logic. Among the six "critical" level vulnerabilities, most fall into these categories.

In simple terms, you've hired a highly efficient but accident-prone newbie. They work fast, but you need to constantly watch for traps they might leave behind.

Another issue: AI lacks the ability to plan for code's "sustainable development." It can generate functionally correct code, but it may not consider maintainability. One benchmark found that AI models achieved a first-time compilation success rate of only 27.38% for complete project builds (involving multiple files).

That means out of ten projects, seven and a half won't run on the first try.

In other words, you've paid a lot for a genius intern, but you have to check every piece of homework they turn in.

---

People Who Really Know How to Use AI Act as "Digital Project Managers"

So how do you use AI effectively?

My experience: Don't be a hands-off boss; be a good "digital project manager."

Many people misunderstand "AI writing code." They think you can just say, "Write me a customer service system," and AI will conjure up a complete product. That's not how it works.

AI can write code, but it won't proactively understand your business. It won't ask you: "What are your company's customer service assignment rules? By region or by issue type? How many inquiries can one agent handle at a time?"

These questions need you to answer. And only someone who truly understands the business can answer them clearly.

I've summarized a universal formula for commanding AI: Context + Goal + Input + Output + Constraints.

Here's a real example.

❌ Vague request: "Write me an OA approval system."

AI will generate a bunch of stuff you can't use.

✅ Standard request: "My company has three approval processes: leave, reimbursement, and procurement. For leave under 3 days, the department manager approves; for 3 days or more, the general manager also needs to approve. For reimbursement under 5,000, the department manager approves; for 5,000 or more, the finance director also needs to approve. Write me a web page where employees can submit requests and check approval progress, and approvers can approve or reject with one click. The page should be clean and deployable directly in a browser."

AI generates the complete code, deploy it on a VPS, open it in a browser, and it works. A functional OA approval system without writing a single line of code.

See? It's not that AI isn't capable; you just didn't command it correctly.

---

What AI Does Best Is Going from 1 to N, Not from 0 to 1

I've tested many AI tools and found a pattern: AI excels at optimization and improvement, not creation from scratch.

For example, I recently wrote an information aggregation tool. It crawls content from mainstream media, tech media, Twitter, Reddit, HuggingFace, GitHub, HackerNews, ProductHunt, arXiv, etc., based on topics I set, then uses an LLM to filter and create one-sentence briefings.

There's an EWMA (Exponentially Weighted Moving Average) heat algorithm in it: if a certain KOL talks a lot but mostly about trivial matters with little high-value information, the probability of crawling that account again decreases dynamically. Each account gets a heat score based on content quality after each crawl. Accounts that post frequently with high-quality content are always extracted often, while chatterboxes appear less and less in my feed.

I had AI help me write this feature, and it did a great job. But if I had to design the entire system architecture from scratch, AI wouldn't be much help—because it doesn't understand my business scenario, doesn't know which information is valuable to me and what's noise.

---

What Disturbs Me Most Isn't What AI Can Do, but What It Reveals to Me

Over the past three months, I conducted an experiment: I handed over all tasks that could be done by AI to AI.

Result—my workload dropped by 80%.

But what disturbs me most isn't "so many tasks can be done by AI," but seeing the documents, analyses, and plans AI produced, I realized: These tasks never should have taken so much time in the first place.

So what's the remaining 20%? That's the key.

Let me give you three real scenarios.

Scenario 1: Two conflicting requirements—which one to implement?

Operations says, "Users want a one-click share feature." Data says, "The share feature usage rate is only 3%." AI can analyze pros and cons, but it won't tell you—what's your core metric right now: user acquisition or retention? AI can't make that decision because it doesn't know what your CEO promised at the board meeting last week.

Scenario 2: What users say and what they do are completely different.

In surveys, 80% of users say they want "more features." But when you look at behavioral data, the most active users only use two core features and never touch the others. Should AI trust the survey or the data? Neither. This conclusion requires you to combine survey results, data, industry experience, and even your gut feeling to make a judgment.

Scenario 3: The tech team says, "We can't do it."

You know they can, but they think it's not worth it. How do you judge whether their "it'll take two weeks" estimate is real or just a guess? You need a basic understanding of technical solutions, and you need to squeeze out the real timeline without damaging team relationships.

These three scenarios share one thing: incomplete information, conflicting stakeholders, and irreversible decisions.

What AI excels at is the opposite: complete information, clear goals, and the ability to iterate repeatedly.

---

My Judgment: Optimistic About the Tool, Pessimistic About the Narrative

Why optimistic about the tool? The speed of capability improvement is real. What it can't do well today might actually be possible in five years. Especially in closed, standardized, testable scenarios, AI programming might truly change how development is done.

Why pessimistic about the narrative? "AI will replace programmers," "Anyone can develop complex software with AI"—these narratives severely overestimate near-term capabilities and fundamentally misunderstand the nature of complex engineering.

This kind of hype isn't new. When 3D printing was hot, people predicted