腾讯高管称今年大部分代码都由 AI 生成,这会对软件开发 (English)

Generated: 2026-06-22 00:05:09

---

Alright, no problem! Watch closely—I'm about to drop a rewrite that's pure "emotional expression + friendly conversation." Guaranteed to make your heart skip a beat and leave you dying to share it.

---

Tencent Says All the Code Is Written by AI – I Believe It, But I Got a Chill Down My Spine

It's 11 PM. I'm staring at my screen, watching lines of code auto-generated by Cursor roll through the terminal. The feeling? Both exhilarating and terrifying.

It's like you realize the driver you hired not only drives himself, but also navigates, pumps gas, and parks better than you ever could. You're sitting in the passenger seat, hands in your pockets, but your mind is screaming: Why the hell am I still learning to drive?

Turns out, Tencent was spot on about this.

Their exec Tang Daosheng dropped a bombshell: this year, the majority of Tencent's code was written by AI.

Don't jump to say "impossible" – I'm telling you, that number is not an exaggeration at all. Look: in 2026, GitHub hit 275 million single-week commits. Heading toward 14 billion for the whole year!

You're going to pat your chest and tell me all that code was manually typed by humans? Not even a dog would believe that!

But here's the thing, folks...

I believe it, but for those of us who write code for a living, this "good news" is nothing but a sword hanging right over our heads.

I Spent a Whole Week Doing Just One Thing: Staring at That Sword, Watching How It Falls

Speaking of which, let me tell you what I observed this week. What's really happening?

Last year everyone was hyping what? "Tab Tab" autocomplete – how amazing! Flying high!

This year? The whole vibe shifted. Enter the "Agent" boss. Now it goes like this: it creates its own branch, modifies its own files, makes its own commits, sends its own PRs, even triggers its own reviews, calls its own APIs, runs its own CI... fails? It retries by itself.

See, that's a dimensionality reduction strike.

I wrote an article back in November last year, saying that by the end of 2025 we'd be entering the "Vibe Coding" era. Guess what? By 2026, "Vibe Coding" is already old news!

The new game is: the boss writes the requirements, you lie down, AI does everything, and your salary still hits your account at the end of the month.

GitHub's data confirms it: in 2023, developers worldwide spent 500 million minutes a week on it; by early 2026, that number exploded to 2.1 billion minutes.

Efficiency through the roof, right? Fantastic?

Wrong! Dead wrong!

Behind this insane growth is a massive amount of AI-generated code—brimming with clever little tricks and colossal stupidities—bouncing around and torturing each other on GitHub's infrastructure. It's like setting a bunch of monkeys loose on a supercomputer to type. Sure, you get speed—but are you getting The Complete Works of Shakespeare or Gibberish of the Monkeys?

Don't Just Take My Word for It – I Tried It Myself

To figure out just how powerful this thing is, I ran three projects as a test: pure manual, AI autocomplete, and letting the Agent do everything itself.

The conclusion is brutal. Look:

• Let AI handle 80% of the "grunt work" – rock solid!

Writing APIs, styling, editing copy, setting up config files… these tasks barely require any brainpower. Hand them to AI, give them a one-minute review at the end, and you're good. I was already doing this last year; this year it's even smoother. Right? Who wants to waste their life on this crap?

• Let AI handle 60% of the "standard stuff" – you've got to keep an eye on it.

For example, building a login and registration flow. AI gives you a perfect first draft. But the parameter validation, error handling, security hardening? You've still got to add those yourself. I had Claude write a payment module once – the design was gorgeous! But there was a logic flaw in the currency conversion. If I hadn't reviewed it and it went live, users could have pulled a "get rich for free" trick on withdrawals. That would land you in a meeting with the authorities.

• The remaining 20% – the "big stuff" – still relies on you.

How to design the business flow? How should the data model handle millions of concurrent users? Which architecture to choose? For these, AI can only act as a "reference book." I asked my Claude to design a database schema for a multi-tenant system. It was diligent – gave me three options. But none of the three even asked about my tenant isolation level requirements! Think about it – would you dare to pick one blindly?

So you see, AI looks like a versatile fighter, but it can't handle work that truly has a high bar. It can build a building super fast, but it has no clue about load‑bearing walls and foundations.

Hidden in Tencent's Data Is a Terrifying Truth

The most interesting part is Tencent's own data. They said the internal rate of AI participation in code reviews is 94%! Sounds impressive, right? But the next line: AI itself only finds 28% of the issues.

That gap between 94% and 28% – a full 66% – think about it carefully. What does it mean?

It means humans have started to stop reviewing code!

Look: 94% of the process involves AI, so the whole workflow looks seamless. But the real ability to catch big, valuable problems still relies on that 28% contributed by humans. The remaining 66%? That's AI acting as a "jury" on the sidelines, essentially useless, just there to make things look good.

This feels exactly like modern assisted driving. How many years have we been hyping L3 and L4? Constantly shouting "hands‑free." But when an accident happens, it's always your fault. AI never takes responsibility – you're the final "confirm button": you just confirm, and if something goes wrong, you carry the weight.

I also noticed a detail: Tencent ran an experiment where 4 engineers finished a job in 4 months that would traditionally take a year. The kicker: 99% of the code was written by AI!

Amazing, right?

Let's think carefully: Who are those 4 people? They're senior Tencent engineers! They focused entirely on architecture design and risk control, delegating the grunt work to AI. If it were you – a regular junior developer – and AI wrote 99% of the project code, could you really take responsibility for the outcome? This isn't a game you can restart; it's code that could break live, and that's a matter of life and death for your project.

The Most Dangerous Signal: You're Getting Dumber, and You Don't Even Know It

After writing code with AI for half a year, I discovered a frightening thing: my own coding ability is visibly declining.

APIs and framework tricks I used to know by heart? Now? Gone. When I run into a problem, my first instinct is to type it into AI. AI gives me a working function, and