Name: Anthropic 打造編譯器的真實原因！ (The Real Reason Anthropic built a Compiler)
Uploaded: 2026-02-19T01:18:46.000Z
Duration: 8 min 18 s
Description: 【看影片學英語】數萬部 YouTube 影片，搭配英漢字典即點即查，輕鬆掌握單字發音與用法，長久累積看電影不必再看字幕。

程度副詞

Anthropic just released a marketing video proving that Claude, their flagship programming model, can produce a C compiler from scratch.

Anthropic 最近發布了一個行銷影片，證明瞭他們旗艦級的程式模型 Claude，可以從零開始產生一個 C 編譯器。

And they can do it in parallel and it can compile Linux, SQLite, Redis, Lua and even Doom.

而且他們可以平行處理，編譯 Linux、SQLite、Redis、Lua，甚至 Doom。

This is going to be a weird video because I'm going to say something positive about Anthropic and then I'm also going to call Anthropics one of the most deceitful companies I have ever seen.

這會是一個有點奇怪的影片，因為我一方面要說 Anthropic 的好話，另一方面又要說 Anthropic 是我看過最騙人的公司之一。

This little product demo that they did, the language inside the video, it is the most dishonest framing of what they actually did versus what they're saying they did.

他們做的這個小產品展示，影片裡的說法，對於他們實際做的事情和他們說他們做的事情之間的呈現，是最不誠實的。

Anthropic and its non stop hyping of the AI models is perhaps one of the most annoying features of 2025 and it's coming into 2026 with a whole new level of rigor and, and

Anthropic 和他們不停炒作的 AI 模型，可能是 2025 年最讓人討厭的特點之一，而且進入 2026 年時，還會帶著全新的嚴謹度和完整性。

All right, so let's go over what they did, what the actual cool part and actually why.

好了，那我們就來看看他們做了什麼，什麼才是真正酷的地方，以及為什麼。

What they did was complete hype, nonsense.

他們做的事情完全是炒作、胡說八道。

Software engineering is in fact not over quite yet.

軟體工程其實還沒那麼快結束。

Now that video that you just got done saying there's so many lies into it.

所以剛才你說那個影片充滿謊言。

So we're going to just try, we're going to try our best to kind of take apart the lie.

那我們就來試試看，我們會盡力拆穿這些謊言。

So first off, the thing that they were actually building was, was trying to see can they get CLAUDE to run by itself long enough to produce a functional piece of software.

所以首先，他們實際在做的，是想看看能不能讓 CLAUDE 自己跑足夠長的時間，來產出一個功能性的軟體。

Functional of course just means that it can do what it says it does regardless of how good it does it.

功能性當然就是指它能做它說它做的事，不管做得好不好。

Long as at the end of the day you say do X, it does X, you compile Linux, it should produce Linux as an artifact.

只要最後你說做 X，它就做 X，你編譯 Linux，它就應該產出 Linux 這個成品。

Also put a little star on that because that even that not so true.

這點也請畫個小重點，因為這也不是那麼絕對真實。

All right, the article kind of kicks off with two test dress it.

好了，文章一開始就用兩個測試案例來開場。

Now it being this multi agent harness to be able to have Claude run autonomously for multiple weeks, I tasked 16 agents with writing a Rust based C compiler from scratch capable of compiling the Linux kernel over nearly 2,000 Claude code sessions and $20,000 worth of API cost.

所以現在，它作為一個多代理的系統，能夠讓 CLAUDE 自主運行好幾個星期，我指派了 16 個代理，從零開始撰寫一個基於 Rust 的 C 編譯器，能夠在將近 2,000 次 CLAUDE 的程式碼對話和 20,000 美元的 API 費用中編譯 Linux 核心。

The agent produced a 100,000 line compiler that can build Linux 6.9 on X, 86 ARB and RISC V.

該代理產出了一個 10 萬行的編譯器，可以在 X、86 ARB 和 RISC V 上建構 Linux 6.9。

再次強調，沒有任何星號標示。

那麼，這些所謂的星號標示又是什麼呢？

Well first off, I don't know if you know this, but

嗯，首先，我不知道你們知不知道，

gcc, which is what they're basing this off of, not Only is it open source, which means that they've already trained on top of it.

gcc，也就是他們以此為基礎的，它不僅是開源的，這意味著他們已經在上面進行了訓練。

And we all know if, in case you didn't know that these models you can produce nearly perfect identical copyright works back out of them if you can convince them to produce it.

而且我們都知道，如果你不知道的話，這些模型如果你能讓它們產出，它們幾乎可以產出近乎完美的相同版權作品。

Like this one right here, showing that Claude3.7 can produce 95.8% of the first book of Harry Potter only almost a near verbatim reproduction, but

就像這裡這個例子，顯示 CLAUDE3.7 可以產出《哈利波特》第一集的 95.8%，幾乎是逐字重現，但是

When starting from scratch, scratch meant 37 years of developed torture.

從零開始，所謂的從零開始意味著 37 年的開發磨難。

Test suite from GCC yes, you know this classic way of doing software development.

GCC 的測試套件，對，你懂的，這是軟體開發的經典方式。

Usually when I start a new project, I first get handed 37 years of prior art in which contains the perfect golden test suite so that I can test my new project in every conceivable way possible.

通常我開始一個新專案時，我會先拿到 37 年的先期技術，裡麵包含了完美的黃金測試套件，這樣我就可以用各種可能的方式來測試我的新專案。

And then on top of it I have the actual working version of the thing I'm about to build that I can also test my working version against.

然後在我手上的是我即將建置的實際可用版本，我也能用它來測試我的可用版本。

And that's known as the online Oracle, as we, as we say in this software biz.

這就是我們在軟體界常說的「線上神諭」。

Also, when I say I can compile Linux from scratch, I want you to understand that it lacks a 16 bit x86 compiler necessary to boot Linux out of the real mode.

另外，當我說我可以從頭開始編譯 Linux 時，我希望你明白，它缺少一個必要的 16 位元 x86 編譯器，以便在實模式下啟動 Linux。

That's because OPUS was unable to implement a 16bit x86 code generated needed to boot into 16bit real mode.

這是因為 OPUS 無法實作啟動 16 位元實模式所需的 16 位元 x86 程式碼產生。

While the compiler can output the correct 16bit x86 via the 66 and 67 opcode prefixes, the resulting compiled output is over 60kb, far exceeding the 32k code limit enforced by Linux.

雖然編譯器可以透過 66 和 67 的操作碼前綴輸出正確的 16 位元 x86，但產生的編譯輸出超過 60KB，遠遠超過 Linux 強制執行的 32KB 程式碼限制。

Like they have this stupid 32 bit maximum number on.

他們有個愚蠢的 32 位元最大數量限制。

All right, so to say that all in a quick sentence, they started a project from scratch, having every single conceivable test and a way to test it under all conditions.

總之，用一句話來說，他們從零開始了一個專案，包含了所有能想到的測試，以及在各種條件下進行測試的方法。

They had an online version of it, GCC online, to be able to call out to, to validate everything they've done.

他們有一個線上版的，GCC online，可以隨時調用，來驗證他們所做的一切。

And lastly, they couldn't actually make Linux boot because they were unable to produce an artifact small enough.

最後，他們無法真正讓 Linux 啟動，因為他們無法產生足夠小的產物。

So they built a compiler completely from scratch in two weeks and they just had to walk away.

所以他們在兩週內完全從零開始建構了一個編譯器，然後就不得不放棄了。

They also couldn't walk away during those two weeks because sometimes it did things and it would crash itself and we had to restart it and make sure things were working.

在這兩週內他們也無法真正走開，因為有時候它會出現一些狀況然後崩潰，我們必須重新啟動它，確保一切正常運作。

And this is exactly what I mean by dishonesty.

It's like, yeah, dude, we said, hey, Claude, you'll build this from scratch.

就像，對啦，老兄，我們說，嘿，Claude，你會從零開始建構這個。

Two weeks later, Linux was working, Sequel Light was working, Redis was working.

兩週後，Linux 跑起來了，Sequel Light 跑起來了，Redis 跑起來了。

然後對，我們甚至玩了 Doom Loser。

Like, okay, so the real takeaway, which everybody should have, which is what is triggering me and making me so angry again at Anthropic, is that the real cool thing is that they figured out how to get 16 agents to mostly run by themselves for two weeks straight and actually produce something given the correct guidelines.

所以，重點來了，大家應該要知道的重點，也就是讓我很生氣、又對 Anthropic 感到憤怒的點，就是他們竟然能讓 16 個代理人自己跑兩個禮拜，而且在給予正確訓示後，真的能產出東西。

At the end of the day, hey, that's pretty cool.

It shows that models are in fact improving and can handle confidence, kind of context up to this size and able to kind of work together given the right orchestration.

這顯示模型確實正在進步，並且能夠處理到這個規模的信心、類型的上下文，並且在適當的協調下一起工作。

Like, that's a neat takeaway, but that's not what they said.

嗯，這是一個很棒的收穫，但這不是他們說的。

Which of course led to hilarious issues on GitHub.

這當然也導致了 GitHub 上一些爆笑的問題。

Like the very first issue, hello world does not compile.

像是第一個問題，hello world 無法編譯。

Because it turns out the example program that they put in their readme does not compile.

因為事實證明，他們放在 README 裡的範例程式根本無法編譯。

你根本沒提供 linker 啊，兄弟。

它不是 assembler，也不是 linger。

Okay, you think you just get those for free?

喔，你以為這些都是免費的喔？

Which, by the way, there's like 300 and some comments on here of just people arguing back and forth.

順帶一提，這裡大概有三百多則留言，大家就在那邊爭論不休。

And of course I'm not the only person saying this.

當然，說這話的也不只我一個人。

There are some other people saying, hey, this is kind of ridiculous.

有些人也在說，嘿，這有點太扯了吧。

This is a little out of, like out of pocket here, calling this from scratch and no human intervention.

這樣說有點離譜，說這是從零開始、完全沒有人類介入。

So at the end of the day, this could have been an article about how we got Claude to run for multiple weeks and produce something with very little to.

所以說到底，這本來可以是一篇關於我們怎麼讓 Claude 跑了好幾個星期，而且幾乎沒有人為介入就能產出東西的文章。

To almost none human interaction and being able to fulfill a spec.

But instead it became a spectacle about how anthropic, yet again is completely dishonest.

但結果卻變成了一場關於 Anthropic 又一次完全不誠實的鬧劇。

This is just another example of the AI hype cycle.

這只是 AI 炒作週期的又一個例子。

They're just here to try to raise money and to take every piece of novel work possible and sell it back to you.

他們只是想來籌錢，然後把所有新穎的作品都拿去賣給你。

Now, this is a future I'm personally excited about.

嗯，這是我個人很期待的未來。

If you were just, like, a little bit more honest about what you're doing, honestly, I think a lot of people would be much more excited about it and probably cheer you on from the sideline.

如果你們對自己做的事情誠實一點點，我真的覺得很多人會更興奮，甚至會在旁邊為你們加油。

There's a lot of cool stuff about AI we don't have to be AI haters.

AI 有很多很酷的地方，我們不必成為 AI 的仇恨者。

But sometimes you just say the stupidest crap ever and just do the stupidest stuff and make everybody hate you.

但你們有時候就是會說一些超蠢的話，做一些超蠢的事，讓大家討厭你們。

How about you just don't do that anymore?

你們乾脆不要再這樣做了，好不好？