So here I pasted in the prompt, and I'm asking the model to translate this badly corrupted Korean sentence to English.

是以，我在這裡粘貼了提示，要求模型將這個嚴重損壞的韓語句子翻譯成英語。

And as you can see, this is not an invalid Korean sentence.

正如你所看到的，這並不是一個無效的韓語句子。

So let's start with the existing model, GPT-40, and see how it does.

是以，讓我們從現有型號 GPT-40 開始，看看它的表現如何。

The model is just not able to understand this text, which is a valid response because this is not a valid language.

模型只是無法理解這段文字，這是一個有效的回答，因為這不是一種有效的語言。

So what's happening here?

這裡發生了什麼？

So Korean is an interesting language in that when we form a character, we can combine vowels and consonants, sometimes put the consonant at the bottom, and so on.

是以，韓語是一種有趣的語言，當我們組成一個字元時，我們可以將元音和輔音結合起來，有時還可以將輔音放在底部，等等。

One way to corrupt this character is to add in some extra unnecessary consonants to it.

破壞這種字元的一種方法是在其中添加一些不必要的輔音。

And that combination is so unnatural to native speakers, so they can just, when they see it, just automatically undo that change and understand the text.

對於母語使用者來說，這種組合非常不自然，所以他們看到這種情況時，就會自動撤銷這種改動，然後理解文本。

So this is character-level corruption.

所以，這是人物級別的腐敗。

We can do that at the phrase level, we can also do that at the sound level, and so on.

我們可以在詞組層面做到這一點，也可以在聲音層面做到這一點，等等。

People have come up with various methods like this, and I found it really interesting, so I adopted a few of them to create this example.

人們想出了各種類似的方法，我覺得非常有趣，是以採用了其中的一些方法來創建這個示例。

So if you understand Korean, this part that I'm highlighting now, you can read it off as a I'm not going to read off the whole thing, but this is the idea.

所以，如果你懂韓語，我現在強調的這部分內容，你可以把它作為一個我不打算讀完的內容來讀，但這就是大意。

Koreans can read it, but the models find it so difficult to understand.

韓國人可以讀懂，但模特們卻很難理解。

So now let's go on to our new model, O1 Preview, and see if reasoning can help solve this problem.

現在，讓我們來看看我們的新機型 O1 Preview，看看推理是否能幫助我們解決這個問題。

So I typed in the same thing.

於是我輸入了同樣的內容。

Unlike GPT-4.0, this model starts thinking through this problem before outputting the answer.

與 GPT-4.0 不同的是，該模型在輸出答案之前就開始思考這個問題。

So it's decoding the garbled text.

所以它在解碼亂碼文本。

So that's actually the right approach, because I gave a translation task, but the underlying task is actually decoding this problem.

是以，這種方法其實是正確的，因為我給出了一個翻譯任務，但根本任務其實是解碼這個問題。

So it started off with the right path, and then I'm examining this text, deciphering the text.

是以，一開始我走的是正確的道路，然後我在研究這段文字，解讀這段文字。

Deciphering is actually the right verb to use here, enhancing the translation.

實際上，"破譯 "是一個正確的動詞，可以增強翻譯的效果。

And then actually, it starts unpacking some part of it.

然後，實際上，它開始解開其中的一部分。

So here, and so on.

這裡也是，等等。

This is already a decrypted part of this.

這已經是解密的一部分了。

And then once the model figures it out, that part, everything else is just easy enough.

然後，一旦模型弄明白了這部分，其他一切就都很簡單了。

So it does the other sentence too.

另一句也是如此。

And so let's close this thought.

讓我們結束這一思考吧。

So it thought for 15 seconds.

所以它想了 15 秒鐘。

The final translation, the model output is, No translator on earth can do this, but Koreans can easily recognize it.

最後的翻譯、模型輸出是：世界上沒有一個翻譯能做到這一點，但韓國人很容易就能識別出來。

There's a method of encrypting Hangul by inputting various transformations of vowels and consonants.

有一種通過輸入元音和輔音的各種變換來加密韓文的方法。

It creates a way to make it look different on the surface.

它創造了一種方法，讓它在表面上看起來與眾不同。

It can even confuse AI models.

它甚至會迷惑人工智能模型。

I think this is a perfect translation of the sentence.

我認為這句話翻譯得非常完美。

So this illustrates how general purpose reasoning models like O1 Preview can help seemingly unrelated questions like this, which is almost like a code cracking.

是以，這說明了像 O1 預覽這樣的通用推理模型是如何幫助像這樣看似不相關的問題的，這幾乎就像是密碼破解。

So I hope this illustrates how reasoning can be a powerful tool for solving your problems.

是以，我希望這能說明推理是如何成為解決問題的有力工具的。