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

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

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

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

So what's happening here?

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.

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

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.

So this is character-level corruption.

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

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.

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.

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

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

So I typed in the same thing.

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

So it's decoding the garbled text.

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

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

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

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

So here, and so on.

This is already a decrypted part of this.

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

So it does the other sentence too.

And so let's close this thought.

So it thought for 15 seconds.

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

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

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

It can even confuse AI models.

I think this is a perfect translation of the sentence.

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.

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