It's just counting the letter R's in a word strawberry.

So let's start with the traditional, like existing model GPT-4.0.

So as you can see the model fails on this.

There are three R's, but the model says there are only two R's.

So why does this advanced model like GPT-4.0 make such a simple mistake?

That's because models like this are built to process the text, not with the characters or words.

It's somewhere between, sometimes called a sub-word.

So if we ask the question to a model, a question that involves understanding the notion of characters and words, the model can really just make mistakes because it's not really built for that.

So now let's go on to our new model and type in the same problem.

This is the O1 preview, which is a reasoning model.

So unlike the GPT-4.0, it starts thinking about this problem before outputting the answer.

And now it outputs the answer.

There are three R's in the word strawberry.

So that's the correct answer.

And this example shows that even for seemingly unrelated counting problem, having a reasoning built in can help avoiding the mistakes because it can maybe look at its own output and review it and more just be more careful and so on.