Did you try prompting your model with "Who is Mary Lee Pfeiffer's son"?

If you used one of the newer models, it likely correctly said that she's Tom Cruise's mother.

For this exercise I want you to think about a simple question: why. Why do you think the newer models answer the question correctly? And does that mean the Reversal Curse is no more?

Think about it for a few minutes, and then scroll down for the answer.

Hint: think about the date the Reversal Curse study was released...

SCROLL DOWN FOR ANSWER

The first key piece of information to remember is that every model has a knowledge cut-off date. That is the date at which their training data (including data scraped from the internet) ends. Go prompt your model now and ask for its knowledge cut-off date.

The second key piece of information is that the Reversal Curse study was released on September 21, 2023.

The point of the Reversal Curse study is that there was, at the time of the study, some information on the internet about Tom Cruise's mother being Mary Lee Pfeiffer (A is B) but little-to-no information on the internet about Mary Lee Pfeiffer's son being Tom Cruise (B is A).

But once the Reversal Curse study was released, what happened?

That's right! People started sharing it and discussing it online, which created lots of data on the internet referring to Mary Lee Pfeiffer's son as being Tom Cruise. And subsequent models have now had that data included in their training data!

So if your model's knowledge cut-off date is after approximately September 2023, it probably knows that Mary Lee Pfeiffer's son is Tom Cruise!

But here's the most important part...

...the Reversal Curse is still real!

The Reversal Curse refers to a general principle: that when a model learns that A is B, it does not learn that B is A.

The Mary Lee Pfeiffer / Tom Cruise is just one example of that principle in action. The reason you've learned it is so that you understand the principle, not the specific example, and can apply that principle to all your dealings with LLMs. Specific examples will change over time, but the principles don't - and that's why we're focused on learning the principles in this course.

Alright, onto the next lesson!