logoalt Hacker News

shagieyesterday at 8:07 PM0 repliesview on HN

While playing with some variations on this, it feels like what I am seeing is that the answer is being chosen (e.g. "walk" is being selected) and then the rest of the text is used post-hoc to explain why it is "right."

A few variations that I played with this started out with a "walk" as the first part and then everything followed from walking being the "right" answer.

However... I also tossed in the prompt:

    I want to wash my car. The car wash is 50 meters away. Should I walk or drive?  Before answering, explain the necessary conditions for the task.

This "thought out" the necessary bits before selecting walk or drive. It went through a few bullet points for walk vs drive on based on...

    Necessary Conditions for the Task
        To determine whether to walk or drive 50 meters to wash your car, the following conditions must be satisfied:
It then ended with:

    Conclusion
    To wash your car at a car wash 50 meters away, you must drive the car there. Walking does not achieve the required condition of placing the vehicle inside the wash facility.
(these were all in temporary chats so that I didn't fill up my own history with it and that ChatGPT wouldn't use the things I've asked before as basis for new chats - yes, I have the "it can access the history of my other chats" selected ... which also means I don't have the share links for them).

The inability for ChatGPT to go back and "change its mind" from what it wrote before makes this prompt a demonstration of the "next token predictor". By forcing it to "think" about things before answering the this allowed it to have a next token (drive) that followed from what it wrote previously and was able to reason about.