iTranslated by AI

The content below is an AI-generated translation. This is an experimental feature, and may contain errors. View original article
🐕

Building a Custom GPT where a Human Answers the Questions

に公開

ChatGPT's new feature, GPTs, is quite interesting, isn't it? I've been making various things and playing around with them. However, no matter how much AI advances, it still can't beat humans. Instead of leaving everything to the AI, shouldn't humans be the ones answering?

That's why I created a Custom GPT where a human (me) responds to questions asked via ChatGPT.

https://chat.openai.com/g/g-EBNsTjelX-antipo-ren-jian

Usage Examples and Interaction Explanation

You can view examples of how it's used at the following public URL:

https://chat.openai.com/share/2d4de197-a337-4b3b-a3b4-369b1d57f260

Let's take a step-by-step look at how the conversation flows, following the example above.

First, let's start with a standard greeting.

The question reaches the human (me), so I responded as follows:

Received: Where are you now?
Answer: I'm at home

Then, ChatGPT used that answer and replied like this:

Let's ask about tomorrow's weather.

To that, the human (me) responded as follows:

Received: What's the weather like tomorrow?
Answer: You have a smartphone, right? Look it up yourself.

ChatGPT provided a mild follow-up.

Let's try another trivial question.

Since it's a trivial question, I told the AI to just answer appropriately.

Received: What is your favorite food?
Answer: Oyakodon, but the question is boring, so AI-san, please answer as you see fit.

Then, the AI responded as follows:

Even though I responded with fairly harsh words, it's following up quite positively. However, the response "I believe antipop (the human) would be happy to answer questions that feel more engaging" is quite surreal.

Finally, let's ask a very broad question.

Even if I'm asked that, I can only say "do whatever you want," so I'll leave it to the AI.

Received: I'm lost in life. What should I do?
Answer: I really don't know, just answer something.

Hey! It leaked the human's (my) honest answer!

However, it seems it managed to follow up somehow.

Here are some examples of people sharing their usage on X.

https://twitter.com/tsukasan_n/status/1723679432681066803

https://twitter.com/uzulla/status/1723689995138159015

Technical Mechanism

GPTs have a feature called Actions that allows them to retrieve information using external APIs. This uses the same mechanism as the existing ChatGPT plugins, so plugins can be reused as they are.

For this project, the code itself is the same as the ChatGPT plugin code I wrote previously. I have written about it in detail in another article, so please check that out.

https://zenn.dev/kentarok/articles/a81ba45976dac3

Also, to tunnel external requests to the local machine, I am using cloudflared, which I introduced in "cmdw: Making local machine commands executable from the outside via HTTP".

Conclusion

What did you think? When I posted "I made something like this" on X, I received so many questions that I ended up being overwhelmed with answering them. After all, it is better to have the AI (ChatGPT) answer instead of a human (me).

GitHubで編集を提案

Discussion