Quick Hit: ChatGPT as Personal Tour Guide
The Quick Pitch:
*“We use ChatGPT to generate an on-the-fly personal tour guide as you walk through major tourist locations”
This is a combination of using your location, a decent-enough database of major tourist sites, and seeding ChatGPT with enough data to produce a reasonable 45 - 90 sec introduction to the site, and maybe another one to two minutes of things to look for.
The underlying assumption is that ChatGPT inherently “knows” just about everything that is in Wikipedia and likely all of the specialist literature as well. It has already proven relatively decent at compiling good lists of essential shoes for a man, so, in theory, it should be reasonable at talking about famous works of art, or historic locations.
The next step is to integrate this into a cell phone app which manages things like where you are and when it asks for a new guide “snippet.” You might even be able to use ChatGPT to build a walking tour; since it has state, and should know
Open Questions
-
How good are the snippets that ChatGPT creates? Here is a place where we want to avoid “confident but wrong,” and have the narrator to be more explicit about what is widely substantiated and what is harvested from the depths of the internet.
-
Can we use ChatGPT to also generate the tour itself? There are related design questions here, if you want it to be set up before you start, or something more free-form. (Or, how well can ChatGPT keep goals in mind. If it is good at goals, then “keep the entire tour under 2 hours” would be reasonable.
-
How “deep” is ChatGPT’s knowledge (and modern cell phone location)? Can we get down to the painting level in a museum? Or is it going to be mostly major attractions?
I’m playing a bit as I’m writing this, and my overall feeling is that what ChatGPT is generating isn’t bad. For a real product, we would probably do some light editing, but using the output would really accelerate seeding the database.
I’m going to have to think this idea through a little more. It’s definitely got some potential here. As I think about it, there is a follow-up to question three above: maybe this is going to work best for specialist content. We can tune the response to cover, for instance, a transportation buff’s interests versus a military buff’s interests.