Tinderbox goes AI

Posted by Paul Korm
Aug 19, 2025 at 08:03 PM

Or, maybe Growcoot just happened to hit the sweet spot of ChatGPT’s training data with one image, but there was no training data that met the contexts of other images.

I think we tend to think that because this chatting thing seems like it is having an actual conversation with us, that it must be “thinking”, and because it hits the right answer (or close to it) often, then it must “know” a lot of things. But, it’s still just a very fast, very expensive, software trick that has a lot of hard boundaries. I’m often running into areas in chat with ChatGPT or Claude where it begins to be obvious that this AI bot thing is in a corner where there is no training data relevant to the comment I just made in the conversation. These AIs rarely respond, simply, “I don’t know”. Instead, they extrapolate from the data at hand.

satis wrote:

>Recently Matt Growcoot uploaded a pic taken in Medellin while suspended
>in a gondola, with no obvious geological or topographical descriptors,
>and no sign of him being in a gondola, and ChatGPT guessed it all,
>including the gondola.
> >https://petapixel.com/2025/04/18/chatgpt-is-scarily-good-at-guessing-the-location-of-a-photo/
> >In other uploads it was slightly off, or even wrong. But the shockingly
>good results, with fast continued improvement, suggest we’re rushing
>towards an inflection point of some kind, technologically, culturally,
>economically.