12
Watching o3 guess a photo’s location is surreal, dystopian and wildly entertaining
(simonwillison.net)
Technology for a Solar-Punk future.
Airships and hydroponic farms...
I'm going to leave this here. TL;DR modern AI chatbots have learned the same cold reading techniques as fake psychics. This is almost textbook: A generic guess followed by some more specific guesses, framed up in a way that the person will automatically glom onto the most accurate of the three and conclude that the thing is a genius whichever one is correct, and happily subconsciously add things like "200 miles is not a big deal" to the model when it needs the help, and not regard it as all that big a deal if the thing is just 1,000% wrong some of the time. And also, it will not mention when it is digging through EXIF tags or its prompted location data. Of course. It's not trained for honesty, it's specifically trained to generate pleasingly misleading answers.
For all I know, the "chain of thought" stuff is going to evolve into AGI and this location identification is working gangbusters. I have no idea, I haven't tried. But be careful with assigning success to this stuff.
The model output also circles around the initial guess. So the final result central ca is also the one shot result without the "reasoning" non sense. So instead of wasting, who knows how much power, a much simpler model to determine the vegetation would have given the same result.