🧵 View Thread
🧵 Thread (28 tweets)
I’ve independently verified this, twice, and have to spoil the thread to say “Any photo outdoors can be trivially location matched in a fashion you’d assume would imply capabilities of an intelligence agency. If that is news to you, feel free to incorporate it into decisions.”
@harryh Yes; screenshotted to strip EXIF data and accessing from an IP in Chicago re: photos in Florida and Tokyo. (It did not articulate usage of knowledge I’m associated with my Tokyo neighborhood to inform guess, but that fact is abundantly publicly available.)
@patio11 Recent viral thread with a ton of examples: https://t.co/tfGozthH6r
@patio11 One of the takeaways was if your photo contains any mountains, the model is *extremely* good, i.e. it seems to be able to identify basically any ridgeline in the world https://t.co/DpoqXPzc7Y
@patio11 If you have an iPhone, check your photo app settings. Apple (and likely everyone else) is building/maintaining a database of photos to match features to known coordinates. This reverse location lookup will only get more accurate over time.
The prompt is very good, but I tried using this exact prompt and photo and O3 kept guessing Cape Cod for me https://t.co/mKCwOROeQJ Which it does look a lot like the National Seashore, but that's also relatively local for me. o3 apparently guesses the West coast more often for people on the west coast.
@KelseyTuoc I just tried again with your prompt (which is excellent) and it narrowed it down to Cannon Beach, Oregon or Nauset Cape Cod, and went with the Cape after studying the pole and doing a lot of searches for that detail. I have to agree, this looks exactly like the outer cape :) https://t.co/PATTSsfsP5
o3 deep research got this location within 25 miles. Even verified I stripped all metadata before uploading. Its thought process spent a lot of time thinking about tree species distribution and snow depth climatology. It did also give a list of possible places rather than a single location, but the first place it listed as the most likely was the correct answer.
@patio11 I tried this vulture photo, which is from an obscure tiny wildlife refuge, hard to geolocate. I'm watching o3 go to crazy lengths to pin down the location -- writing OCR script in python, reasoning about sign text (which is a red herring), &c Finally picked one ~200 miles away https://t.co/88XVi2wWyJ
I spent an hour last night going thru my photos where I had perfect recall or GPS metadata to test with o3. It got it 65% of the time, roughly. I was ACTIVELY trying to trick it. The ones that were failures, were often directionally correct. The worst failure was an airplane photo of the ground, which they pegged as on the way to SeaTac, instead of on the way to SFO. Most failures were off by no more than 3-4 miles.
@patio11 I just gave it a pretty detailed photo of downtown Berkeley taken from my front porch and it was unable to identify the location. It's interesting how even in this narrow domain it's effectiveness is very spikey/hit-and-miss.