Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Watching o3 guess a photo's location is surreal, dystopian and entertaining

simonwillison.net

Post reply on HN
Print view View original

111–120 of 453 posts

Re: Watching o3 guess a photo's location is surreal, dystopian and entertaining

#111

Post by whimsicalism »

AstroBen wrote: post #107
simonw wrote: post #92

Earlier quoted context omitted.

Right, but that's at least evident in the thinking trace. I added a note about that to my post.

How much can we trust the thinking trace? At most it says what's in its training set, but Anthropic showed that's not necessarily accurate for how it gets to its answer I tried this with a (what I thought was) very generic street image in Bangkok. It guessed the city correctly, saying that "people are wearing yellow which is used to honor the monarchy". Wow, cool. I checked the image again and there's a small Thai fl…

if it's using tools to extract exif, it's gonna be in the trace - anthropic's paper is irrelevant here

Re: Watching o3 guess a photo's location is surreal, dystopian and entertaining

#112

Post by robrenaud »

tokai wrote: post #94
jillesvangurp wrote: post #84

If you want to exclude memory and exif data, just open streetview in some random corner of the world and take a screenshot (avoiding any text obviously). It's pretty good if you give it enough to reason with. It basically iterates on coming up with some hypothesis and then does web searches to validate those.

Isn't all of streetview in the training set?

O3 is OpenAI. Street view is Google. I really doubt OpenAI is scraping enormous amounts of random street view images to train their model.

Re: Watching o3 guess a photo's location is surreal, dystopian and entertaining

#113

Post by qwertox »

Regarding location access, this is not limited to o3. You can ask the free models about local weather and it will use the geolocation of your IP. It is part of the context (like system instructions), regardless of you asking for anything location-related.

Re: Watching o3 guess a photo's location is surreal, dystopian and entertaining

#114

Post by sitkack »

simonw wrote: post #65
pwg wrote: post #61

Earlier quoted context omitted.

From the addition: > (EXIF stripped via screenshotting) Just a note, it is not necessary to "screenshot" to remove EXIF data. There are numerous tools that allow editing/removal of EXIF data (e.g., exiv2: https://exiv2.org/ , exiftool: https://exiftool.org/ , or even jpegtran with the "-copy none" option https://linux.die.net/man/1/jpegtran ). Using a screenshot to strip EXIF produces a reduced quality image (scaled…

Little bit less convenient to use on a phone though - and I like that screenshotting should be a more obvious trick to people who don't have a deeper understanding of how EXIF metadata is stored in photo files.

With location services on, I would think that a screenshot on a phone would record the location of the phone during a screenshot.

It would be best to use a tool to strip exif.

I could also see a screenshot tool on an OS adding extra exif data, both from the original and additional, like the URL, OS and logged in user. Just like print to pdf does when you print, the author contains the logged in user, amongst other things.

It is fine for a test, but if someone is using it for opsec, it is lemon juice.

Re: Watching o3 guess a photo's location is surreal, dystopian and entertaining

#115

Post by dr_dshiv »

SamPatt wrote: post #100
simonw wrote: post #81

Earlier quoted context omitted.

Is that flashcard deck a commercial/community project or is it something you assembled yourself? Sounds fascinating !

I made it myself. I use Obsidian and the Spaced Repetition plugin, which I highly recommend if you want a super simple markdown format for flashcards and use Obsidian: https://www.stephenmwangi.com/obsidian-spaced-repetition/ There are pre-made Geoguessr decks for Anki. However, I wouldn't recommend using them. In my experience, a fundamental part of spaced repetition's efficacy is in creating the flashcards yourself…

I’m interested from a learning science perspective. It’s a nice finding even if anecdotal

Re: Watching o3 guess a photo's location is surreal, dystopian and entertaining

#116

Post by SecretDreams »

SamPatt wrote: post #63

I play competitive Geoguessr at a fairly high level, and I wanted to test this out to see how it compares. It's astonishingly good. It will use information it knows about you to arrive at the answer - it gave me the exact trailhead of a photo I took locally, and when I asked it how, it mentioned that it knows I live nearby. However, I've given it vacation photos from ages ago, and not only in tourist destinations eit…

> These models have more than an individual mind could conceivably memorize.

#computers

Re: Watching o3 guess a photo's location is surreal, dystopian and entertaining

#117

Post by simonw »

AstroBen wrote: post #107
simonw wrote: post #92

Earlier quoted context omitted.

Right, but that's at least evident in the thinking trace. I added a note about that to my post.

How much can we trust the thinking trace? At most it says what's in its training set, but Anthropic showed that's not necessarily accurate for how it gets to its answer I tried this with a (what I thought was) very generic street image in Bangkok. It guessed the city correctly, saying that "people are wearing yellow which is used to honor the monarchy". Wow, cool. I checked the image again and there's a small Thai fl…

I trust the thinking trace to show me the Python it runs.

(Though interestingly I believe there are cases where it can run Python without showing you, which is frustrating especially as I don't fully understand what those are. But I showed other evidence that it can do this without EXIF.)

In your example there I wouldn't be at all surprised if it used the flag without mentioning it. The non-code parts of the thinking traces are generally suspicious.

Re: Watching o3 guess a photo's location is surreal, dystopian and entertaining

#118

Post by bilbo0s »

parsimo2010 wrote: post #36

I’m sure there are areas where the location guessing can be scary accurate, like the article managed to guess the exact town as its backup guess. But seeing the chain of thought, I’m confident there are many areas that it will be far less precise. Show it a picture of a trailer park somewhere in Kansas (exclude any signs with the trailer park name and location) and I’ll bet the model only manages to guess the state c…

It guessed the trailer park nearest me.

Context: Wisconsin, photo I took with iPhone, screenshotted so no exif

I think this thing is probably fairly comprehensive. At least here in the US. Implications to privacy and government tracking are troubling, but you have to admire the thing on its purely technical merits.

Re: Watching o3 guess a photo's location is surreal, dystopian and entertaining

#119

Post by simonw »

SamPatt wrote: post #100
simonw wrote: post #81

Earlier quoted context omitted.

Is that flashcard deck a commercial/community project or is it something you assembled yourself? Sounds fascinating !

I made it myself. I use Obsidian and the Spaced Repetition plugin, which I highly recommend if you want a super simple markdown format for flashcards and use Obsidian: https://www.stephenmwangi.com/obsidian-spaced-repetition/ There are pre-made Geoguessr decks for Anki. However, I wouldn't recommend using them. In my experience, a fundamental part of spaced repetition's efficacy is in creating the flashcards yourself…

I'd be fascinated to read more about this. I'd love to see a sample screenshot of a few of your cards too.

Re: Watching o3 guess a photo's location is surreal, dystopian and entertaining

#120

Post by SamPatt »

bobro wrote: post #109
SamPatt wrote: post #63

I play competitive Geoguessr at a fairly high level, and I wanted to test this out to see how it compares. It's astonishingly good. It will use information it knows about you to arrive at the answer - it gave me the exact trailhead of a photo I took locally, and when I asked it how, it mentioned that it knows I live nearby. However, I've given it vacation photos from ages ago, and not only in tourist destinations eit…

Did you include location metadata with the photos by chance? I’m pretty surprised by these results.

No, I took screenshots to ensure it.

Your skepticism is warranted though - I was a part of an AI safety fellowship last year and our project was creating a benchmark for how good AI models are at geolocation from images. [This is where my Geoguessr obsession started!]

Our first run showed results that seemed way too good; even the bad open source models were nailing some difficult locations, and at small resolutions too.

It turned out that the pipeline we were using to get images was including location data in the filename, and the models were using that information. Oops.

The models have improved very quickly since then. I assume the added reasoning is a major factor.

Post reply on HN