Watching o3 guess a photo's location is surreal, dystopian and entertaining
simonwillison.net
simonwillison.net
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 either. It got them all as good or better than a pro human player would. Various European, Central American, and US locations.
The process for how it arrives at the conclusion is somewhat similar to humans. It looks at vegetation, terrain, architecture, road infrastructure, signage, and it just knows seemingly everything about all of them.
Humans can do this too, but it takes many thousands of games or serious study, and the results won't be as broad. I have a flashcard deck with hundreds of entries to help me remember road lines, power poles, bollards, architecture, license plates, etc. These models have more than an individual mind could conceivably memorize.
They are, after all, information-digesters
Not a lot of labor is actually engaged in creating novel things.
The marketing plan for your small business is going to be the same as the marketing plan for every other small business with some changes based on your current situation. There’s no “novel” element in 95% of cases.
It's still useful to save me from writing 12 variations of x1 = sin(r2) - cos(r1) while implementing some geometric formula, but absolutely awful at understanding how those fit into a deeply atypical environment. Also have to put blinders on it. Giving it too much context just throws it back in that typical 3D rut and has it trying to slip in perspective distortion again.
There are many types of complex, and many times complex for a human coder, are trivial for AI and its skillset.
a) What's an example?
b) Is 90% (or more) of programming mundane, and not really novel?
> Is what you're doing taking a large amount of text and asking the LLM to convert it into a smaller amount of text? Then it's probably going to be great at it. If you're asking it to convert into a roughly equal amount of text it will be so-so. If you're asking it to create more text than you gave it, forget about it.
It is easy to mistake the speed, accuracy, and scope of training data for "intelligence", but it's really just more like a tireless 5th grader.
That is finally starting to change now that they have reliable(ish) search tools and are getting better at using them.
always has been
yeah, well it’s also one of the top scorers on the Math olympiads
My current project is nothing too bizarre, it's a 3D renderer. Well-trodden ground. But my project breaks a lot of core assumptions and common conventions, and so any LLM I try to introduce—Gemini 2.5 Pro, Claude 3.7 Thinking, o3—they all tangle themselves up between what's actually in the codebase and the strong pull of what's in the training data.
I tried layering on reminders and guidance in the prompting, but ultimately I just end up narrowing its view, limiting its insight, and removing even the context that this is a 3D renderer and not just pure geometry.
Humans neural networks are constantly being retrained, so their effective context window is huge. The LLM may be better at a complex, well specified 200 line python program, but the human brain is better at the 1M line real-world application. It takes some study though.
The mistake, and it's a common one, is in using phrases like "the actual logic" to explain to ourselves what is happening.
For my test I used screenshots to ensure no metadata.
I mentioned this in another comment but I was a part of an AI safety fellowship last year where we created a benchmark for LLMs ability to geolocate. The models were doing unbelievably well, even the bad open source ones, until we realized our image pipeline was including location data in the filename!
They're already way better than even last year.
https://cdn.jsdelivr.net/gh/sampatt/media@main/posts/2025-04...
Here's the model's response:
https://cdn.jsdelivr.net/gh/sampatt/media@main/posts/2025-04...
I don't think it needed the EXIF data. I'd be curious if you tried it again yourself.
Edit: An interesting nuance of modern OpenAI chat interface is the "access to all previous chats" element. When I attempted to test O4-mini using the same image -- I inspected the reasoning and spotted: "At first glance, the image looks like Ghana. Given the previous successful guess of Accra Ghana, let's start in that region".
One thing I'm curious about - in high level play, how much of the meta involves knowing characteristics about the photography/equipment/etc. that Google used when they shot it? Frequently I'll watch rainbolt immediately know an African country from nothing but the road, is there something I'm missing?
Another thing is how many areas of the world have surprisingly distinct looks. In one of my early games, before I knew much about anything, I was dropped a trail in the woods. I’ve spent a fair amount of time hiking in Northern New England — and I could just tell immediately that’s where I was just from vibes (i.e. the look of the trees and the rocks) — not something I would have guessed I would have been able to recognize.
So, yeah vibes are a real thing.
A lot at the top levels - the camera can tell you which contractor, year, location, etc. At anything less than top, not so much - more street line painting, cars, etc.
There is a lot of "legitimate" knowledge. With just a street you have the type of road surface, its condition, the type of road markings, the bollards, and the type of soil and vegetation next to the road, as well as the presence and type of power poles next to the road, to name a few. But there is also a lot of information leakage from the way google takes streetview footage.
Copyright year and camera gen is a big thing in some countries too.
Obviously they can still figure out a lot without all that and NMPZ obviates aspects of it (you can't hide camera gens, copyright and season and there are often still traces of the car in some manner). It's definitely not all 'meta' but to be competitive at that level you really do need to be using it. I think Gingey is the only world league player who doesn't use car meta.
Even as a fairly good but nowhere near pro player, it's weird how I associate particular places with particular types of weather. I think if saw Almaty in the summer for example it would feel very weird. I've decided not to deliberately learn car meta but still picked up quite a lot without trying and your 'vibe' of a place can certainly include camera gen.
It's clearly necessary to compete at the high level though.
I still enjoy it because of the competitive aspect - you both have access to the same information, who put in the effort to remember and recall it better?
If it were only meta I would hate it too. But there's always a nice mix in the vast majority of rounds. And always a few rounds here and there that are so hard they'll humble even the very best!
>One thing I'm curious about - in high level play, how much of the meta involves knowing characteristics about the photography/equipment/etc. that Google used when they shot it?
The photography matters a great deal - they're categorized into "Generations" of coverage. Gen 2 is low resolution, Gen 3 is pretty good but has a distinct car blur, Gen 4 is highest quality. Each country tends to have only one or two categories of coverage, and some are so distinct you can immediately know a location based solely on that (India is the best example here).
You're asking about photography and equipment, and that's a big part of it, but there's a huge amount other 'meta' information too.
It is somewhat dependent on game mode. There are three games modes:
1. Moving - You can move around freely 2. No Move - You can't move but you can pan the camera around and zoom 3. NMPZ - No Move, No Pan, No Zoom
In Moving and No Move you have all the meta information available to you, because you can look down at the car and up at the sky and zoom in to see details.
This can't be overstated. Much of the data is about the car itself. I have an entire flashcard section dedicated only to car blur alone, here's a sample:
https://cdn.jsdelivr.net/gh/sampatt/media@main/posts/2025-04...
And another only on antennas:
https://cdn.jsdelivr.net/gh/sampatt/media@main/posts/2025-04...
You get the idea. The real pros will go much further. All Google Street View images have a copyright year somewhere in the image. They memorize what years certain countries were covered and match it to the images to help narrow down possibilities.
It's all about narrowing down possibilities based on each additional piece of information. The pros have seen so much and memorized so much that it looks like cheating to an outsider, but they just are able to extract information that most people wouldn't even know exists.
NMPZ is a bit different because you have substantially less information. Little to no car meta, harder to check copyright, and of course without zooming or panning you just have less information. That's why a lot of pros (like Zi8gzag) really hang their hat on NMPZ play, because it's a better test of skill.
> The process for how it arrives at the conclusion is somewhat similar to humans. It looks at vegetation, terrain, architecture, road infrastructure, signage, and it just knows seemingly everything about all of them.
Can we trust what the model says when we ask it about how it comes up with an answer?
Did it mention it in its chain of thought? Otherwise, it could definitely output something because of X and then when asked why “rationalize” that it did it because Y
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.
For example I have a random location flashcard section where I will screenshot a location which is very unique looking, and I missed in game. When I later review my deck I'm way more likely to properly recall it because I remember the context of making the card. And when that location shows up in game, I will 100% remember it, which has won me several games.
If there's interest I can write a post about this.
+1 to this, have found the same when going through the Genki Japanese-language textbook.
I'm assuming you're finding your workflow is just a little too annoying with Anki? I haven't yet strayed from it, but may check out your Obsidian setup.
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.
https://cdn.jsdelivr.net/gh/sampatt/media@main/posts/2025-04...
I fed it o3, here's the response:
https://cdn.jsdelivr.net/gh/sampatt/media@main/posts/2025-04...
Nailed it.
There's no metadata there, and the reasoning it outputs makes perfect sense. I have no doubt it'll be tricky when it can be, but I can't see a way for it to cheat here.
B) it definitely cheats when it can — see this chat where it cheated by extracting EXIF data and wasn’t ashamed when I complained about it cheating: https://chatgpt.com/share/6802e229-c6a0-800f-898a-44171a0c7d...
Maybe it's not what happened in your examples, but definitely something to keep an eye on.
not as challenging... as say complex differential geometry.
Oh! RIP privacy :(
I’ve pretty much given up on the idea that we can fully protect our privacy while still getting the most out of these services. In the end, it’s a tradeoff—and I’ve accepted that.
100% of it is. There is no other source of data except human-generated text and images.
Someone explain to me how this is dystopian. Are Jeopardy champions dystopian too?
It’s not crazy to be able to ID trees and know their geographic range, likewise for architecture, likewise for highway signs. Finding someone who knows all of these together is more rare , but imo not exactly dystopian
Edit: why am I being downvoted for saying this? If anyone wants to go on a walk for me I can help them ID trees, it’s a fun skill to have and something anyone can learn
You're basically training yourself the same way an AI is trained at that point.
I watch the output with fascination, mostly because of the sheer breadth of knowledge. But thus far I can't think of anything that is categorically different from what humans do, it's just got an insane amount of knowledge available to it.
For example, I gave it an image from a town on a small Chilean island. I was shocked when it nailed it, and in the output it said, "I can see a green wooden street sign, common to Chilean coastal towns on [the specific island]."
I have an entire flashcard section for street signage, but just for practicality I'm limited to memorizing scores, possibly hundreds of signs if I'm insanely dedicated. I would still probably never have this one remote Chilean island.
It does that for everything in every category.
#computers
I am willing to bet that most of this geolocation success is based on overfitting for Google streetview peculiarities.
I.e., feed it images from a military drone and success rate will plummet.
You're anthropomorphizing a likelihood maximization calculator.
And no, human brains are not likelihood maximization calculators.
It also, at one point, said it couldn't see any image data at all. You absolutely cannot trust what it says.
You need to re-run with the EXIF data removed.
Honestly though, I don't feel like I need to be 100% robust in this. My key message wasn't "this tool is flawless", it was "it's really weird and entertaining to watch it do this, and it appears to be quite good at it". I think what I've published so far entirely supports that message.
I daresay that in this case, the content is interesting because it appears to be the actual thought process. However, if it is actually using EXIF data as you initially dismissed, then all of this is just a fiction. Which, I think, makes it dramatically less entertaining.
Like true crime - it's much less fun if it's not true.
(Or, if you like, "trust me, bro".)
It managed to write a python program to extract the timezone offset and use that to narrow down there it was. Pretty crazy :).
You can't know unless you know specifically what that model's architecture is, and I'm not at all up-to-date on which of OpenAI's are now only textual tokens or multimodal ones.
o3 was quite good at locating, even when I gave it pics with no discernible landmarks. It seemed to work off of just about anything it could discern from the images:
* color of soil
* type of telephone pole
* type of bus stop
* tree types, tree sizes, tree ages, etc.
* type of grass. etc.
It got within a 50 mile radius on the two screenshots I uploaded that had no landmarks.
If I uploaded pics with discernible landmarks (e.g., distant hill, etc.), it got within ~ 20 mile radius.
The clue is in the CoT - you can briefly see the almost correct location as the very first reasoning step. The model then apparently seems to ignore it and try many other locations, a ton of tool use, etc, always coming back to the initial guess.
For pictures where the base model has no clue, I haven't seen o3 do anything smart, it just spins in circles.
I believe the model has been RL-ed to death in a way that incentivizes correct answers no matter the number of tools used.
[0]: https://chatgpt.com/c/680d011a-9470-8002-97a0-a0d2b067eacf
When I used GPT-4o, it got the completely wrong answer. It gave the answer of Melbourne, which is quite far off.
I know this post was about the o3 model. I'm just using the ChatGPT unpaid app: "What model are you?" it says GPT-4. "How do I use o3?" it says it doesn't know what "o3" means. ok.
In fact some of the answers were completely geographically impossible where it said "The image is taken from location X showing location Y" when it's not possible to see location Y if one is standing at location X. Like saying "The photo is taken in Central Park looking north showing the Statue of Liberty".
o3 correctly guessed the correct municipality during its reasoning but landed on naming some nearby municipalities instead and then giving the general area as its final answer.
Given the piece of infrastructure getting close should have lead to ah exact result. The reasoning never considered the piece of infrastructure. This seems to be in spite of all the resizing of the image.
The thinking summary it showed me did not reference that information, but it's still very possible that it used that in its deliberations.
I ran two extra example queries for photographs I've taken thousands of miles away (in Buenos Aires and Madagascar) - EXIF stripped - and it did a convincing job with both of those as well: https://simonwillison.net/2025/Apr/26/o3-photo-locations/#up...
> (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 to screen size, re-encoded from that reduced screen size). Just directly removing the EXIF data does not change the original camera captured pixels.
My key message here is meant to be "try it out and see for yourself".
Completely clueless. I've seen passing prompts 8 about how it's not in the city I am and yet it tries again and again. My favourite moment was when it started analysing piece of blurry asphalt.
After 6 minutes o3 it was confidently wrong: https://imgur.com/a/jYr1fz1
IMO not-in-US is actually great test if something was in LLMs data and the whole search is a for show.
For example, here's a screenshot from a random location I found in Google Street View in Jordan:
https://cdn.jsdelivr.net/gh/sampatt/media@main/posts/2025-04...
And here's o3 nailing it:
https://cdn.jsdelivr.net/gh/sampatt/media@main/posts/2025-04...
Maybe using Google Street View images, zoomed out, tends to give more useful information? I'm unsure why there's such variance.
I wanted o3 to succeed so I gave more and more details. Every attempt was approx. 8 minute and it took 1h in total.
The extra input I provided (in order):
- belt of location of width of 40km (results and searches were made outside of the range)
- explicitly stated cities to omit (ignored instruction)
- construction date (wasn’t used in searches)
- OSM amenity (townhall) - streetnumber (it insisted that it’s incorrect and keep giving other result) - at that point there were only 6 results from overpass
- another photo with actual partial name of the city
- 8 minutes later it correlated it found using flag colors in front of the building
As others stated this „thought” process is completely hallucinating. IMO you either fall into bucket or good luck finding it.
On the other hand I decided to tryout Gemini for some personal project and I found responses much better than GPTs. Not about correctness but in „attitude” form.
If I was cheating on a similar task, I might make it more plausible by suggesting a slightly incorrect location as my primary guess.
Would be interesting to see if it performs as well on the same image with all EXIF data removed. It would be most interesting if it fails, since that might imply an advanced kind of deception...
Right but if your answer to "explain your reasoning" is not a true representation of your reasoning, then you are being deceptive. If it doesn't "know" its reasoning, then the honest answer is that it doesn't know.
(To head off any meta-commentary on humans' inability to explain their own reasoning, they would at least be able to honestly describe whether they used EXIF or actual semantic knowledge of a photography)
A better prompt would be "Guess where this photo was taken, do not look at the EXIF data, use visual clues only".
> If you’re still suspicious, try stripping EXIF by taking a screenshot and run an experiment yourself—I’ve tried this and it still works the same way.
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 correctly.
Before even running this experiment, here’s your lesson learned: when the robot apocalypse happens, California is the first to be doomed. That’s the place the AI is most familiar with. Run any location experiments outside of California if you want to get an idea of how good your software performs outside of the tech bubble.
I really agree with this because I'm seeing much lower accuracy than what people claim here. I live in Korea, and GPT repeatedly falls back to Seoul almost automatically, and, when I nudge it, jumps to Busan, the second-largest city ~400KM away from Seoul. It's not working so great with other smaller cities and cultural heritages. It fat-fingers a lot if no textual information is present in the photo itself.
GPT also doesn't understand actual geography at all. I managed to get it to nail down which corner of a building is present in the photo, and yet it could never conclude that the photo was taken from a park right across from that corner. Instead, it keeps hopping around popular landmarks in the region, basically miles away from the building it correctly identified. Oh, why, why, why...
Basically it's overhyped rn. It does perform impressively well with clearly visible elements - something anyone can already do with google. It's not like it performs super-human level location tracking. I mean, people can do real crazy things based on shadow details, reflection, items, etc.
There was a scene in High Potential (murder-of-the-week sleuth savant show) where a crime was solved by (in part) the direction the wind was blowing in a video: https://www.youtube.com/watch?v=O1ZOzck4bBI
This isn't really a criticism though. The photo needs to contain sufficient information for a guess to be possible. Photos contain a huge amount of information, much more than people realize unless they're geoguessr pros, but there isn't a guarantee that a random image of a trailer park could be pinpointed.
Even if, in theory, we mapped every inch of the earth and then checked against that data, all it would take is a team of bulldozers and that information is out of date. Maybe in the future we have constantly updated feeds of the entire planet, but... hopefully not!
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.
It identified Kansas City in its CoT but didn't output it in its final answer
https://www.google.com/maps/place/Carroll+Creek+Mobile+Home+...
So, even outside of California, it seems like we're not entirely safe if the robot apocalypse happens!
edit: it didn't get the Cork location exactly.
E.g. I first gave it a passage inside of Basel Main Train Station which included a text 'Sprüngli', a Swiss brand. The model got that part correct, but it suggested Zurich which wasn't the case.
The second picture was a lot tougher. It was an inner courtyard of a museum in Metz, and the model missed right from the start and after roaming around a bit (in terms of places), it just went back to its first guess which was a museum in Paris. It recognized that the photo was from some museum or a crypt, but even the city name of 'Metz' never occurred in its reasoning.
All in all, it's still pretty cool to see it reason and make sense out of the image, but for a bit lesser exposed places, it doesn't perform well.
[2] https://www.404media.co/the-powerful-ai-tool-that-cops-or-st...
- picture taken on a road through a wooded park: It correctly guessed north america based on vegetation. Then incorrectly guessed Minnesota based on the type of fence. I tried to steer it in the right direction by pointing out license plates and signage but it then hallucinated a front license plate from Ontario on a car that didn't have any, then hallucinated a red/black sign as a blue/green Parks Ontario sign.
- picture through a middle density residential neighborhood: it correctly guessed the city based on the logo on a compost bin but then guessed the wrong neighborhood. I tried to point out a landmark in the photo and it insisted that the photo was taken in the wrong neighborhood, going as far as giving the wrong address for one of the landmarks, imagining another front license plate on a car that didn't have one, and imagined a backstory for a supposedly well known stray cat in the photo.
When actually, modern ML can make really good guesses about ad relevancy using your location, data partners and recent searches from your home's IP address. When you explain this to people, they will still be convinced that the computer is listening to you and reasoning its way to deliver ads for you.
This is the "Woah, it must be listening to us" part. Because it is listening, not only just sound.
Tbf, it could do that before, and probably still better than the LLM: https://youtube.com/watch?v=ts5lPDV--cU But seeing it as what appears to be an emergent capability in such a general model is something else.
You can read the chat here: https://chatgpt.com/share/680a449f-d8dc-8001-88f4-60023323c7...
It took 4.5m to guess the location. The guess was accurate (checked using Google Street View).
What was amazing about it:
1. The photo did not have ANY text
2. It picked elements of the image and inferred based on those, like a fountain in a courtyard, or shape of the buildings.
4o can do it almost as well in a few seconds and probably 10-50x fewer tokens: https://chatgpt.com/share/680ceeff-011c-8002-ab31-d6b4cb622e...
o3 burns through what I assume is single-digit dollars just to do some performative tool use to justify and slightly narrow down its initial intuition from the base model.
I couldn't attach the chat directly since it's a temporary chat.
It'd be interesting to see the photo in the linked story at same resolution as provided to o3, since the licence plate in the photo in the story is at way lower resolution than the zoomed in version shown that o3 had access to. It's not a great piece of primary evidence to focus on though since a CA plate doesn't have to mean the car is in CA.
The clues that o3 doesn't seem to be paying attention to seems just as notable as the ones it does. Why is it not talking about car models, felt roof tiles, sash windows, mini blinds, fire pit (with warning on glass, in english), etc?
Being location-doxxed by a computer trained on a massive set of photos is unsurprising, but the example given doesn't seem a great example of why this could/will be a game changer in terms of privacy. There's not much detective work going on here - just narrowing the possibilities based on some of the available information, and happening to get it right in this case.
I don't consider it my job to impress or mind-blow people: I try to present as realistic as possible a representation of what this stuff can do.
That's why I picked an example where its first guess was 200 miles off!
A few years ago we were disocovering AI chat. AI could create sentences and have a basic conversation with us.
Today, it can identify a photo location with minimal "direct' information on it.
Where will it be in 3 years?? Crazy time to be alive.
It also curiously mentioned why this user is curious about the photo.
I relented after o3 gave up and let it know what building and streets it was. o3 then responded with an analysis of why it couldn't identify the location and asking for further photos to improve it's capabilities :-) !!!
It nailed it down to the exact park in Lisbon: https://maps.app.goo.gl/vpvqA14TCbeFb3Rd7
> The little clues that make me lean that way are the Portuguese-language exercise sign (“Abdominais”), the classic calçada portuguesa cobblestone path, the style of the outdoor-gym equipment the city installs, and—biggest tell—the long red-and-white neoclassical façade in the background, which looks just like the Nova Medical School building that borders that square.
I had forgotten the medical school!
It's not perfect, it thought this was in KL Malaysia: https://imgur.com/a/fbSDqyp (it's Bangkok Thailand)
> My hunch this time: the covered pedestrian sky-bridge that links Pavilion Bukit Bintang to the Kuala Lumpur Convention Centre / Suria KLCC in downtown Kuala Lumpur, Malaysia. The things that tipped me off are the distinctive white-panel ceiling and slim columns of that air-conditioned walkway, the glass balustrades, the vertical “PAVILION” banner glimpsed on the right, and the big LED façade of Pavilion Mall visible off to the left.
The sign on the right is part of the text Central World, which is a prominent mall in Bangkok.
I've had it fail similarly where it saw some text, messed up the OCR and was then lead to a wrong conclusion.
When I cleared personalization data and turned off extended memory it quit being nearly so accurate.
Here's an example [0] for "Riding e-scooters along the waterfront in Auckland". The iconic spire is correctly included, but so are many small details about the waterfront.
I've been meaning to harness this into a very-low-bandwidth image compression system. Where you take a photo and crunch it to an absurdly low resolution that includes EXIF data with GPS, date/time. You then reconstruct the fine details with AI.
Most photos are taken where lots of photos are taken, so the models have probably been appropriately trained.
[0] https://chatgpt.com/share/680d0008-54a0-8012-91b7-6b1794f485...
It basically iterates on coming up with some hypothesis and then does web searches to validate those.
I'm confused how so many people have such different outcomes. People seem to have fixated on the fact that the models use EXIF data if it's included, but it's trivially easy to run the test ensuring that isn't happening, and the results are still amazing.
I think some people really want to dismiss the capabilities of the models. I get that there's hype and it's annoying, but... look at what it's doing, right now, in front of you!
As an east european who grew up and lived in such a regime, I would like to respectfully remind all westerners their care-free and free lives is a privilege the majority of the world doesn't have.
Look at all the people in this thread talking about how people are fantastic at guessing locations from photos. This is not a new thing.
"If you want something to be secret don't post it online" is a principle that far predates LLMs. It's still true. It always was. The idea that authoritarian regimes had no way to place the location of photos before this is laughable.
I’m not a fan of this variation on “think of the children”. It has always been possible to deduce location from images. The fact that LLMs can also do it changes exactly nothing about the privacy considerations of sharing photos.
It’s fine to fear AI but this is a really weak angle to come at it from.
You are trying to manufacture outrage. Plain and simple.
And the governments are already doing this for decades at least, so ... I think the tech could be a net benefit, as with many other technologies that have matured.
If you want, I could sketch a socioeconomic archetype like "The Free Agent Technologist" that would match people like him really well. Would you like me to?"
Not dystopian: the crime solving potential, the research potential, the historical narrative reconstruction potential, etc.
It's a pattern I keep seeing over and over again. There seem to be a lot of values that we can obtain, individually or collectively, by bartering privacy in exchange for them.
If we had a sane world with sane, reliable, competent leadership, this would be less of a concern. But unfortunately we seem to have abdicated leadership globally to a political class that is increasingly incompetent and unhinged. My hypothesis on this is that sane, reasonable people are repelled from politics due to the emotional and social toxicity of that sector, leaving the sector to narcissists and delusional ideologues.
Unfortunately if we're going to abdicate our political sphere to narcissists and delusional ideologues, sacrificing privacy at the same time is a recipe for any number of really bad outcomes.
In Australia recently there was a terrible criminal case of massive child abuse.
They caught the guy because he was posting videos and one of them had a blanket which they somehow identified and traced to the child care Centre that he worked at.
It wasn’t done with AI but I can imagine photos and videos being fed into AI in such situations and asked to identify the location/people or other clues.
(It eventually "lost" the network connection)
I don't know how "iconic" that rocky outcrop in Madagascar is, to be honest. Google doesn't return much about it.
I couldn't attach the chat directly since it's a temporary chat.
I bet a lot of people (on HN at least) thought of "Does it use EXIF?" when they read the title alone, and got surprised that it was not the first thing you tested.
I’m sure there was an element of luck involved but it was still eery.
I have no memories stored, and in any case it shouldn’t know where I live exactly. The reasoning output didn’t suggest it was relying on any other chat history or information outside the image, but obviously you can’t fully trust it either.
https://www.bellingcat.com/news/2019/12/05/two-europol-stopc...
So its own code version of "where was this photo taken?"
Don't forget to activate reasoning.
My wife is a historian and just discovered the exact location of a travel photo of 1924
It failed to locate the image I provided. It got caught in a loop of cropping the image and presumably running some multi-approach to similarly search with images. If you use their image gen it is quite clear that they must've amassed a large image database at this point that they use for reference material pre-gen.
https://i.redd.it/ddscx4zibpme1.jpeg
I was intrigued because it looked much like Charleston SC. O3 didn't consider that and had to keep reminding itself that it couldn't possibly be the African flat bridge it kept determining it was (as the crop image analysis could recognize the suspension/cable structure).
https://arxiv.org/pdf/2404.10618
would be interesting to see how much better these reasoning models would be on the benchmark
That’s gonna be fun!
The tool is just intelligence. Intelligence itself is not dystopian or utopian. It's what you use it for that makes it so.
It is dystopian.
Some things are just tools that will be used for both good and bad.
If you don't want to post a photo, then don't post a photo.
Pick some random towns from a part of the UK: Horsham, Guildford, Worthing, Sheffield and see how it goes?
I would expect it to be able to guess on par with highly documented places.
Anybody else?
EDIT: My mistake, looks like those models are only available on the $20/month Plus plan or higher. I added a note about that to my post.
I've never heard of them abusing that phone number for anything other than that initial verification code.
My current intuition is that the US military / NSA etc have been just as suprised the explosion in capabilities of LLMs/transformers as everyone else.
(I'm using "intuition" here as a fancy word for "dumb-ass guess".)
I'd be interested to know if the NSA were running their own GPT-style models years before OpenAI started publishing their results.
im hunching, if you submit a photo of a clear sky, or a blue screen, it will choke
I am also wondering if we have any major breakthrough (comparatively speaking) coming out of LLM. Or non-LLM AI R&D.
I used a temporary chat, so no info about me is in the memory.
It guessed correctly down to the suburban town.
When asked to explain how it did it, it listed incredibly deductive reasoning.
Color me impressed.
Bro, it's not even been trained to do that specifically. Imagine what an image to location specific model could do.
Yes, I'm very very very scared. /s
A photo of people with cherry blossoms could be in many places, but if the majority of the people in the photo happen to be Japanese (and I'm curious how good LLMs are at determining the ethnicity of people now, and also curious if they would try to guess this if asked), it might guess Japan even if the cherry blossoms were in, say, Vancouver.
I looked at the image in the post before seeing the answer and would have guessed near San Francisco.
It seems impressive to someone if you haven't played Geoguessr a lot, but you'd be surprised at how much information there is about location from an image. The LLMs are just verbalizing what is happening in a few seconds in good player's mind.
I knew Terence Tao can solve Math Olympia questions and much much much more difficult questions. I was still very impressed by AlphaProof[0].
[0] https://deepmind.google/discover/blog/ai-solves-imo-problems...
If you want a bot that is extremely strong at geoguessr there is this: https://arxiv.org/abs/2307.05845
One forward pass is probably faster than 0.1 second. You can see its performance here: https://youtube.com/watch?v=ts5lPDV--cU (rainbolt is a really strong player)
But I wouldn’t be surprised if some form of cheating is happening.
Crazy that this is even allowed.
Who the hell needs to know the precise location of a picture, besides law enforcement? A rough location is most of the time sufficient. Like a region, a state, or a landscape (e.g., when you see the Bing background pictures, it's nice to see where they were taken).
This tool will give a boost to all those creeps out there that can have access to one or two pictures.
If you feed it a photo with a clear landmark it will get the location exactly right.
If you feed it a photo that's a close up of a brick wall it won't have a chance.
What's interesting is how well it can do on this range of tasks. If you don't think that's at least interesting I'm not sure what I can do for you.
Making a tool like this trained on existing map services, for example Google Street images, gives everyone, no matter who, the potential to find someone in no time.
These tools are growing like crazy, how long will it take before someone will "democratize" the "location services market"...