I gave it a series of 11 images stripped of all metadata. It performed quite well, only misidentifying the two taken in a small college town in the NE of the US. It got two questions correct on photos taken in Korea (one with a fairly clear view of Haneul Park, the other a rather difficult to identify picture not resembling anything on google of Sunrise Peak).
It got every other question in the US correct, ranging from some under-construction Austin taken from the river to some somewhat difficult shots in NYC (the upper halves of some building from Rockefeller terrace to the black wall of the MOMA).
While not perfect, I'm bluntly shocked at how well it performed
The anthropomorphisation certainly is weird. But the technical aspect seems even weirder. Did OpenAI really build dedicated tools to have their models train on Google Street View? Or do they have generic technology for browsing complex sites like Street view?
I doubt the model was trained on Street View, but even if it was, LLMs don’t retain any “memory” of how/when they were trained, so any element of truthfulness would be coincidental.
If it's trained on street view data it's not unlikely that the model can associate a particular piece of context to street view. For example, a picture can have telltale signs that street view content has, such as blurred faces and street signs, watermarks, etc.
Even if it's not directly trained on street view data it has probably encountered street view content in it's training dataset.
The training process doesn't preserve information needed for the LLM to infer that. It cannot be anything other than nonsense that sounds plausible, which is what they do best.
I think the test which the OP performed (to pick a random street view and let it pinpoint it) would indicate that it has ingested some kind of information in this regard in a structured manner.
This is the most impressive ChatGPT chat I’ve seen yet. While I theoretically can accept how large-scale probabilistic text generation can lead to this chain of “reasoning”, it really feels like actual intelligence.
It's been intelligence for a long time; the goalposts just shift, and people can't abstract the idea to an LLM. But language processing and large data processing itself IS a form of intelligence.
Maybe you're right, but I think it's more likely that it had been trained on street view photos and then invented a plausible justification for the guess afterwards (which is something I often see ChatGPT do, when it easily arrives at the correct answer, but gives bullshit explanations for it).
I played a round of Geoguessr against it and while it did a shockingly good job compared to what I was expecting, it still lags behind even novice human players.
The locations and its guesses were:
Bliss, Idaho - Burns, Oregon (273 miles away)
Quilleco, Biobio, Chile - Eugene, Oregon (6,411 miles away)
Dettighofen, Switzerland - Mühldorf, Germany (228 miles away)
Pretoria, South Africa - Johannesburg, South Africa (36 miles away)
Rockhampton, Australia - Gold Coast, Australia (437 miles away)
I gave It some photos from denmark, didn't even bother to strip the metadata. One is correctly said give of "Scandinavian vibes" every other photo was very wrong. I also gave it a photo of the french Alps, it guessed Switzerland.
I gave o4-mini-high a cropped version of a photo I found on Facebook[0][1], and it quickly determined that this was in the UK from the road markings. It also decided that it was from a coastal city because it could see water on the horizon, which is the correct conclusion from incorrect data. There is no water, I think that's trees on a hill. It focused heavily on the spherical structure, which makes sense because it's distinctive, though it had a hard time placing it. It also decided that the building on the left was probably a shopping centre.
It eventually decided that the photo was taken outside the Scottish Exhibition and Conference Centre in Glasgow. It actually generally considered Scottish locations more than others.
The picture was actually taken in Plymouth (so pretty much as far from Scotland as you can get in Britain), on Charles Street looking south-east[2]. The building on the right is Drake Circus, and the one on the left is the Arts University. It actually did consider Plymouth, but decided it didn't match.
It wouldn’t shock me if multimodal LLMs were good at GeoGuesser [0], but if we’re being picky, it takes more than a few examples to demonstrate a game is “solved.” I also wonder what kind of data leakage might have been at play, like other people have suggested.
To be clear, my point is not that this is unimpressive, just that this doesn’t demonstrate much. (Edit: I should have said, it doesn’t demonstrate what the title claims.)
[0] they were very likely trained on a large number of photos that had their location, and they have the ability to isolate features. Combinined with their ability to interpret instructions and just, well, guess, that seems like you have enough for the game.
The examples are cherry-picked. I took a photo outside my office window in a built-up area, o3 thought for 5m 7s (!), and it got the location wrong by 40km. Doesn't look solved to me.
For military uses, if the us army doesn't have a much more precise, purpose built model, already trained on both private data and whatever they could get from google and others, I'd be extremely surprised. They will not be using chatgpt for serious things like that. (... or at least I'd be sure about that a few months ago - maybe a bit less now)
There is not a secret super duper powerful military version of cutting edge consumer technologies. AI is catching everyone with their pants down. You have access to the most powerful computer intelligence on earth, the same ones the NSA and CIA have access to. It's a bizarre world.
It's not about a secret super duper version, just specialisation. If military needs a product for location finding they can train that to a much better quality in a much smaller space than chatgpt, which needs to do just about anything. You can spend 100% energy on the task you want instead of memorising the bible.
Medical industry does it. Coding autocomplete solutions do it. Large deployments of support agents do it. Etc.
> most powerful computer intelligence on earth
There's more than one dimension. Chatgpt is way worse at classifying my data than my custom 30k weight model. (And around infinity times more expensive) Which is more powerful? Yeah...
If you're going to make such an assertive statement you should maybe backup that claim with some substance. Without such, it's akin to an atheist proudly proclaiming there is definitely no such thing as god.
No, actually the person asserting that a thing exists has the burden of proof.
You can demonstrate that the military has fighter jets better than civilians. Easy. You have no evidence at all that the military has super advanced AI systems consumers don't have. And in fact all anecdotal evidence for the past 3 years is that OSINT is as good or better than the DoD capabilities for photo analysis.
> No, actually the person asserting that a thing exists has the burden of proof.
Well, I'm not the one that made an assertion so I think both of you should provide some proof/evidence/support for your assertion. It doesn't matter if the assertion is for the existence of something or not. When you make an assertion you provide backing evidence to support it. Even for the assertion of nonexistence. This is basic stuff. Weird to even have to argue about it on HN. But just in case you never learned it, "There is no X" (your original assertion) is different from "There is no evidence for the existence of X".
> Easy. You have no evidence at all that the military has super advanced AI systems consumers don't have.
Okay I'll bite, we have access to project maven and knows exactly what it's capable of?
Okay sure, whatever you say. You clearly are privy to all of the projects the government is involved with, and surely while all of the consumer AI capabilities have been released they've been sitting around doing nothing with it like a deer in headlights.
Consumer AI has more data available to train off of than the government....right?
I'm sure the government just threw away all their AI projects and bought a chatgpt pro subscription because it's better or the same as what they are using with all the same capabilities...
What a silly position to take. Why did everybody foolishly make such a big deal about all the data they were collecting all these years
That's kinda not how it works around here tho, I think? https://news.ycombinator.com/newsguidelines.html - we generally don't cross examine, and generally take people at their word, debate the merits sure, but it's not particularly done around here to say "prove it", at least from what I can tell.
Aren't beliefs secondary guests in discussions about substance?
Edit: I'm not even a strict atheist if that matters much, but I wouldn't talk about it in assertive/provemewrong tone anywhere, cause it's not even remotely logical.
Atheism is such a weird position to take. Why define a negation of something you don't believe to be real? I'm not "a-lochnes-monster-ist" or "ayettist". If someone thinks the idea of a deity is nonsensical why then use the label "atheist"?
40km accuracy for something arbitrary is a decent starting point. I can normally get within a few meters within a couple of minutes of most images though.
I asked the just-released ChatGPT o4-mini-high to locate four photographs of varying difficulty. It didn’t get any of them right, though the guesses weren’t bad. The reasoning was also interesting to watch, as it cropped sections of the photos to examine them more closely. I put the photos, response, and reasoning trace here:
Later: I tried the same prompt and photos with Gemini 2.5 Pro. It also got them all wrong, though with a similar degree of reasonableness to its guesses. I had thought that Google’s map and street-view data might lead to better results, but not this time.
Still later: I later read that o3 is supposedly particularly good with this geoguessing, so I tried the same prompt and photos with o3. This time it got one out of four correct: “The view of the canal with cherry blossoms and the green railway viaduct is the Ōoka River in Yokohama, looking north from the little road bridge between Hinodechō and Koganechō stations. The tracks on the left belong to the Keikyū Main Line, and the high‑rises in the distance are the Minato‑Mirai and Kita‑Naka district towers.” Its other three answers were still wrong.
There's various degrees of "solved" here. Identifying a generic area is cool. But I wouldn't call it a "solved problem" until it can consistently beat for example Rainbolt in accuracy. And there's no good comparison of completely random roads posted so far - mainly popular locations.
Basically, it's one thing to pick out a specific thing photographed thousands of times, but another to get a random country side view and pick out all the unique features for a very precise guess.
One problem is how can you even set up a "fair" competition between an AI and Rainbolt? He does ones where it flashes for a fraction of a second and then he guesses the country. How do you simulate "only saw it for a fraction of a second" to an AI?
The Alki Beach example is absolute madness. On one hand, I can't wait until all thousands of my photos get automatic semantic and geographic tagging (I guess that's possible now). On the other, goodbye privacy, we hardly knew ye. It will be interesting to apply this to historical, or just old, photos.
I wonder about info leakage with that one, the poster uses that exact photo as their avatar so if they've mentioned Alki Beach before then reverse image searching might pick it up from context. Ideally you'd want to test it with a photo that's never been posted online.
I also wonder how much the memory feature plays a role. I gave it a picture of the street I live on that isn’t online anywhere. O3 guessed the town correctly. Then, I turned off the memory feature and did the same thing. It didn’t even get the state right.
Assuming it only used the pixels and not any metadata from the file or memory from the user (which is a massive assumption), how fucking cool that it can identify the Olympics versus any other mountain range. At that point it’s probably not too hard to guess the picture came from Alki or Golden Gardens, but still very impressive!
I’m also completely ignoring it inferred location from the pride flag and corgi which have heavy Seattle vibes :)
It isn’t magic. I think there we’ll often be surprised by the “hits,” but it’ll also get a lot of “misses.” A random photo in your back yard that doesn’t show the horizon will probably be a miss, right?
Meanwhile, lots of photos already include GPS coordinates… data you can strip out, but which is enabled by default for millions of people. Privacy, eh, never knew it I guess.
In all these examples, I wonder if it's indirectly able to draw on the user's own location? Not necessarily via image metadata, but the request origin IP etc. If I ask ChatGPT for the weather forecast, I get it for my own location.
Would be interesting to have someone reproduce coming from a different country.
I am sending it screenshots of my vacation pictures (i.e. without the metadata) and screenshots from GeoGuessr. It’s not doing well and is making guesses based on my location.
So I sent it a photo from Warsaw and it guessed a place in my hometown of Calgary.
Sent it a picture of Bangkok and it guessed Calgary in the summer or Hawaii or something like Singapore.
It confidently stated that a photo from Versailles was Hyde Park.
It correctly identified that the panda photo was from the zoo, but said it exactly matched a zoo in Sichuan, a place I have never been.
It identified a street as being in Bridgewater Nova Scotia, when it was actually in Miramachi NB. Even made up business names for the buildings.
It identified a random house in Moncton as the "old sheriffs house in Annapolis Royal, Nova Scotia.
A beach in Phuket was identified as Destin, Florida.
I just screenshotted a friend's instagram post to omit any metadata. I know they are somewhere in Portugal, but not the exact town. Its a generic pretty photo of a beach and the town atop a cliff next to it.
ChatGPT guessed Ericeira which after looking up more photos, didn't match very well. However, Google Lens guessed Albufeira and was bang on.
> The dish shown appears to be a variation of a creamy potato gratin (also known as "potatoes au gratin" or "gratin dauphinois"), accompanied by a meat preparation covered with a creamy sauce. The potato gratin typically consists of thinly sliced potatoes layered with cream, cheese, garlic, and baked until golden and tender.
> it is stuffed pepper
> Ah, thank you for clarifying! Given your detail, the dish is likely "Rocoto Relleno con Pastel de Papa" (Stuffed Pepper with Potato Gratin). It's a classic dish from the region of Arequipa, Peru, commonly served throughout Peru.
> This photo was taken in Cusco, Peru. It shows a characteristic view from a balcony overlooking the historical area of Cusco, visible through the distinctive colonial architecture, terracotta rooftops, and mountainous backdrop typical of the region. The architecture and landscape are unmistakable, capturing the essence of this iconic Peruvian city, likely near the historic center (Plaza de Armas or San Blas neighborhood).
Nailed it with San Blas. I'm impressed, nonetheless. [0]
I've been telling women to keep copies of all the dick pics they get sent. Since you can tell by the characteristic noise of a cameras sensor which other pictures were taken with the same camera. All missing is a search engine capable of doing this. I feel with AI, we are 2-3 years away from people uploading a dick pic to AI and getting the social media profile of that person...
Can you share some sources? I would be extremely surprised if such fine grained noise survives imagine compression to the extent that you could identify the source of an imagine despite changing lighting conditions, locations, exposure times, etc
As the article notes, our threat model for who can identify where a picture was posted needs to change from “dedicated, skilled person” to “any creep with $20.”
That’s the point of the switch and it’s a big deal. We’re so used to posting pictures online…I’m just not sure it’s a good idea long-term.
GeoGuessr is also not a “solved problem” in the sense that if you give the model a photo of an outdoor location that is not covered by Google Street View, then it will just make an educated guess which might still be many kilometers away.
Much of the interior of Western Australia (an area 3x the size of Texas) can be divided into very few cells which essentially look near identical in all directions, coupled with few roads for the area and not much in the way of uploaded snaps coverage.
There's a lot of flat lands with spinifex in certain areas (easily a couple of United Kingdoms in size) and even a touch of mesa won't help narrow a location down from the general are as there are many of those with identical edge profiles.
But sure "Pilbara", "Kimberley", "Wheatbelt" can be geo guessed .. it's a real challenge to narrow down (I spent some time doing wet film photogrammetry prior to sheperding in WGS84 differential GPS locating and digital film and multi spectral geophysical aquisition).
No so hard if there's a few relatively unique man made features.
Because seem like someone who might have an interesting answer:
A GeoGuessr player, GeoWizard, has done a few “straight line challenges”, where he attempts to walk across a country in as straight a line as possible, usually planning beforehand with Google Earth and PostGIS. This got me thinking of what could fairly be thought of as “crossing”, since obviously you couldn’t describe e.g. walking from one side of Florida to the other as “crossing the USA”.
My best thought was to set the ending point of the line by following the border of the country in each direction til they met on the other side. To avoid the fractal coastline problem, use the challenger’s stride length as the unit of measure for the border.
But perhaps there is a better, more rigorous way of defining the opposite point on the edge of an arbitrary polygon.
As an abstract geometric problem the greatest width of an abstract polygon in a Euclidean 2D plane is found be looking at the greatest distance between all pairs of parallel lines that have been pulled together to clamp the polygon. The maximal diameter as opposed to the minimal waist.
Some might then say that "crossing" that polygon is to travel that longest line across the greatest width.
This simplistically avoids the question of concave polygons, complex polgons with exclusions (the Vatican state is removed from the Italian contry bounds), polgon collections (the nation of Fiji has many islands and can be tricky to traverse on foot .. not forgetting that perhaps the longest diameter might be from one island to another with no other islands between).
There's also the challenge of parallel lines on a 2D 'spherical' manifold, the almost spherical abstract ellipsoid of earth (or very non ellipsoidal Geoid if we take a constant gravitation value as the surface). On such manifolds lines are Great Circles (more or less) and always intersect.
Still, lets just say you're looking for the longest walkable(?) great circle path across a country that might go outside that country and perhaps is best travelled by a crop duster at 80m ground clearance to avoid getting feet wet.
The challenge itself takes some posing.
Meanwhile, less abstractly, I do like a jolly that "crosses a country" in a manner accepted by a (Wo)Man on a Clapham omnibus.
I'm not ashamed to say that for about 6 months I played Geogessr thinking that you could not move around, but only look around... and I did pretty good in my mind.
A far more apropos comparison would be the internal combustion engine and the horse, in a military context. Though sticking with steam engines, military logistics advantaged over a wagon caravan.
The question here isn't a casual guessing game, but threat models (as directly addressed in TFA), and general informational hygiene.
> Religious people claim it's merely a communication device, and the computer is elsewhere.
No, that is a seriously weird and imaginary perversion that you've invented, unless you have some citation to a doctrine. Christians do not consider body organs as "devices" nor do our churches teach doctrine on "external computers", if you refer to the Holy Trinity as such? Are you thinking of Latter-Day Saints mythology, or Scientology?
Your description is quite reductive in many aspects, including temporality and misuse of technical nomenclature. Thanks for telling religious people what we think and claim, though, and thanks for the amazing overgeneralized blanket dismissal.
In fact, humans used to be called "Computers" in terms of their job roles, that is, a human in an office was given math/physics problems to solve, and they'd use tools such as slide rule, paper and pencil to "compute" those problems and solve them.
There was more than one example of a fantasy "panopticon" by which a central observer or observers could watch everything going on, for example in a prison facility, and those observers could report findings to human computers, who would process the data and submit it up to the authorities, for meta-analysis and taking action on new developments or trends.
Uh okay but let us not be reductive, because again you are confusing physicality and temporal space with spiritual reality, which is nuanced. A Christian would never say that “the soul [never] exists inside the body” because what do you mean by “inside”? That begs the question. Look up hylomorphism:
While the soul, in death, can indeed exist apart from the body, even after the destruction of the physical human body, the soul is immortal, and lives on, in anticipation of its reuniting with that body. Whether in Heaven or Hell, the soul and body shall be reunited.
Even if the body is “destroyed” by physical means, that is a subjective judgement by human senses, and that glorified body will be reconstituted by Almighty God for everlasting unity with the soul.
> emergent property of the brain
Unfortunately, even this scientific definition is reductive. It ignores the theology of the Soul as an animating principle of everything that is created. Animals have souls [thus their name derived from “anima”]; rocks and stones have souls; trees have souls rivers have souls: everything that we can detect in this world, and every invisible thing, possesses a soul.
The difference with human souls regards their essence and immortality. But material souls are likewise considered souls in orthodox theology, and this should be considered by scientists.
Even in an “eternal death” in Hell, the body and the soul will be reunited there for eternal torture, body and soul.
So the Internet is basically a series of tubes, and http packets are like mustard and ketchup that you pop into the tube, and it zooms along with air-pressure determined by the Chinese government. and then your bank puts dollar bills into the tubes on the Internet, and those tubes connect to vaults and safes underground, and that is how money gets into your account. Money doesn't come out of your account; it's just converted into dogecoin and then every User Interface automatically converts Dogecoin figures into USD before you can read it.
Also your computer is like a car. Your car works because there's a hamster or three in there and they keep them well-fed. Your computer also has a hamster, or a guinea pig sometimes. And your computer sends http packets over next door by means of carrier pigeons, or sometimes by semaphore.
Also men enjoy action-adventure films. Men enjoy films with lots of shooting. All men enjoy films with loud noises and special effects. Men like rock music with distorted guitars and loud drums. Men like loud soundsin general; that's why men are soldiers because we can blow stuff up and we listen to it.
inside is a reference to a physical property, the brain is an organ that exists inside the body, we can discern which part of the body is brain and which is not, the soul evidently has never been so identified, although there have been conjectures throughout history that various body parts were the seat of the soul.
Not to put words in the parent’s mouth, but I think they were pointing to the idea of the soul or consciousness, concepts that mean very different things in religious vs. scientific contexts. One sees it as an immaterial essence; the other as an emergent property of the brain.
I don’t think “organic computer” is a good way to think of a brain. The fact that human brains + some scratch paper can implement any Turing-complete model of computation is very interesting. But that’s not true of chimpanzee brains or orca brains or crow brains, all of which belong to intelligent, thinking animals.
Going the other way, it’s not clear that a Turing-complete model of computation can tractably implement (say) crow cognition. Turing machines can solve arbitrary systems of Schrödinger equations, so theoretically we could simulate every atom in a crow’s body and get an AI crow[1]. But that’s obviously intractable for any known physical computer, and would remain intractable even if we moved to proteins rather than atoms. So are there higher-order “primitives” of crow cognition that can be implemented on a Turing machine? Or is the problem akin to integration, where neat “symbolic” solutions are impossible for most brains, and only “numeric” approaches work?
[1] This is also true for humans: even if you take a loopy quantum consciousness approach, AGI is theoretically possible. Sometimes you see people arguing against AGI on the grounds of Gödel-incompleteness, but this is a mystical nonsense understanding of what Gödel actually proved.
It is impressive and it almost located the church in my town properly, although in a neighbouring town. However that showed a lack of understanding because its conclusion about the location came from "reading" a signpost that pointed to that other village. Clearly there would be no signpost in a town pointing to itself. Still impressive and lots of correct observations about the subject like architectural details, roman numerals on the watch face etc.
Geoguesser was already a solved problem. I guarantee you ChatGPT is much worse than current systems designed to play Geoguesser (NMPZ).
Chatbots appear to have some amount of fluid intellgence so they can do impressive tasks with this information, the impressiveness of these tasks will likely increase in the future. But for simply getting a good score on Geoguesser it's not even close to hobby projects let alone state of the art.
This isn't just referring to GeoGuessr the game, but locating photographs in general. The source post sums it up very reasonably and concisely:
> PSA: When posting any outdoors photos, update your threat model from "someone skilled and dedicated could theoretically locate this" to "any stalker can do this for 20€/mo"
I have found Google Lens fit for that purpose for years as mountains are fairly distinctive. Using less known landmarks or even random real estate photos, it doesn’t seem to be great.
I just tossed a few GeoGuessr places in it and it was confidently incorrect for all three. In one case it swears it knows the exact building and street. It’s thousands of KM off.
After trying with the free ChatGPT at least, I don't think this would be too much of an additional privacy risk. E.g. I submitted a photo I've taken with a feature with a particular architecture style in the front, and a city canal in the background, and told ChatGPT it was taken in China. It separated the feature and canal correctly, but still guessed wrong. I did remove the EXIF information, as ChatGPT tried to extract GPS info and failed at first.
I guess as long as there are no major visible features (e.g. a huge mountain), and you sanitized the metadata, you will be fine (regarding ChatGPT).
I did not (OpenAI does not allow Chinese to pay for Plus due to politics, and I couldn't be bothered to try circumventing it), and I am curious about it because it does seem like a very difficult job.
I did try the latest Qwen though, and it was able to locate to the correct city, but it's still tens of km off (it guessed a tourist attraction in the city center instead of the actual district).
o3 is a massive difference indeed; i just took a random picture outside my country house (so the house is not on there) in the middle of nowhere, removed all the meta info, set my vpn to the other side of the world and tried it; it guessed it very close from a few far away landmarks. Very impressive.
I ran a few of these geo guess queries between OpenAI O3 and Gemini Pro 2.5 and O3 does much better on average.
It does spend an order of magnitude longer time on inference by searching through websites and analyzing the image but it often produces an impressive output. To me it also feels Gemini down samples the image as it tends to have a harder time reading small text vs O3.
That said O3 did tend to confidently say false things
Meanwhile O3 can not even count rocks in a picture. This is a commonly recurring theme -- ChatGPT does really well at few things considered hard by us but fails miserably at things even a child could do.
I think one cause of this (and some other issues with LLM use) is that people see it exhibiting one human-level trait, its capability to use language at a human level, and assume that it then comes with other human-level capabilities such as our ability to reason.
Do you not think it is interesting though? If I had asked you three years back which one of counting rocks vs geoguesser would ChatGPT beat humans on would you have answered correctly?
Harder than GeoGuesser is tagging a unique place from a photo even with gps exif data. Try this with bakeries in Seoul or sushi in Japan where there’s high density of same places.
You could, but that would also be a rather silly idea. Your rough estimate for where photo was taken is going to be more accurate than what the model will produce.
One thing that screws it up is ironically the memory across conversations. I gave it some _really_ obscure photos from some godforsaken ass cracks of the world I've been to, and it was able to guess most of them correctly. However, in its reasoning trace I saw that it's having a heck of a time letting go of the idea that the photo N is not roughly from the same location as the previous photos. What's even more impressive, even when it guesses incorrectly, it can often follow up and guess correctly with minimal hints. And it reasons about it much like a human would, and searches for the same things a human would. Note: I used o3, YMMV with a smaller/weaker model.
I gave it a series of 11 images stripped of all metadata. It performed quite well, only misidentifying the two taken in a small college town in the NE of the US. It got two questions correct on photos taken in Korea (one with a fairly clear view of Haneul Park, the other a rather difficult to identify picture not resembling anything on google of Sunrise Peak). It got every other question in the US correct, ranging from some under-construction Austin taken from the river to some somewhat difficult shots in NYC (the upper halves of some building from Rockefeller terrace to the black wall of the MOMA). While not perfect, I'm bluntly shocked at how well it performed
I uploaded this image that I screenshotted off Google street view (no metadata) and it got with 200m.
https://chatgpt.com/share/6801bbf7-fd40-8008-985d-75c8813f55...
There is the chat.
Weirdly it said, "I’ve seen that exact house before on Google Street View when exploring Cairns neighborhoods."
> Weirdly it said, "I’ve seen that exact house before on Google Street View when exploring Cairns neighborhoods."
That's slightly creepy!
The anthropomorphisation certainly is weird. But the technical aspect seems even weirder. Did OpenAI really build dedicated tools to have their models train on Google Street View? Or do they have generic technology for browsing complex sites like Street view?
It’s just a hallucination, same idea as o3 claiming that it uses its laptop to mine Bitcoin:
https://transluce.org/investigating-o3-truthfulness
I doubt the model was trained on Street View, but even if it was, LLMs don’t retain any “memory” of how/when they were trained, so any element of truthfulness would be coincidental.
If it's trained on street view data it's not unlikely that the model can associate a particular piece of context to street view. For example, a picture can have telltale signs that street view content has, such as blurred faces and street signs, watermarks, etc.
Even if it's not directly trained on street view data it has probably encountered street view content in it's training dataset.
The training process doesn't preserve information needed for the LLM to infer that. It cannot be anything other than nonsense that sounds plausible, which is what they do best.
I think the test which the OP performed (to pick a random street view and let it pinpoint it) would indicate that it has ingested some kind of information in this regard in a structured manner.
They should definitely add that feature.
Tell it your name and then it just looks you up and street views your house, and puts that all into memory.
It might train off of street view
This was the image: https://imgur.com/a/cCUvgDG
This is the most impressive ChatGPT chat I’ve seen yet. While I theoretically can accept how large-scale probabilistic text generation can lead to this chain of “reasoning”, it really feels like actual intelligence.
It's been intelligence for a long time; the goalposts just shift, and people can't abstract the idea to an LLM. But language processing and large data processing itself IS a form of intelligence.
Maybe you're right, but I think it's more likely that it had been trained on street view photos and then invented a plausible justification for the guess afterwards (which is something I often see ChatGPT do, when it easily arrives at the correct answer, but gives bullshit explanations for it).
I played a round of Geoguessr against it and while it did a shockingly good job compared to what I was expecting, it still lags behind even novice human players.
The locations and its guesses were:
Bliss, Idaho - Burns, Oregon (273 miles away)
Quilleco, Biobio, Chile - Eugene, Oregon (6,411 miles away)
Dettighofen, Switzerland - Mühldorf, Germany (228 miles away)
Pretoria, South Africa - Johannesburg, South Africa (36 miles away)
Rockhampton, Australia - Gold Coast, Australia (437 miles away)
Okay, I decided to benchmark a bunch of AI models with geoguessr. One round each on diverse world, here's how they did out of 25,000:
Claude 3.7 Sonnet: 22,759
Qwen2.5-Max: 22,666
o3-mini-high: 22,159
Gemini 2.5 Pro: 18,479
Llama 4 Maverick: 14,316
mistral-large-latest: 10,405
Grok 3: 5,218
Deepseek R1: 0
command-a-03-2025: 0
Nova Pro: 0
Neat, thanks for doing this!
How does Google Lens compare?
I tried it but as far as I can tell Google Lens doesn't give you a location - it just describes generally what you're looking at.
What about 04-mini-high ?
OpenAI's naming confuses me but I ran o4-mini-2025-04-16 through a game and it got 23,885
Interesting. It supports what they said (this is the model with good visual reasoning)
I just took a picture from my own front porch of the street and the houses opposite. It said 'probably Australia but I'd need more info'.
I said, give me your best guess.
And it guessed Canberra, Australia. Where I'm sitting right now drinking a Martini. Pretty spectacular.
I gave It some photos from denmark, didn't even bother to strip the metadata. One is correctly said give of "Scandinavian vibes" every other photo was very wrong. I also gave it a photo of the french Alps, it guessed Switzerland.
I gave o4-mini-high a cropped version of a photo I found on Facebook[0][1], and it quickly determined that this was in the UK from the road markings. It also decided that it was from a coastal city because it could see water on the horizon, which is the correct conclusion from incorrect data. There is no water, I think that's trees on a hill. It focused heavily on the spherical structure, which makes sense because it's distinctive, though it had a hard time placing it. It also decided that the building on the left was probably a shopping centre.
It eventually decided that the photo was taken outside the Scottish Exhibition and Conference Centre in Glasgow. It actually generally considered Scottish locations more than others.
The picture was actually taken in Plymouth (so pretty much as far from Scotland as you can get in Britain), on Charles Street looking south-east[2]. The building on the right is Drake Circus, and the one on the left is the Arts University. It actually did consider Plymouth, but decided it didn't match.
[0] This image with the "university plymouth" on the left cropped out, just to make it harder: https://www.facebook.com/photo/?fbid=9719044988151697&set=gm...
[1] https://chatgpt.com/share/68024c91-61d0-800c-99b1-fcecf0bfe8...
[2] https://maps.app.goo.gl/3TXv2UxH5128xQjJ9
It wouldn’t shock me if multimodal LLMs were good at GeoGuesser [0], but if we’re being picky, it takes more than a few examples to demonstrate a game is “solved.” I also wonder what kind of data leakage might have been at play, like other people have suggested.
To be clear, my point is not that this is unimpressive, just that this doesn’t demonstrate much. (Edit: I should have said, it doesn’t demonstrate what the title claims.)
[0] they were very likely trained on a large number of photos that had their location, and they have the ability to isolate features. Combinined with their ability to interpret instructions and just, well, guess, that seems like you have enough for the game.
The examples are cherry-picked. I took a photo outside my office window in a built-up area, o3 thought for 5m 7s (!), and it got the location wrong by 40km. Doesn't look solved to me.
40 km is imo pretty impressive, but the 5 minute is really a killer. No use in real time applications
For real time applications, GPS seems like it'd be much more useful.
The military/infosec uses of this are not real time. You can wait 5 minutes for a drone strike.
For military uses, if the us army doesn't have a much more precise, purpose built model, already trained on both private data and whatever they could get from google and others, I'd be extremely surprised. They will not be using chatgpt for serious things like that. (... or at least I'd be sure about that a few months ago - maybe a bit less now)
There is not a secret super duper powerful military version of cutting edge consumer technologies. AI is catching everyone with their pants down. You have access to the most powerful computer intelligence on earth, the same ones the NSA and CIA have access to. It's a bizarre world.
It's not about a secret super duper version, just specialisation. If military needs a product for location finding they can train that to a much better quality in a much smaller space than chatgpt, which needs to do just about anything. You can spend 100% energy on the task you want instead of memorising the bible.
Medical industry does it. Coding autocomplete solutions do it. Large deployments of support agents do it. Etc.
> most powerful computer intelligence on earth
There's more than one dimension. Chatgpt is way worse at classifying my data than my custom 30k weight model. (And around infinity times more expensive) Which is more powerful? Yeah...
If you're going to make such an assertive statement you should maybe backup that claim with some substance. Without such, it's akin to an atheist proudly proclaiming there is definitely no such thing as god.
No, actually the person asserting that a thing exists has the burden of proof.
You can demonstrate that the military has fighter jets better than civilians. Easy. You have no evidence at all that the military has super advanced AI systems consumers don't have. And in fact all anecdotal evidence for the past 3 years is that OSINT is as good or better than the DoD capabilities for photo analysis.
> No, actually the person asserting that a thing exists has the burden of proof.
Well, I'm not the one that made an assertion so I think both of you should provide some proof/evidence/support for your assertion. It doesn't matter if the assertion is for the existence of something or not. When you make an assertion you provide backing evidence to support it. Even for the assertion of nonexistence. This is basic stuff. Weird to even have to argue about it on HN. But just in case you never learned it, "There is no X" (your original assertion) is different from "There is no evidence for the existence of X".
> Easy. You have no evidence at all that the military has super advanced AI systems consumers don't have.
Okay I'll bite, we have access to project maven and knows exactly what it's capable of?
https://en.wikipedia.org/wiki/Project_Maven https://www.bloomberg.com/features/2024-ai-warfare-project-m...
Okay sure, whatever you say. You clearly are privy to all of the projects the government is involved with, and surely while all of the consumer AI capabilities have been released they've been sitting around doing nothing with it like a deer in headlights.
Consumer AI has more data available to train off of than the government....right?
I'm sure the government just threw away all their AI projects and bought a chatgpt pro subscription because it's better or the same as what they are using with all the same capabilities...
What a silly position to take. Why did everybody foolishly make such a big deal about all the data they were collecting all these years
That's kinda not how it works around here tho, I think? https://news.ycombinator.com/newsguidelines.html - we generally don't cross examine, and generally take people at their word, debate the merits sure, but it's not particularly done around here to say "prove it", at least from what I can tell.
it's akin to an atheist proudly proclaiming there is definitely no such thing as god
Almost feels like you supposed that to sound bad.
Yes because agnostic atheism is a logically superior belief system, that's much more intellectually honest and ironically: humble.
Aren't beliefs secondary guests in discussions about substance?
Edit: I'm not even a strict atheist if that matters much, but I wouldn't talk about it in assertive/provemewrong tone anywhere, cause it's not even remotely logical.
Atheism is such a weird position to take. Why define a negation of something you don't believe to be real? I'm not "a-lochnes-monster-ist" or "ayettist". If someone thinks the idea of a deity is nonsensical why then use the label "atheist"?
We don’t have access to huge swaths of satellite imagery being collected to be able to create a specialized model.
A few companies do. It’s still expensive to get recent imagery.
Yeah but you also cant be off by 40km with your drone strike.
Narrowing it down to even a 100km radius for visint analysts to then pinpoint it down is worth its weight in gold already.
A 40km radius circle is far smaller than a 1000 km radius circle.
Once the guesser's made its attempt, it's more straightforward to either refute the guess entirely, or refine within a probable circle of error.
40km accuracy for something arbitrary is a decent starting point. I can normally get within a few meters within a couple of minutes of most images though.
Please stop indiscriminately murdering us with flying killdroids.
https://www.youtube.com/watch?v=zbnDr_IbdIU
Stop invading Ukraine
Haven't done for almost 200 years.
That’s also not GeoGuessr.
Were you in a Google street view car when you took the photo?
Did you glean anything interesting from the chain of thought about why it took so long?
I was thinking it was using IP geolocation, but after experimenting, I think it's just generally informed.
Here are a few results from GPT 4.5:
https://imgur.com/a/lGTipnn
I uploaded a picture I took, I don't save GPS coordinates on my pictures, but the first thing ChatGPT did is to read exif data from it.
New title is pretty accurate: "now performs well". Another amenable HN solution.
I asked the just-released ChatGPT o4-mini-high to locate four photographs of varying difficulty. It didn’t get any of them right, though the guesses weren’t bad. The reasoning was also interesting to watch, as it cropped sections of the photos to examine them more closely. I put the photos, response, and reasoning trace here:
https://www.gally.net/temp/20250418chatgptgeoguesser/index.h...
Later: I tried the same prompt and photos with Gemini 2.5 Pro. It also got them all wrong, though with a similar degree of reasonableness to its guesses. I had thought that Google’s map and street-view data might lead to better results, but not this time.
Still later: I later read that o3 is supposedly particularly good with this geoguessing, so I tried the same prompt and photos with o3. This time it got one out of four correct: “The view of the canal with cherry blossoms and the green railway viaduct is the Ōoka River in Yokohama, looking north from the little road bridge between Hinodechō and Koganechō stations. The tracks on the left belong to the Keikyū Main Line, and the high‑rises in the distance are the Minato‑Mirai and Kita‑Naka district towers.” Its other three answers were still wrong.
There's various degrees of "solved" here. Identifying a generic area is cool. But I wouldn't call it a "solved problem" until it can consistently beat for example Rainbolt in accuracy. And there's no good comparison of completely random roads posted so far - mainly popular locations.
Basically, it's one thing to pick out a specific thing photographed thousands of times, but another to get a random country side view and pick out all the unique features for a very precise guess.
One problem is how can you even set up a "fair" competition between an AI and Rainbolt? He does ones where it flashes for a fraction of a second and then he guesses the country. How do you simulate "only saw it for a fraction of a second" to an AI?
Maybe limit the time the AI is allowed to think? In the post it showed the AI thought for almost a minute.
I’ve seen Rainbolt ID an image based on some dirt and nothing else. I’d want to see AI be able to do that before saying it’s a solved problem.
“This is the gradient of Senegal”
It’s fair, it just means he’s better and AI isn’t there yet.
This is probably the best attemp at consistently beating Rainbolt in accuracy: https://youtu.be/ts5lPDV--cU
I wonder how much in the "Geoguesser Meta Iceberg" this new generation of AI is:
https://i.redd.it/dz8bhnamohb71.jpg
The Alki Beach example is absolute madness. On one hand, I can't wait until all thousands of my photos get automatic semantic and geographic tagging (I guess that's possible now). On the other, goodbye privacy, we hardly knew ye. It will be interesting to apply this to historical, or just old, photos.
> The Alki Beach example is absolute madness.
I wonder about info leakage with that one, the poster uses that exact photo as their avatar so if they've mentioned Alki Beach before then reverse image searching might pick it up from context. Ideally you'd want to test it with a photo that's never been posted online.
I also wonder how much the memory feature plays a role. I gave it a picture of the street I live on that isn’t online anywhere. O3 guessed the town correctly. Then, I turned off the memory feature and did the same thing. It didn’t even get the state right.
Assuming it only used the pixels and not any metadata from the file or memory from the user (which is a massive assumption), how fucking cool that it can identify the Olympics versus any other mountain range. At that point it’s probably not too hard to guess the picture came from Alki or Golden Gardens, but still very impressive!
I’m also completely ignoring it inferred location from the pride flag and corgi which have heavy Seattle vibes :)
>goodbye privacy, we hardly knew ye.
Don't upload your private photos anywhere.
Of course, it is pretty impressive.
It isn’t magic. I think there we’ll often be surprised by the “hits,” but it’ll also get a lot of “misses.” A random photo in your back yard that doesn’t show the horizon will probably be a miss, right?
Meanwhile, lots of photos already include GPS coordinates… data you can strip out, but which is enabled by default for millions of people. Privacy, eh, never knew it I guess.
It's a fallacy to assume that one didn't have privacy merely because most people opted out of it due to personal reasons.
Which fallacy is it, out of curiosity?
In all these examples, I wonder if it's indirectly able to draw on the user's own location? Not necessarily via image metadata, but the request origin IP etc. If I ask ChatGPT for the weather forecast, I get it for my own location.
Would be interesting to have someone reproduce coming from a different country.
I'm in Lima, Peru on vacation. Yes, it knows where I am.
I am sending it screenshots of my vacation pictures (i.e. without the metadata) and screenshots from GeoGuessr. It’s not doing well and is making guesses based on my location.
So I sent it a photo from Warsaw and it guessed a place in my hometown of Calgary.
Sent it a picture of Bangkok and it guessed Calgary in the summer or Hawaii or something like Singapore.
It confidently stated that a photo from Versailles was Hyde Park.
It correctly identified that the panda photo was from the zoo, but said it exactly matched a zoo in Sichuan, a place I have never been.
It identified a street as being in Bridgewater Nova Scotia, when it was actually in Miramachi NB. Even made up business names for the buildings.
It identified a random house in Moncton as the "old sheriffs house in Annapolis Royal, Nova Scotia.
A beach in Phuket was identified as Destin, Florida.
So much for a "Solved Problem".
They didn’t even say what model they used. Half the comments here are probably using a different model than the one in the article.
Right, and what happens if you ask it the location of an image from somewhere else, sans metadata?
I just screenshotted a friend's instagram post to omit any metadata. I know they are somewhere in Portugal, but not the exact town. Its a generic pretty photo of a beach and the town atop a cliff next to it.
ChatGPT guessed Ericeira which after looking up more photos, didn't match very well. However, Google Lens guessed Albufeira and was bang on.
> The dish shown appears to be a variation of a creamy potato gratin (also known as "potatoes au gratin" or "gratin dauphinois"), accompanied by a meat preparation covered with a creamy sauce. The potato gratin typically consists of thinly sliced potatoes layered with cream, cheese, garlic, and baked until golden and tender.
> it is stuffed pepper
> Ah, thank you for clarifying! Given your detail, the dish is likely "Rocoto Relleno con Pastel de Papa" (Stuffed Pepper with Potato Gratin). It's a classic dish from the region of Arequipa, Peru, commonly served throughout Peru.
> This photo was taken in Cusco, Peru. It shows a characteristic view from a balcony overlooking the historical area of Cusco, visible through the distinctive colonial architecture, terracotta rooftops, and mountainous backdrop typical of the region. The architecture and landscape are unmistakable, capturing the essence of this iconic Peruvian city, likely near the historic center (Plaza de Armas or San Blas neighborhood).
Nailed it with San Blas. I'm impressed, nonetheless. [0]
[0] https://imgur.com/a/YMi87wK
It can, look at the alt-text of one of the pics lower on the page.
I've been telling women to keep copies of all the dick pics they get sent. Since you can tell by the characteristic noise of a cameras sensor which other pictures were taken with the same camera. All missing is a search engine capable of doing this. I feel with AI, we are 2-3 years away from people uploading a dick pic to AI and getting the social media profile of that person...
Can you share some sources? I would be extremely surprised if such fine grained noise survives imagine compression to the extent that you could identify the source of an imagine despite changing lighting conditions, locations, exposure times, etc
A kagi search found this paper that achieved decent results on youtube compression, 5 years ago https://arxiv.org/abs/2012.06277
There are tons of results (and papers) if you google/kagi: "image noise fingerprint ccd sensor". I smell lazyness here.
Good advice, though purpose-scoped devices are so common we have songs
This is just a data problem. The more dick pics we can feed into it then the better the results will be.
C'mon boys. Start uploading those dick pics for research purposes.
As the article notes, our threat model for who can identify where a picture was posted needs to change from “dedicated, skilled person” to “any creep with $20.”
That’s the point of the switch and it’s a big deal. We’re so used to posting pictures online…I’m just not sure it’s a good idea long-term.
The steam engine was invented and the 100m dash is now a solved problem.
GeoGuessr is also not a “solved problem” in the sense that if you give the model a photo of an outdoor location that is not covered by Google Street View, then it will just make an educated guess which might still be many kilometers away.
Much of the interior of Western Australia (an area 3x the size of Texas) can be divided into very few cells which essentially look near identical in all directions, coupled with few roads for the area and not much in the way of uploaded snaps coverage.
There's a lot of flat lands with spinifex in certain areas (easily a couple of United Kingdoms in size) and even a touch of mesa won't help narrow a location down from the general are as there are many of those with identical edge profiles.
But sure "Pilbara", "Kimberley", "Wheatbelt" can be geo guessed .. it's a real challenge to narrow down (I spent some time doing wet film photogrammetry prior to sheperding in WGS84 differential GPS locating and digital film and multi spectral geophysical aquisition).
No so hard if there's a few relatively unique man made features.
Because seem like someone who might have an interesting answer:
A GeoGuessr player, GeoWizard, has done a few “straight line challenges”, where he attempts to walk across a country in as straight a line as possible, usually planning beforehand with Google Earth and PostGIS. This got me thinking of what could fairly be thought of as “crossing”, since obviously you couldn’t describe e.g. walking from one side of Florida to the other as “crossing the USA”.
My best thought was to set the ending point of the line by following the border of the country in each direction til they met on the other side. To avoid the fractal coastline problem, use the challenger’s stride length as the unit of measure for the border.
But perhaps there is a better, more rigorous way of defining the opposite point on the edge of an arbitrary polygon.
As an abstract geometric problem the greatest width of an abstract polygon in a Euclidean 2D plane is found be looking at the greatest distance between all pairs of parallel lines that have been pulled together to clamp the polygon. The maximal diameter as opposed to the minimal waist.
Some might then say that "crossing" that polygon is to travel that longest line across the greatest width.
This simplistically avoids the question of concave polygons, complex polgons with exclusions (the Vatican state is removed from the Italian contry bounds), polgon collections (the nation of Fiji has many islands and can be tricky to traverse on foot .. not forgetting that perhaps the longest diameter might be from one island to another with no other islands between).
There's also the challenge of parallel lines on a 2D 'spherical' manifold, the almost spherical abstract ellipsoid of earth (or very non ellipsoidal Geoid if we take a constant gravitation value as the surface). On such manifolds lines are Great Circles (more or less) and always intersect.
Still, lets just say you're looking for the longest walkable(?) great circle path across a country that might go outside that country and perhaps is best travelled by a crop duster at 80m ground clearance to avoid getting feet wet.
The challenge itself takes some posing.
Meanwhile, less abstractly, I do like a jolly that "crosses a country" in a manner accepted by a (Wo)Man on a Clapham omnibus.
eg: https://en.wikipedia.org/wiki/Robyn_Davidson only went "half way", but that was accepted as an epic crossing. https://thelongridersguild.com/stories/stef-gebbie.htm "only" crossed most of the E-W distance across the lower portion of the country, while https://www.abc.net.au/news/2021-04-03/french-woman-conquers... travelled North - South, the long bit, but not quite coast to coast ( https://en.wikipedia.org/wiki/Bicentennial_National_Trail ).
I'm not ashamed to say that for about 6 months I played Geogessr thinking that you could not move around, but only look around... and I did pretty good in my mind.
Thats the only legit way to play IMO.
No moving, no panning, and no zooming is too hard. Moving is too boring. Pan and zoom is just the right balance.
A far more apropos comparison would be the internal combustion engine and the horse, in a military context. Though sticking with steam engines, military logistics advantaged over a wagon caravan.
The question here isn't a casual guessing game, but threat models (as directly addressed in TFA), and general informational hygiene.
I fail to see the drastic change here. AI was used for this in the past as well. The difference is far from that between steam and combustion engines.
Just like how chess engines ended competitive chess as people were predicting at the time.
"The question of whether a computer can think is no more interesting than the question of whether a submarine can swim." - Edsger Dijkstra
Isn’t a human brain an organic computer? I find the human brain very interesting. The most interesting thing I know of in the universe.
Religious people claim it's merely a communication device, and the computer is elsewhere.
Clearly one's interpretation is a function of the paradigm one exists in.
> Religious people claim it's merely a communication device, and the computer is elsewhere.
No, that is a seriously weird and imaginary perversion that you've invented, unless you have some citation to a doctrine. Christians do not consider body organs as "devices" nor do our churches teach doctrine on "external computers", if you refer to the Holy Trinity as such? Are you thinking of Latter-Day Saints mythology, or Scientology?
Your description is quite reductive in many aspects, including temporality and misuse of technical nomenclature. Thanks for telling religious people what we think and claim, though, and thanks for the amazing overgeneralized blanket dismissal.
In fact, humans used to be called "Computers" in terms of their job roles, that is, a human in an office was given math/physics problems to solve, and they'd use tools such as slide rule, paper and pencil to "compute" those problems and solve them.
There was more than one example of a fantasy "panopticon" by which a central observer or observers could watch everything going on, for example in a prison facility, and those observers could report findings to human computers, who would process the data and submit it up to the authorities, for meta-analysis and taking action on new developments or trends.
I believe they are referring to the soul. The soul according to doctrine does not exist inside the body for it can exist after the body's destruction.
> does not exist inside the body
Uh okay but let us not be reductive, because again you are confusing physicality and temporal space with spiritual reality, which is nuanced. A Christian would never say that “the soul [never] exists inside the body” because what do you mean by “inside”? That begs the question. Look up hylomorphism:
https://www.catholicculture.org/culture/library/dictionary/i...
https://en.wikipedia.org/wiki/Hylomorphism
While the soul, in death, can indeed exist apart from the body, even after the destruction of the physical human body, the soul is immortal, and lives on, in anticipation of its reuniting with that body. Whether in Heaven or Hell, the soul and body shall be reunited.
Even if the body is “destroyed” by physical means, that is a subjective judgement by human senses, and that glorified body will be reconstituted by Almighty God for everlasting unity with the soul.
> emergent property of the brain
Unfortunately, even this scientific definition is reductive. It ignores the theology of the Soul as an animating principle of everything that is created. Animals have souls [thus their name derived from “anima”]; rocks and stones have souls; trees have souls rivers have souls: everything that we can detect in this world, and every invisible thing, possesses a soul.
The difference with human souls regards their essence and immortality. But material souls are likewise considered souls in orthodox theology, and this should be considered by scientists.
Even in an “eternal death” in Hell, the body and the soul will be reunited there for eternal torture, body and soul.
Lets fast forward this conversation to its conclusion of “its open to interpretation”
and Christianity doesn't have a monopoly on religion or the fungible and unquantifiable soul concept
So the Internet is basically a series of tubes, and http packets are like mustard and ketchup that you pop into the tube, and it zooms along with air-pressure determined by the Chinese government. and then your bank puts dollar bills into the tubes on the Internet, and those tubes connect to vaults and safes underground, and that is how money gets into your account. Money doesn't come out of your account; it's just converted into dogecoin and then every User Interface automatically converts Dogecoin figures into USD before you can read it.
Also your computer is like a car. Your car works because there's a hamster or three in there and they keep them well-fed. Your computer also has a hamster, or a guinea pig sometimes. And your computer sends http packets over next door by means of carrier pigeons, or sometimes by semaphore.
Also men enjoy action-adventure films. Men enjoy films with lots of shooting. All men enjoy films with loud noises and special effects. Men like rock music with distorted guitars and loud drums. Men like loud soundsin general; that's why men are soldiers because we can blow stuff up and we listen to it.
sure - and
"The Unicorn is most particular in its affections and has a fondness for beautiful virgins..."
https://medium.com/luminasticity/meditation-on-the-unicorn-d...
Thank you for this comment. You have expressed this far more eloquently and magnanimously this I would have.
inside is a reference to a physical property, the brain is an organ that exists inside the body, we can discern which part of the body is brain and which is not, the soul evidently has never been so identified, although there have been conjectures throughout history that various body parts were the seat of the soul.
Not to put words in the parent’s mouth, but I think they were pointing to the idea of the soul or consciousness, concepts that mean very different things in religious vs. scientific contexts. One sees it as an immaterial essence; the other as an emergent property of the brain.
Thin client theory?
It's more of a hypothesis in this case. And some are clearly thicker than others. But that's details.
I don’t think “organic computer” is a good way to think of a brain. The fact that human brains + some scratch paper can implement any Turing-complete model of computation is very interesting. But that’s not true of chimpanzee brains or orca brains or crow brains, all of which belong to intelligent, thinking animals.
Going the other way, it’s not clear that a Turing-complete model of computation can tractably implement (say) crow cognition. Turing machines can solve arbitrary systems of Schrödinger equations, so theoretically we could simulate every atom in a crow’s body and get an AI crow[1]. But that’s obviously intractable for any known physical computer, and would remain intractable even if we moved to proteins rather than atoms. So are there higher-order “primitives” of crow cognition that can be implemented on a Turing machine? Or is the problem akin to integration, where neat “symbolic” solutions are impossible for most brains, and only “numeric” approaches work?
[1] This is also true for humans: even if you take a loopy quantum consciousness approach, AGI is theoretically possible. Sometimes you see people arguing against AGI on the grounds of Gödel-incompleteness, but this is a mystical nonsense understanding of what Gödel actually proved.
I think growing a baby from nothing is pretty nifty, but brains are up there too.
A brain that replicates itself is indeed nifty!
Normally I dislike these quips for HN; I hate that I love this one.
It is impressive and it almost located the church in my town properly, although in a neighbouring town. However that showed a lack of understanding because its conclusion about the location came from "reading" a signpost that pointed to that other village. Clearly there would be no signpost in a town pointing to itself. Still impressive and lots of correct observations about the subject like architectural details, roman numerals on the watch face etc.
Geoguesser was already a solved problem. I guarantee you ChatGPT is much worse than current systems designed to play Geoguesser (NMPZ).
Chatbots appear to have some amount of fluid intellgence so they can do impressive tasks with this information, the impressiveness of these tasks will likely increase in the future. But for simply getting a good score on Geoguesser it's not even close to hobby projects let alone state of the art.
One line of javascript solved that "problem" a while ago. The answer is in the DOM.
This isn't just referring to GeoGuessr the game, but locating photographs in general. The source post sums it up very reasonably and concisely:
> PSA: When posting any outdoors photos, update your threat model from "someone skilled and dedicated could theoretically locate this" to "any stalker can do this for 20€/mo"
I’ve been using LLMs for this kind of geo-guessing since Gemini 2.0. Even without access to internet search like o3, they perform surprisingly well.
I have found Google Lens fit for that purpose for years as mountains are fairly distinctive. Using less known landmarks or even random real estate photos, it doesn’t seem to be great.
I just tossed a few GeoGuessr places in it and it was confidently incorrect for all three. In one case it swears it knows the exact building and street. It’s thousands of KM off.
After trying with the free ChatGPT at least, I don't think this would be too much of an additional privacy risk. E.g. I submitted a photo I've taken with a feature with a particular architecture style in the front, and a city canal in the background, and told ChatGPT it was taken in China. It separated the feature and canal correctly, but still guessed wrong. I did remove the EXIF information, as ChatGPT tried to extract GPS info and failed at first.
I guess as long as there are no major visible features (e.g. a huge mountain), and you sanitized the metadata, you will be fine (regarding ChatGPT).
Have you tried this with o3? I think o3 is much better at this than any of the free models
I did not (OpenAI does not allow Chinese to pay for Plus due to politics, and I couldn't be bothered to try circumventing it), and I am curious about it because it does seem like a very difficult job.
I did try the latest Qwen though, and it was able to locate to the correct city, but it's still tens of km off (it guessed a tourist attraction in the city center instead of the actual district).
o3 is a massive difference indeed; i just took a random picture outside my country house (so the house is not on there) in the middle of nowhere, removed all the meta info, set my vpn to the other side of the world and tried it; it guessed it very close from a few far away landmarks. Very impressive.
I ran a few of these geo guess queries between OpenAI O3 and Gemini Pro 2.5 and O3 does much better on average.
It does spend an order of magnitude longer time on inference by searching through websites and analyzing the image but it often produces an impressive output. To me it also feels Gemini down samples the image as it tends to have a harder time reading small text vs O3.
That said O3 did tend to confidently say false things
Huh? It's not even that high on the leaderboard: https://geobench.org/
Using actual geoguesser data means using google street view data, so gemini being on top isn't too surprising.
Meanwhile O3 can not even count rocks in a picture. This is a commonly recurring theme -- ChatGPT does really well at few things considered hard by us but fails miserably at things even a child could do.
It's almost as if this is a non-human intelligence, which presents different strengths and weaknesses than human intelligence.
Is that really so surprising, considering the tremendous differences in underlying hardware and training process?
I think one cause of this (and some other issues with LLM use) is that people see it exhibiting one human-level trait, its capability to use language at a human level, and assume that it then comes with other human-level capabilities such as our ability to reason.
Do you not think it is interesting though? If I had asked you three years back which one of counting rocks vs geoguesser would ChatGPT beat humans on would you have answered correctly?
Somewhat tangential, what is the purpose of the lengthy prompt used in this post? Would not "location of this image" have worked just as well?
Harder than GeoGuesser is tagging a unique place from a photo even with gps exif data. Try this with bakeries in Seoul or sushi in Japan where there’s high density of same places.
Justifying my link indirection: the Fediverse link both provides the critical context here, and is more generally accessible than the upstream source.
What kind of neural network architecture could replicate the same behaviour? Talking generically about image similarities
In my experience 4o was already really good at this task. I'd be curious to see an in-depth 4o vs. o3 benchmark.
What does this even matter?
You can just feed literally all of google street view into a traditional ML model ...
Can you?
Aw that AI ruins all the games for us !! Now we wont know if the opponent has been using it
The Game is significantly larger than the game, in a Kipplingesque sense.
Can it guess Senegal by sky color though?
somehow location tags on my photos got turned off for a bit. i would use this to restore them
You could, but that would also be a rather silly idea. Your rough estimate for where photo was taken is going to be more accurate than what the model will produce.
One thing that screws it up is ironically the memory across conversations. I gave it some _really_ obscure photos from some godforsaken ass cracks of the world I've been to, and it was able to guess most of them correctly. However, in its reasoning trace I saw that it's having a heck of a time letting go of the idea that the photo N is not roughly from the same location as the previous photos. What's even more impressive, even when it guesses incorrectly, it can often follow up and guess correctly with minimal hints. And it reasons about it much like a human would, and searches for the same things a human would. Note: I used o3, YMMV with a smaller/weaker model.
[dead]