Remix.run Logo
actuallyalys 2 days ago

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.

Kolya 2 days ago | parent | next [-]

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.

sadeshmukh 2 days ago | parent | next [-]

40 km is imo pretty impressive, but the 5 minute is really a killer. No use in real time applications

antonvs 2 days ago | parent | next [-]

For real time applications, GPS seems like it'd be much more useful.

bpodgursky 2 days ago | parent | prev [-]

The military/infosec uses of this are not real time. You can wait 5 minutes for a drone strike.

viraptor 2 days ago | parent | next [-]

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)

bpodgursky 2 days ago | parent [-]

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.

viraptor 2 days ago | parent | next [-]

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...

abletonlive 2 days ago | parent | prev | next [-]

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.

wruza 2 days ago | parent | next [-]

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.

abletonlive a day ago | parent [-]

Yes because agnostic atheism is a logically superior belief system, that's much more intellectually honest and ironically: humble.

wruza a day ago | parent [-]

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.

namaria 14 hours ago | parent [-]

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"?

bpodgursky 2 days ago | parent | prev [-]

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.

abletonlive 2 days ago | parent [-]

> 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

neom 2 days ago | parent [-]

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.

hhh 2 days ago | parent | prev [-]

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.

tempest_ 2 days ago | parent | prev | next [-]

Yeah but you also cant be off by 40km with your drone strike.

dogma1138 a day ago | parent | next [-]

Narrowing it down to even a 100km radius for visint analysts to then pinpoint it down is worth its weight in gold already.

dredmorbius 2 days ago | parent | prev | next [-]

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.

hhh 2 days ago | parent | prev [-]

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.

swarnie 2 days ago | parent | prev [-]

Please stop indiscriminately murdering us with flying killdroids.

TiredOfLife 2 days ago | parent | next [-]

Stop invading Ukraine

swarnie a day ago | parent [-]

Haven't done for almost 200 years.

DonHopkins 2 days ago | parent | prev [-]

https://www.youtube.com/watch?v=zbnDr_IbdIU

paulcole 2 days ago | parent | prev | next [-]

That’s also not GeoGuessr.

casey2 2 days ago | parent | prev | next [-]

Were you in a Google street view car when you took the photo?

nxobject 2 days ago | parent | prev [-]

Did you glean anything interesting from the chain of thought about why it took so long?

xzjis a day ago | parent | prev | next [-]

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.

echelon 2 days ago | parent | prev | next [-]

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

cwmoore 2 days ago | parent | prev [-]

New title is pretty accurate: "now performs well". Another amenable HN solution.