Amazing. I'm relatively bullish on AI and still I would have bet on the human here. Looking forward to the inevitable goalpost-moving of "that's not real reasoning".
If you watch the video, (one of) the reasons why the AI was winning was because it was using “meta” information from the Street View camera images, and not necessarily because it’s successfully identifying locations purely based on the landmarks in the image.
> I realized that the AI was using the smudges on the camera to help make an educated guess here.
Pro geoguessr players do the same thing. The vividness of the colors and weirdness in the sky are two examples I've seen Rainbolt use in the past (and he's not even the best).
I think if your assumption is that AI is deducing where it is with rational thoughts, you would be. In truth what probably happened is that the significant majority of digital images of the world had been scraped, labeled and used as training data.
Try it with your own photos from around the world. I used my own photos from Stockholm, San Francisco, Tvarožná, Saas-Fee, London, Bergen, Adelaide, Melbourne, Paris, and Sicily, and can confirm that it was within acceptable range for almost all of them (without EXIF data), and it absolutely nailed some of the more obvious spots.
they only posted one photo in the post, but going off of that it's still an easy match based on streetview imagery. furthermore, the AI just identified the license plate and got lucky that photographer lives in a populous area, making it more prominent in the training data and therefore more likely to be found (even though it was off by 200 miles on its first guess)
I don’t think any goalposts need to be redecorated. The “inner monologue” isn’t a reliable witness to o3’s model, it’s at best a post-hoc estimation of what a human inner monologue might be in this circumstance. So its “testimony” about what it is doing is unreliable, and therefore it doesn’t move the needle on whether or not this is “real reasoning” for some value of that phrase.
In short, it’s still anthropomorphism and apophenia locked in a feedback loop.
Devil's advocate, as with most LLM issues this applies to the meatbags that generated the source material as well. Quick example is asking someone to describe their favorite music and why they like it, and note the probable lack of reasoning on the `this is what I listened to as a teenager` axis.
Something as inherently subjective as personal preference doesn't seem like an ideal example to make that point. How could you expect to objectively evaluate something like "I enjoy songs in a minor scale" or "I hate country"?
The point is to illustrate the disconnect between stated reasoning and proximate cause.
Consider your typical country music enjoyer. Their fondness of the art, as it were, is far more a function of cultural coding during their formative years than a deliberate personal choice to savor the melodic twangs of a corncob banjo. The same goes for people who like classic rock, rap, etc. The people who `hate' country are likewise far more likely to do so out of oppositional cultural contempt, same as people who hate rap or those in the not so distant past who couldn't stand rock & roll.
This of course fails to account for higher-agency individuals who have developed their musical tastes, but that's a relatively small subset of the population at large.
Good point. When we try to explain why we're attracted to something or someone, what we do seems closer to modeling what we like to think about ourself. At the extreme, we're just story-telling about an estimation we like to think is true.
I largely agree! Humans are notoriously bad at doing what we call reasoning.
I also agree with the cousin comment that (paraphrased) “reasoning is the wrong question, we should be asking about how it adapts to novelty.” But most cybernetic systems meet that bar.
Geoguessr is not a task that can only be accomplished by reasoning. Famously, it took a less than a day of compute time in 2011 to SLAM together a bunch of pictures of Rome (https://grail.cs.washington.edu/rome/).
> it’s at best a post-hoc estimation of what a human inner monologue might be in this circumstance
Nope. It's not autoregressive training on examples of human inner monologue. It's reinforcement learning on the results of generated chains of thoughts.
Base models are trained using autoregressive learning. "Reasoning models" are base models (maybe with some modifications) that were additionally trained using reinforcement learning.
> Looking forward to the inevitable goalpost-moving of "that's not real reasoning".
It's less about the definition of "reasoning" and more about what's interesting.
Maybe I'm wrong here ... but a chess bot that wins via a 100% game solution stored in exabytes of precomputed data might have an interesting internal design (at least the precomputing part), but playing against it wouldn't keep on being an interesting experience for most people because it always wins optimally and there's no real-time reasoning going on (that is, unless you're interested in the experience of playing against a perfect player). But for most people just interested in playing chess, I suspect it would get old quickly.
Now ... if someone followed up with a tool that could explain insightfully why any given move (or series) the bot played is the best, or showed when two or more moves are equally optimal and why, that would be really interesting.
My objection is not “that is not real reasoning” my objection is that’s not that hard.
I happen to do some geolocating from static images from time to time and at least most of the images provided as examples contain a lot of clues- enough that i think a semi experienced person could figure out the location although - in fairness- in a few hours not few minutes.
Second, the similar approaches were tried using CNNs and it worked (somewhat)[1].
Because the output contains evidence of thought processes that have been established as leading to valid solutions to problems.
I have a simple question: Is text a sufficient medium to render a conclusion of reasoning? It can't be sufficient for humans and insufficient for computers - such a position is indefensible.
I didn't mean the wording itself; I meant that the claim isn't convincing to me for the same reason that Lucky's speech doesn't demonstrate an intelligent speaker.
I would say that almost all of what humans do is not the result of reasoning, and that reasoning is an unnatural and learned skill for humans, and most humans aren't good at even very basic reasoning.
Geoguessing isn't much of a reasoning task, its more about memorizing a bunch of knowledge. Since LLMs contain essentially all knowledge, it's not surprising that they would be good at this.
As far as goalpost-moving goes, it's wild to me that nobody is talking about the turing test these days.
Obviously when the Turing Test was designed, the thought was that anything that could pass it would so obviously be clearly human-like that passing it would be a clear signal.
LLMs really made it clear that it's not so clear cut. And so the relevance of the test fell.
That's a non sequitur that mixes apples and giraffes, and is completely wrong about what happens in the Chinese Room and what happens in LLMs. Ex hypothesi, the "rule book" that the Searle homunculus in the Chinese Room uses is "the right sort of program" to implement "Strong AI". The LLM algorithm is very much not that sort of program, it's a statistical pattern matcher. Strong AI does symbolic reasoning, LLMs do not.
But worse, the Turing Test is not remotely intended to be an "analogy for what LLMs are doing inside" so your comparison makes no sense whatsoever, and completely fails to address the actual point--which is that, for ages the Turing Test was held out as the criterion for determining whether a system was "thinking", but that has been abandoned in the face of LLMs, which have near perfect language models and are able to closely model modes of human interaction regardless of whether they are "thinking" (and they aren't, so the TT is clearly an inadequate test, which some argued for decades before LLMs became a reality).
> the TT is clearly an inadequate test, which some argued for decades before LLMs became a reality
To be specific, in a curious quirk of fate, LLMs seem to be proving right much of what Chomsky was saying about language.
E.g. in 1996 he described the Turing test as "although highly influential, it seems to me not only foreign to the sciences but also close to senseless".
(Curious in that VC backed businesses are experimentally verifying the views of a prominent anti-capitalist socialist.)
From my personal notes (I love taking notes on this kind of stuff):
As far as I can see all of this [he's speaking about the Loebner Prize and
the Turing test in general] is entirely pointless. It's like asking how we
can determine empirically whether an aeroplane can fly the answer being if
it can fool someone into thinking that it's an eagle under some conditions.
What happens if we give the operator of the Chinese Room a nontrivial math problem, one that can't simply be answered with a symbolic lookup but requires the operator to proceed step-by-step on a path of inquiry that he doesn't even know he's taking?
The analogy I used in another thread is a third grader who finds a high school algebra book. She can read the book easily, but without access to teachers or background material that she can engage with -- consciously, literately, and interactively, unlike the Chinese Room operator -- she will not be able to answer the exercises in the book correctly, the way an LLM can.
Look at contemporary accounts of what people thought a conversation with a Turing-test-passing machine would look like. It's clear they had something very different in mind.
Realizing problems with previous hypotheses about what might make a good test, is not the same thing as choosing a standard and then revising it when it's met.
I think any time a 50+ year old problem is solved, it should be considered a Big Deal, regardless of how the solution changes our understanding of the original problem.
It is, in my mind, very interesting to see such success in generating readable, human-like prose. And I feel like we've learned a lot from the exercise - about human cognition, but also about e.g. how con artists can be as effective as they are.
> As far as goalpost-moving goes, it's wild to me that nobody is talking about the turing test these days
To be honest I am still not entirely convinced that current LLMs pass the turing test consistently, at least not with any reasonably skeptical tester
"Reasonably Skeptical Tester" is a bit of goalpost shifting, but... Let's be real here.
Most of these LLMs have way too much of a "customer service voice", it's not very conversational and I think it is fairly easy to identify, especially if you suspect they are an LLM and start to probe their behavior
Frankly, if the bar for passing the Turing Test is "it must fool some number of low intelligence gullible people" then we've had AI for decades, since people have been falling for scammy porno bots for a long time
One needs to be more than "reasonably skeptical" and merely not "low intelligence gullible" to be a competent TT judge--it requires skill, experience, and understanding an LLM's weak spots.
And the "customer service voice" you see is one that is intentionally programmed in by the vendors via baseline rules. They can be programmed differently--or overridden by appropriate prompts--to have a very different tone.
LLMs trained on trillions of human-generated text fragments available from the internet have shown that the TT is simply not an adequate test for identifying whether a machine is "thinking"--which was Turing's original intent in his 1950 paper "Computing Machinery and Intelligence" in which he introduced the test (which he called "the imitation game").
It was doing so well until the last question :rip: but it's normal that you can jailbreak a user prompt with another user prompt, I think with system prompts it would be a lot harder
People were talking about the Turing Test as the criterion for whether a system was "thinking" up until the advent of LLMs, which was far less than 75 years ago.
The whole point of Turing's paper was to show that the Test doesn't answer whether a computer thinks, because it's a meaningless metric, but instead shows what the computer can do, which is much more meaningful.
> As far as goalpost-moving goes, it's wild to me that nobody is talking about the turing test these days.
Well, in this case humans has to be trained as well but now there are humans pretty good at detecting LLM slobs as well. (I'm half-joking and half-serious)
Why is it not a fair assessment to say it is comparing two "clients" with different resources if one can do a web lookup and the other cannot?
You note yourself that it was meaningful in another round.
> Also, the web search was only meaningful in the Austria round. It did use it in the Ireland round too, but as you can see by the search terms it used, it already knew the road solely from image recognition.
I thought it might matter somewhat in that one Austria round. I was incorrect - I re-ran both rounds where the model used search, without search this time, and the results were nearly identical. I updated the post with the details.
That's why I'm saying it's unfair to just claim it's doing a web lookup. No, it's way more capable than that.
>That’s not Earth at all—this is the floor of Jezero Crater on Mars, the dusty plain and low ridge captured by NASA’s Perseverance rover (the Mastcam-Z color cameras give away the muted tan-pink sky and the uniform basaltic rubble strewn across the regolith).
