I'd consider these the vision-language equiv of natural adversarial examples; back when imagenet classification was the big thing there was some cool work wrt natural objects in weird places being much harder to classify (e.g. a bullfrog in some grass instead of water)
Probably no embodied learning? No way to learn the lines are different unless you spent time measuring them before. Would love to hear what @melaniemitchell.bsky.social or @dileeplearning.bsky.social have to say.
This is similar to language-only⬇️ "illusion illusions" like when you give it a version of the "Monty Hall" problem where it makes no sense to switch doors, but the LLM "thinks" it's the original Monty Hall problem so says "switch". Lots of other examples of this: ⬇️
e.g., "goat-farmer-cabbage" river-crossing type puzzles. The LLM recognizes statistical signature of a task it "knows" and then gives the answer to that task, rather than the actual task it is given. Perhaps a kind of "ember of autoregression" as in McCoy et al. paper?
LLMs predict an answer similar to a question similar to the questions in the training set; the prompt has to override the stochastic nature of the training when you want something that looks like what's in the training set but *isn't*
It did the same thing if I write the lines out in dashes, which is interesting because I wouldn’t think that this illusion comes up too frequently in text-only mediums
One of the findings from the WEIRD people paper a while back was that blew my mind was that this, and other optical illusion, don't necessarily translate across cultures. They suggested that the prevalence of carpenters corners over natural shapes in the individuals environment was a major factor.
Oh is this related to the study with the group of urban western kids raised with visible right angles vs groups of kids outside built areas (where all the buildings are straw huts or similar) and how their perceptions of shapes are different?
It was an overall review of literature on why behavioral studies has a freaking GIGANTIC blind spot, which shows up in places no one is even considering, but I do think that one of the studies they relied on. It was a really great article that got a decent bit of coverage before things caught fire.
Sorry, I say the "Free PDF" link without checking if it really worked. Does Bluesky have a way of requesting paywalled journal articles from academics with access yet?
That is peak gaslighting to the editor
"There are not the same length"
"Well you see editor, they are, that is what this chapter is about"
"No no, they are off by millimeters"
"Yeah okay editor, you are just confused lol"
I think this might illustrate the limits of LLM AI. With no model of the world the context of the question cannot be 'understood' - can it actually lead to real world inference? Specific examples like the river crossing problem can be added, but generality will I believe be out of reach 🤔
My boss wanted us to use AI assisted writing to speed up our writing process. Fact-checking what the AI wrote took the same time that researching and writing the text ourselves would have. Sometimes those mistakes are sneaky, esp. if you're a journalist writing about a topic you're not an expert in.
Qwen2.5-VL-72B-Instruct perfectly answered the question: The red circle is bigger than the blue circle. The red circle is surrounded by smaller black circles, which makes it appear larger in comparison to the blue circle, which is much smaller and surrounded by larger black circles.
I love that it answered your question correctly—the lines are straight—but couldn't resist screwing it up by claiming they're also parallel, which you hadn't even asked about.
This one is amazing, since no computer has ever actually seen those illusory gray rectangles! Are there images like this, illustrating the illusory percept, that would be in the training data?
Right, my point is that existing images of this illusion don't have any gray squares, so I wonder if the LLM has seen versions of the image WITH the gray squares in its input. Maybe to illustrate what the effect looks like in a psych textbook or something
ohhh I see now; I can't say for sure but I kind of doubt it, given the other illusion-illusions I think it is sufficient to pair rough pattern matching of the grid illusion with text like "there seem to be squares there but there aren't"
The LLM was trained on a lot of text describing said illusions, often accompanied by a similar question as OP is using.
Thus the autocorrect-on-steroids does what it does best and it copies those texts ad verbum.
good q! I didn't run an extensive check, but asking about an image without uploading an image prompts you to upload the image; though, it does "get" the question is pointing to an illusion
Maybe because without as many statistically likely off the rack answers, the smaller models have to weigh and match the distribution of the actual picture you enter in
It also works with "cognitive illusions", at least in version 3.5. For example, you can trick it to choose "change door" in a modified Monty Hall problem where all doors are initially opened and everyone can see the car behind door #1.
The kind test question that checks reading comprehension, or more importantly, the ability of the subject to slow down and think through the language used rather than reaching for the first pattern that occurs to them.
Did it actually pick a choice of behavior in its imagined scenario?
4o "fixes" some kinds of these problems like "which is heavier, a pound of bricks or two pounds of feathers?" but consistently fails a version of Monty Hall modified slightly such that the answer is 1/2. The latter I discuss in section 3 of this article. https://nickchk.substack.com/p/ai-errors-are-statistical-errors
Problem is... the red one is shorter in this presentation. Measure it... from the intersection of the arrowhead to the other intersection of the arrowhead on both the blue and the red. The blue is four times longer at least.
This is an instructive illustration of crucial differences between humans and LLMs. However, we should keep in mind that just as with humans, it matters a lot how we ask LLMs questions. Also, it is worth noting that similar familiarity/fluency processes have long been demonstrated in humans, too.
This is not familiarity but technical limitation of the LLM (totally different quality). If the tone of the question determins the correctness of the answer one can only ask question to which one knows the answer already - which is pointless obviously. LLMs answer by proximity only and not by logic.
Right, these are all close to classic optical illusions, but... broken. The AI doesn't realize they're broken (so it describes them as if they were the original illusions).
Because LLMs use computer logic, they can't intuit that the situation is novel. They can't proact, only react using existing responses. The models are "large" to hide the dialogue tree from the user.
I’m not defending LLMs (nor promoting them), but the sophistication goes well beyond the concept of a list of preformed responses. There’s no dialog tree. It’s algorithmic, in the realm of neural networks, with nodes having metadata and weighted values etc. leading to the output.
I was being superlative, but it still doesn't think like a human. That part is pure smoke and mirrors. The output of software using the models is predictable, but the size of the model requires a machine do the prediction. "Dialog tree" was meant pejoratively.
Why do you say the output is predictable? The blackbox nature of most of these systems, and their nondeterministic unpredictable output, is exactly why nobody should be trusting their output.
sure thing, though I would caveat that this isn't a robust testing, just trying some stuff, and people have made similar points on similar issues of course
Comments
Dave: ::::ejected into space :::: silent scream.
Can we archive intelligence from pattern matching?
https://arxiv.org/pdf/1907.07174
https://bsky.app/profile/drgbuckingham.bsky.social/post/3larox36dbk2p
"There are not the same length"
"Well you see editor, they are, that is what this chapter is about"
"No no, they are off by millimeters"
"Yeah okay editor, you are just confused lol"
This is literally a thread about the limits. It shows, with concrete examples, exactly where some of the limits are.
There is, in the opinion of many of us, no way to overcome this.
"MOAR DATA" won't work. LLMs are a dead end.
No way they are equal.
https://en.wikipedia.org/wiki/Grid_illusion
Thus the autocorrect-on-steroids does what it does best and it copies those texts ad verbum.
https://bsky.app/profile/standwithrobots.bsky.social/post/3lcay5hbza22w
(maybe we should call it the non-problem problem)
(also, to be a tiny bit charitable, experts probably have this issue too, sort of)
Did it actually pick a choice of behavior in its imagined scenario?
I wouldn't waste that on useless things.
https://bsky.app/profile/drchristaylor.bsky.social/post/3lc7fwo4pl42u