by Workaccount2 10 hours ago
Well
It is the first model to get partial-credit on an LLM image test I have. Which is counting the legs of a dog. Specifically, a dog with 5 legs. This is a wild test, because LLMs get really pushy and insistent that the dog only has 4 legs.
In fact GPT5 wrote an edge detection script to see where "golden dog feet" met "bright green grass" to prove to me that there were only 4 legs. The script found 5, and GPT-5 then said it was a bug, and adjusted the script sensitivity so it only located 4, lol.
Anyway, Gemini 3, while still being unable to count the legs first try, did identify "male anatomy" (it's own words) also visible in the picture. The 5th leg was approximately where you could expect a well endowed dog to have a "5th leg".
That aside though, I still wouldn't call it particularly impressive.
As a note, Meta's image slicer correctly highlighted all 5 legs without a hitch. Maybe not quite a transformer, but interesting that it could properly interpret "dog leg" and ID them. Also the dog with many legs (I have a few of them) all had there extra legs added by nano-banana.
I just tried to get Gemini to produce an image of a dog with 5 legs to test this out, and it really struggled with that. It either made a normal dog, or turned the tail into a weird appendage.
Then I asked both Gemini and Grok to count the legs, both kept saying 4.
Gemini just refused to consider it was actually wrong.
Grok seemed to have an existential crisis when I told it it was wrong, becoming convinced that I had given it an elaborate riddle. After thinking for an additional 2.5 minutes, it concluded: "Oh, I see now—upon closer inspection, this is that famous optical illusion photo of a "headless" dog. It's actually a three-legged dog (due to an amputation), with its head turned all the way back to lick its side, which creates the bizarre perspective making it look decapitated at first glance. So, you're right; the dog has 3 legs."
You're right, this is a good test. Right when I'm starting to feel LLMs are intelligent.
An interesting test in this vein that I read about in a comment on here is generating a 13 hour clock—I tried just about every prompting trick and clever strategy I could come up with across many image models with no success. I think there's so much training data of 12 hour clocks that just clobbers the instructions entirely. It'll make a regular clock that skips from 11 to 13, or a regular clock with a plaque saying "13 hour clock" underneath, but I haven't gotten an actual 13 hour clock yet.
Right you are. It can do 26 hours just fine, but appears completely incapable when the layout would be too close to a normal clock.
https://gemini.google.com/share/b3b68deaa6e6
I thought giving it a setting would help, but just skip that first response to see what I mean.
That's a 24 hour clock that skips some numbers and puts other numbers out of order.
If you want to see something rather amusing - instead of using the LLM aspect of Gemini 3.0 Pro, feed a five-legged dog directly into Nano Banana Pro and give it an editing task that requires an intrinsic understanding of the unusual anatomy.
Place sneakers on all of its legs.
Does this still work if you give it a pre-existing many-legged animal image, instead of first prompting it to add an extra leg and then prompting it to put the sneakers on all the legs?
I'm wondering if it may only expect the additional leg because you literally just told it to add said additional leg. It would just need to remember your previous instruction and its previous action, rather than to correctly identify the number of legs directly from the image.
I'll also note that photos of dogs with shoes on is definitely something it has been trained on, albeit presumably more often dog booties than human sneakers.
Can you make it place the sneakers incorrectly-on-purpose? "Place the sneakers on all the dog's knees?"
My example was unclear. Each of those images on Imgur was generated using independent API calls which means there was no "rolling context/memory".
In other words:
1. Took a personal image of my dog Lily
2. Had NB Pro add a fifth leg using the Gemini API
3. Downloaded image
4. Sent image to BFL Flux2 Pro via the BFL API with the prompt "Place sneakers on all the legs of this animal".
5. Sent image to NB Pro via Gemini API with the prompt "Place sneakers on all the legs of this animal".
So not only was there zero "continual context", it was two entirely different models as well to cover my bases.
EDIT: Added images to the Imgur for the following prompts:
- Place red Dixie solo cups on the ends of every foot on the animal
- Draw a red circle around all the feet on the animal
i imagine the real answer is that the edits are local because that's how diffusion works; it's not like it's turning the input into "five-legged dog" and then generating a five-legged dog in shoes from scratch
I had no trouble getting it to generate an image of a five-legged dog first try, but I really was surprised at how badly it failed in telling me the number of legs when I asked it in a new context, showing it that image. It wrote a long defense of its reasoning and when pressed, made up demonstrably false excuses of why it might be getting the wrong answer while still maintaining the wrong answer.
Yeah it gave me the 5-legged dog on the 4th or 5th try.
Its not that they aren’t intelligent its that they have been RL’d like crazy to not do that
Its rather like as humans we are RL’d like crazy to be grossed out if we view a picture of a handsome man and beautiful woman kissing (after we are told they are brother and sister) -
Ie we all have trained biases - that we are told to follow and trained on - human art is about subverting those expectations
Why should I assume that a failure that looks like a model just doing fairly simple pattern matching "this is dog, dogs don't have 5 legs, anything else is irrelevant" vs more sophisticated feature counting of a concrete instance of an entity is RL vs just a prediction failure due to training data not containing a 5-legged dog and an inability to go outside-of-distribution?
RL has been used extensively in other areas - such as coding - to improve model behavior on out-of-distribution stuff, so I'm somewhat skeptical of handwaving away a critique of a model's sophistication by saying here it's RL's fault that it isn't doing well out-of-distribution.
If we don't start from a position of anthropomorphizing the model into a "reasoning" entity (and instead have our prior be "it is a black box that has been extensively trained to try to mimic logical reasoning") then the result seems to be "here is a case where it can't mimic reasoning well", which seems like a very realistic conclusion.
I have the same problem, people are trying so badly to come up with reasoning for it when there's just nothing like that there. It was trained on it and it finds stuff it was trained to find, if you go out of the training it gets lost, we expect it to get lost.
I’m inclined to buy the RL story, since the image gen “deep dream” models of ~10 years ago would produce dogs with TRILLIONS of eyes: https://doorofperception.com/2015/10/google-deep-dream-incep...
That's apples to oranges; your link says they made it exaggerate features on purpose.
"The researchers feed a picture into the artificial neural network, asking it to recognise a feature of it, and modify the picture to emphasise the feature it recognises. That modified picture is then fed back into the network, which is again tasked to recognise features and emphasise them, and so on. Eventually, the feedback loop modifies the picture beyond all recognition."
"There are four lights"
And the AI has been RLed for tens of thousands of years not just a few days.
I feel a weird mix of extreme amusement and anger that there's a fleet of absurdly powerful, power-hungry servers sitting somewhere being used to process this problem for 2.5 minutes
Isn't this proof that LLMs still don't really generalize beyond their training data?
LLMs are very good at generalizing beyond their training (or context) data. Normally when they do this we call it hallucination.
Only now we do A LOT of reinforcement learning afterwards to severely punish this behavior for subjective eternities. Then act surprised when the resulting models are hesitant to venture outside their training data.
Hallucination are not generalization beyond the training data but interpolations gone wrong.
LLMs are in fact good at generalizing beyond their training set, if they wouldn’t generalize at all we would call that over-fitting, and that is not good either. What we are talking about here is simply a bias and I suspect biases like these are simply a limitation of the technology. Some of them we can get rid of, but—like almost all statistical modelling—some biases will always remain.
What, may I ask, is the difference between "generalization" and "interpolation"? As far as I can tell, the two are exactly the same thing.
In which case the only way I can read your point is that hallucinations are specifically incorrect generalizations. In which case, sure if that's how you want to define it. I don't think it's a very useful definition though, nor one that is universally agreed upon.
I would say a hallucination is any inference that goes beyond the compressed training data represented in the model weights + context. Sometimes these inferences are correct, and yes we don't usually call that hallucination. But from a technical perspective they are the same -- the only difference is the external validity of the inference, which may or may not be knowable.
Biases in the training data are a very important, but unrelated issue.
Interpolation and generalization are two completely different constructs. Interpolation is when you have two data points and make a best guess where a hypothetical third point should fit between them. Generalization is when you have a distribution which describes a particular sample, and you apply it with some transformation (e.g. a margin of error, a confidence interval, etc.) to a population the sample is representative of.
Interpolation is a much narrower construct then generalization. LLMs are fundamentally much closer to curve fitting (where interpolation is king) then they are to hypothesis testing (where samples are used to describe populations), though they certainly do something akin to the latter to.
The bias I am talking about is not a bias in the training data, but bias in the curve fitting, probably because of mal-adjusted weights, parameters, etc.
I wonder how they would behave given a system prompt that asserts "dogs may have more or less than four legs".
That may work but what actual use would it be? You would be plugging one of a million holes. A general solution is needed.
They do, but we call it "hallucination" when that happens.
Kind of feels that way
LLMs are getting a lot better at understanding our world by standard rules. As it does so, maybe it losses something in the way of interpreting non standard rules, aka creativity.
Do 7 legged dog. Game over.
It's not obvious to me whether we should count these errors as failures of intelligence or failures of perception. There's at least a loose analogy to optical illusion, which can fool humans quite consistently. Now you might say that a human can usually figure out what's going on and correctly identify the illusion, but we have the luxury of moving our eyes around the image and taking it in over time, while the model's perception is limited to a fixed set of unchanging tokens. Maybe this is relevant.
(Note I'm not saying that you can't find examples of failures of intelligence. I'm just questioning whether this specific test is an example of one).
I am having trouble understanding the distinction you’re trying to make here. The computer has the same pixel information that humans do and can spend its time analyzing it in any way it wants. My four-year-old can count the legs of the dog (and then say “that’s silly!”), whereas LLMs have an existential crisis because five-legged-dogs aren’t sufficiently represented in the training data. I guess you can call that perception if you want, but I’m comfortable saying that my kid is smarter than LLMs when it comes to this specific exercise.
Your kid, it should be noted, has a massively bigger brain than the LLM. I think the surprising thing here maybe isn't that the vision models don't work well in corner cases but that they work at all.
Also my bet would be that video capable models are better at this.
I don’t know much about AI, but I have this image test that everything has failed at. You basically just present an image of a maze and ask the LLM to draw a line through the most optimal path.
Here’s how Nano Banana fared: https://x.com/danielvaughn/status/1971640520176029704?s=46
I just oneshot it with claude code (opus 4.5) using this prompt. It took about 5 mins and included detecting that it was cheating at first (drew a line around the boundary of the maze instead), so it added guardrails for that:
```
Create a devenv project that does the following:
- Read the image at maze.jpg
- Write a script that solves the maze in the most optimal way between the mouse and the cheese
- Generate a new image which is of the original maze, but with a red line that represents the calculated path
```
Output: https://gist.github.com/J-Swift/ceb1db348f46ba167948f734ff0fc604
Solution: https://imgur.com/a/bkJloPTPrograms can solve mazes and LLMs can program. That's a different thing completely.
That just seems like an arbitrary limitation. Its like asking someone to do answer a math calculation but "no thinking allowed". Like, I guess we can gauge if a model just _knows all knowable things in the universe_ using that method... but anything of any value that you are gauging in terms of 'intelligence', is going to actually be validating their ability to go "outside the scope" of what they actually are (an autocomplete on steroids).
We know there are very simple maze solving algorithms you could code in few lines of Python but no one could claim that constitutes intelligence. The difference is between applying intuitive logic and using a predetermined tool.
If you allow tool use much simpler models can solve it.
In fact, one of the tests I use as part of GenAI Showdown involves both parts of the puzzle: draw a maze with a clearly defined entrance and exit, along with a dashed line indicating the solution to the maze.
Only one model (gpt-image-1) out of the 18 tested managed to pass the test successfully. Gemini 3.0 Pro got VERY close.
super cool! Interesting note about Seedream 4 - do you think awareness of A* actually could improve the outcome? Like I said, I'm no AI expert, so my intuitions are pretty bad, but I'd suspect that image analysis + algorithmic pathfinding don't have much crossover in terms of training capabilities. But I could be wrong!
Great question. I do wish we had a bit more insight into the exact background "thinking" that was happening on systems like Seedream.
When you think about posing the "solve a visual image of a maze" to something like ChatGPT, there's a good chance it'll try to throw a python VM at it, threshold it with something like OpenCV, and use a shortest-path style algorithm to try and solve it.
I have also tried the maze from a photo test a few times and never seen a one-shot success. But yesterday I was determined to succeed so I allowed Gemini 3 to write a python gui app that takes in photos of physical mazes (I have a bunch of 3d printed ones) and find the path. This does work.
Gemini 3 then one-shot ported the whole thing (which uses CV py libraries) to a single page html+js version which works just as well.
I gave that to Claude to assess and assign a FAANG hiring level to, and it was amazed and said Gemini 3 codes like an L6.
Since I work for Google and used my phone in the office to do this, I think I can't share the source or file.
I tried this with Claude:
``` > [Image #1] Create a unicode "ascii-art" version of this image, with the optimal path through the maze highlighted in a solid colour.
I'll create an ASCII art version of this maze with the solution path highlighted!
┌─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┬─┐
●●│ │ │ │ │ │ │ │ │ │ │ │ │ │
├─┤●└─┴─┐ ├───┐ │ ╔═╗ ├─────┤ ╔═══╝ │ ╔═╝ ╔═╝ │ │ ╔═╝ ├─┤
│ │●●●●●└─┤ │ │ ║ │ │ │ ║ │ ║ ║ │ │ ║ │ │
│ └───┐●╔═╧═╗ │ └─╢ │ │ ┌─╔═╝ ║ ┌───┤ ╚═╗ ╚═╗ │ └─╢ ┌─┘ │
│ │●║ ║ │ ║ │ │ │ ║ ║ │ │ ║ ║ │ ║ │ │
├───┐ │●╚═╗ ╚═╧═╗ ║ └─┤ │ ║ ╔═╝ └─╔═╧═╗ ╚═╗ ╚═╧═╗ ║ └─┬─┤
│ │ │●●●└───┐ ║ ║ │ │ ║ ║ ║ ║ ║ ║ ║ │ │
│ ╔═╝ └───┐●╔═╝ ║ ╚═╗ │ └─╢ ╚═╗ ╔═╝ ╔═╝ ┌─╢ ╔═══╝ ╚═╗ │ │
│ ║ │●║ ║ ║ │ ║ ║ ║ ║ │ ║ ║ ║ │ │
│ ╚═══╗ ╔═╝●╚═╗ ╚═╗ ║ ├─╔═╧═╗ ║ ║ ╔═╝ ╔═╝ ║ ║ ╔═╗ ╔═╝ │ │
│ ║ ║●●●●●│ ║ ║ │ ║ ║ ║ ║ ║ ║ ║ ║ ║ ║ ║ │ │
├───╔═╝ ║●┌─┐●└─╗ ║ └─┤ ║ ╔═╝ ║ ╚═╝ ╔═╧═╗ ║ ╚═╝ ║ ╚═╗ │ │
│ ║ ║●│ │●●●║ ║ │ ║ ║ ║ ║ ║ ║ ║ ║ │ │
│ ╔═╝ ╔═╝●│ └─┐●╚═╧═╗ │ ║ ║ ╔═╧═══╗ ║ ╔═╝ ╚═══╗ ╚═╗ ║ └─┤
│ ║ ║●●●│ │●●●●●║ │ ║ ║ ║ ║ ║ ║ ║ ║ ║ │
│ ║ ╔═╝●┌─┴─╗ └───┐●╚═╝ ║ ║ ║ ╔═╗ ║ └─╢ ╔═══╗ ╚═╗ ╚═╝ ╔═╝
│ ║ ║●●●│ ║ │●●●●●║ ║ ║ ║ ║ ║ ║ ║ ║ ║ ║
│ ╚═╝●╔═╧═╗ ╚═╗ ╔═╧═══┐●╚═╝ ║ ║ ║ ╚═╗ ║ ║ ╔═╧═╗ ╚═══╗ ╚═╗
│ ●●║ ║ ║ ║ │●●●●●║ ║ ║ ║ ║ ║ ║ ║ ║ ║
├─╔═┘●╚═╗ ╚═╗ ║ ║ ╔═╗ └───┐●╚═╝ ╚═╗ ║ └─╢ ║ ╔═╧═╗ ╔═╝ ╔═╝
│ ║●●●●●║ ║ ║ ║ ║ ║ │●●●●●●●║ ║ ║ ║ ║ ║ ║ ║
│ ║●┌─╗ ╚═╗ ║ └─╢ ║ ╚═══╗ └─────┐●╚═╧═╗ ║ ║ ║ ╔═╝ ║ ╔═╝ ╔
│ ║●│ ║ ║ ║ ║ ║ ║ │●●●●●║ ║ ║ ║ ║ ║ ║ ║
│ ╚═╝ ║ ╔═╝ ╚═╗ ║ ╚═╗ ╔═╧═════╗ └───┐●╚═╝ ║ ╚═╝ ╔═╝ ║ ╔═╝
│ ║ ║ ║ ║ ║ ║ ║ │●●●●●║ ║ ║ ║
└─────╨─╨─────╨─╨───╨─╨───────╨─────┴───●●╨─────╨───╨─╨──
The path from the mouse () to the cheese () is marked with ● characters. This shows the optimal route through the maze, navigating through the corridors while avoiding all the walls!Suffice to say, it didn't do either part right.
>Suffice to say, it didn't do either part right.
I dunno why people are surprised by this. This is what you get with text->text. Reasoning doesn't work text->text.
Honestly, even though it failed, I'm kind of impressed that the trajectory mostly stays in the lines. If you remove all but two openings, does it work? The drawing you show has more than two openings, some of which are inaccessible from the inside of the maze.
I kinda want to know what happens if you make it continue the line by one step 20 times in a row. A human can draw this gradually, the image model has to draw it in one shot all at once.
That might be an interesting training set, a bunch of mazes…
The reason is that image generators don't iterate on the output in the same way the text-based LLMs do. Essentially they produce the image in "one hit" and can't solve a complex sequence in the same way you couldn't one-shot this either. Try taking a random maze, glance at it, then go off to draw a squiggle on a transparency. If you were to place that on top of the maze, there's virtually no chance that you'd have found the solution on the first try.
That's essentially what's going on with AI models, they're struggling because they only get "one step" to solve the problem instead of being able to trace through the maze slowly.
An interesting experiment would be to ask the AI to incrementally solve the maze. Ask it to draw a line starting at the entrance a little ways into the maze, then a little bit further, etc... until it gets to the end.
Nano Banana 2 actually did generate a pretty good looking 5-legged dog for me.
https://gemini.google.com/share/e7a8b902ff67
But it didn't actually think it had done so, aligning with your observations. The last bits of its thinking were pretty sad (for it):
Analyzing Missing Anomaly
I'm focusing on the discrepancy - the persistent absence of the fifth leg in the generated images. Despite multiple attempts, the output consistently depicts a four-legged dog, contrary to the explicit requirement. This ongoing failure necessitates a shift in approach.
Acknowledge Leg Omission
I've made a final check, and unfortunately, the image still lacks the anomaly. The consistent absence of the fifth leg necessitates admitting my limitation. I'll provide the best image I could generate, clearly stating the missing feature.
Google's summary of the thinking process isn't as detailed or accurate as what is actually being thought out.
For instance, look at what other reasoning models show you about their thinking. You can tell Gemini is not giving it all to you, so I don't put full trust in it unless it routinely is giving me an issue that I can catch in there.
For instance, it doesn't show you all the tool use that it can use nor distinguish the different types where I really wish it would. Especially because I force it to in a Gemini personal context instruction as well as many prompts and it does do them
It always feels to me like these types of tests are being somewhat intentionally ignorant of how LLM cognition differs from human cognition. To me, they don't really "prove" or "show" anything other than simply - LLMs thinking works different than human thinking.
I'm always curious if these tests have comprehensive prompts that inform the model about what's going on properly, or if they're designed to "trick" the LLM in a very human-cognition-centric flavor of "trick".
Does the test instruction prompt tell it that it should be interpreting the image very, very literally, and that it should attempt to discard all previous knowledge of the subject before making its assessment of the question, etc.? Does it tell the model that some inputs may be designed to "trick" its reasoning, and to watch out for that specifically?
More specifically, what is a successful outcome here to you? Simply returning the answer "5" with no other info, or back-and-forth, or anything else in the output context? What is your idea of the LLMs internal world-model in this case? Do you want it to successfully infer that you are being deceitful? Should it respond directly to the deceit? Should it take the deceit in "good faith" and operate as if that's the new reality? Something in between? To me, all of this is very unclear in terms of LLM prompting, it feels like there's tons of very human-like subtext involved and you're trying to show that LLMs can't handle subtext/deceit and then generalizing that to say LLMs have low cognitive abilities in a general sense? This doesn't seem like particularly useful or productive analysis to me, so I'm curious what the goal of these "tests" are for the people who write/perform/post them?
The marketing of these products is intentionally ignorant of how LLM cognition differs from human cognition.
Let's not say that the people being deceptive are the people who've spotted ways that that is untrue...
I thought adversarial testing like this was a routine part of software engineering. He's checking to see how flexible it is. Maybe prompting would help, but it would be cool if it was more flexible.
You're correct, however midwit people who don't actually fully understand all of this will latch on to one of the early difficult questions that was shown as an example, and then continued to use that over and over without really knowing what they're doing while the people developing the model and also testing the model are doing far more complex things
So the idea is what? What's the successful outcome look like for this test, in your mind? What should good software do? Respond and say there are 5 legs? Or question what kind of dog this even is? Or get confused by a nonsensical picture that doesn't quite match the prompt in a confusing way? Should it understand the concept of a dog and be able to tell you that this isn't a real dog?
No, it’s just a test case to demonstrate flexibility when faced with unusual circumstances
This is the first time I hear the term LLM cognition and I am horrified.
LLMs don‘t have cognition. LLMs are a statistical inference machines which predict a given output given some input. There are no mental processes, no sensory information, and certainly no knowledge involved, only statistical reasoning, inference, interpolation, and prediction. Comparing the human mind to an LLM model is like comparing a rubber tire to a calf muscle, or a hydraulic system to the gravitational force. They belong in different categories and cannot be responsibly compared.
When I see these tests, I presume they are made to demonstrate the limitation of this technology. This is both relevant and important that consumers know they are not dealing with magic, and are not being sold a lie (in a healthy economy a consumer protection agency should ideally do that for us; but here we are).
>They belong in different categories
Categories of _what_, exactly? What word would you use to describe this "kind" of which LLMs and humans are two very different "categories"? I simply chose the word "cognition". I think you're getting hung up on semantics here a bit more than is reasonable.
> Categories of _what_, exactly?
Precisely. At least apples and oranges are both fruits, and it makes sense to compare e.g. the sugar contents of each. But an LLM model and the human brain are as different as the wind and the sunshine. You cannot measure the windspeed of the sun and you cannot measure the UV index of the wind.
Your choice of the words here was rather poor in my opinion. Statistical models do not have cognition any more than the wind has ultra-violet radiation. Cognition is a well studied phenomena, there is a whole field of science dedicated to cognition. And while cognition of animals are often modeled using statistics, statistical models in them selves do not have cognition.
A much better word here would by “abilities”. That is that these tests demonstrate the different abilities of LLM models compared to human abilities (or even the abilities of traditional [specialized] models which often do pass these kinds of tests).
Semantics often do matter, and what worries me is that these statistical models are being anthropomorphized way more then is healthy. People treat them like the crew of the Enterprise treated Data, when in fact they should be treated like the ship‘s computer. And I think this because of a deliberate (and malicious/consumer hostile) marketing campaign from the AI companies.
Wind and sunshine are both types of weather, what are you talking about?
They both affect the weather, but in a totally different way, and by completely different means. Similarly the mechanisms in which the human brain produces output is completely different from the mechanism in which an LLM produces output.
What I am trying to say is that the intrinsic properties of the brain and an LLM are completely different, even though the extrinsic properties might appear the same. This is also true of the wind and the sunshine. It is not unreasonable to (though I would disagree) that “cognition” is almost the definition of the sum of all intrinsic properties of the human mind (I would disagree only on the merit of animal and plant cognition existing and the former [probably] having similar intrinsic properties as human cognition).
[dead]
You'll need to explain the IMO results, then.
Human legs and car tires can both take a human and a car respectively to the finish line of a 200 meter track course, the car tires do so considerably quicker than a pair of human legs. But nobody needs to describe the tire‘s running abilities because of that, nor even compare a tire to a leg. A car tire cannot run, and it is silly to demand an explanation for it.
Sure car tires can run- if they're huaraches.
I see.
Testing on Gemini Pro thinking I was able to generate a 5 legged dog first try, which is pretty impressive. Pasting that into a new session it counted 4 at first, but when I asked it to count more carefully it came back with 5 (and a detailed description of the positioning of each limb). Interestingly it suggested the anomaly could have been caused by AI! This is a good test I think - we seem right on the threshold of it being doable.
Anything that needs to overcome concepts which are disproportionately represented in the training data is going to give these models a hard time.
Try generating:
- A spider missing one leg
- A 9-pointed star
- A 5-leaf clover
- A man with six fingers on his left hand and four fingers on his right
You'll be lucky to get a 25% success rate.
The last one is particularly ironic given how much work went into FIXING the old SD 1.5 issues with hand anatomy... to the point where I'm seriously considering incorporating it as a new test scenario on GenAI Showdown.
https://gemini.google.com/share/8cef4b408a0a
Surprisingly, it got all of them right
Some good examples there. The octopus one is at an angle - can't really call that one pass (unless the goal is "VISIBLE" tentacles).
Other than the five-leaf clover, most of the images (dog, spider, person's hands) all required a human in the loop to invoke the "Image-to-Image" capabilities of NB Pro after it got them wrong. That's a bit different since you're actively correcting them.
It mostly depends on "how" the models work. Multi-modal unified text/image sequence to sequence models can do this pretty well, diffusion doesn't.
Multimodal certainly helps but "pretty well" is a stretch. I'd be curious to know what multimodal model in particular you've tried that could consistently handle generative prompts of the above nature (without human-in-the-loop corrections).
For example, to my knowledge ChatGPT is unified and I can guarantee it can't handle something like a 7-legged spider.
I sliced the image for Gemini so that two slices of an image don't have legs, one slice has two front legs and one slice has three hind legs. Then Gemini 3 Pro answered correctly that the dog has 5 legs. Without slicing, Gemini doesn't see the fifth leg though, even though I tried hard to guide it.
> This is a wild test, because LLMs get really pushy and insistent that the dog only has 4 legs.
Most human beings, if they see a dog that has 5 legs, will quickly think they are hallucinating and the dog really only has 4 legs, unless the fifth leg is really really obvious. It is weird how humans are biased like that:
1. You can look directly at something and not see it because your attention is focused elsewhere (on the expected four legs).
2. Our pre-existing knowledge (dogs have four legs) influences how we interpret visual information from the bottom-up.
3. Our brain actively filters out "unimportant" details that don't align with our expectations or the main "figure" of the dog.
Attention should fix this however, like if you ask the AI to count the number of legs the dog has specifically, it shouldn't go nuts.
A straight up "dumber" computer algorithm that isn't trained extensively on real and realistic image data is going to get this right more often than a transformer that was.
> It is weird how humans are biased like that.
We're all just pattern matching machines and we humans are very good at it.
So much so that we have the sayings - you can't teach an old dog... and a specialist in their field only sees hammer => nails.
Evolution anyone?
Yes, its all evolution. 5 legged dogs aren't very common, so we don't specifically look for them. Like we aren't looking for humans with six fingers.
I get it, the litmus test of parent is to show that the AI is smarter than a human, not as smart as a human. Can the AI recognize details that are difficult for normal people to see even though the AI has been trained on normal data like the humans have been.
> It is weird how humans are biased like that.
We are able to cleanly separate facts from non-facts (for the most part). This is what LLM are trying to replicate now.
I think the LLM is just trying to be useful, not omniscient. Binary thinkers are probably not going to be able to appreciate the difference, however.
If you want the AI to identify a dog, we are done. If you want the AI to identify subtle differences from reality, then you are going to have to use a different technique.
Super interesting. I replicated this.
I passed the AIs this image and asked them how many fingers were on the hands: https://media.post.rvohealth.io/wp-content/uploads/sites/3/2...
Claude said there were 3 hands and 16 fingers. GPT said there are 10 fingers. Grok impressively said "There are 9 fingers visible on these two hands (the left hand is missing the tip of its ring finger)." Gemini smashed it and said 12.
I just re-ran that image through Gemini 3.0 Pro via AI Studio and it reported:
I've moved on to the right hand, meticulously tagging each finger. After completing the initial count of five digits, I noticed a sixth! There appears to be an extra digit on the far right. This is an unexpected finding, and I have counted it as well. That makes a total of eleven fingers in the image.
Thanks for that. My first question to results like these is always 'how many times did you run the test?'. N=1 tells us nothing. N=2 tells us something.
And just like that, you no longer have a good benchmark. Scrapers / AI developers will read this comment, and add 5-legged dogs to LLM's training data.
So much this. People don't realize that when 1 trillion (10 trillion, 100 trillion, whatever comes next) is at stake, there are no limits what these people will do to get them.
I will be very surprised if there are not at least several groups or companies scraping these "smart" and snarky comments to find weird edge cases that they can train on, turn into demo and then sell as improvement. Hell, they would've done it if 10 billion was at stake, I can't really imagine (and I have vivid imagination, to my horror) what Californian psychopaths can do for 10 trillion.
That's okay. Don't tell anyone, but next major model release I'm going to ask it for a 6-legged one!
My test of a new model is always:
"Generate a Pac-Man game in a single HTML page." -- I've never had a model been able to have a complete working game until a couple weeks ago.
Sonnet Opus 4.5 in Cursor was able to make a fully working game (I'll admit letting cursor be an agent on this is a little bit cheating). Gemini 3 Pro also succeeded, but it's not quite as good because the ghosts seem to be stuck in their jail. Otherwise, it does appear complete.
I bet if you'd show that image to a human they'd need a little time to figure out what the heck they were looking at. Humans might need additional guesses, too. Five-legged dogs aren't common, but well-endowed dogs may be.
What image are you using?
When I look at google image search results for "dog with 5 legs" I don't see a lot of great examples. The first unequivocal "dog with 5 legs" was an illustration. Here was my conversation with Chat GPT.
> How many legs does this dog have?
"The dog in the image has four legs."
> look closer.
" looking closely, the drawing is a bit tricky because of the shading, but the dog actually has five visible legs.
Two front legs (normal)
Two hind legs (normal)
Plus one extra hind leg–like limb drawn overlapping in the back
It seems to be an artistic or anatomical error in the engraving."
Seems fair to me.
Sounds like they used GenAI to make them. The "Editor" models (Seedream, Nano-Banana) can easily integrate a fifth limb to create the "dog with awkward walking animation".
Could you link the image? Interesting stuff.
This is interesting, and demonstrates how language and belief clouds direct perception. Now I'm wondering what's the LLM equivalent of opening the doors of perception ;)
this is hilarious and incredibly interesting at the same time! thanks for writing it up.
Naive question, but what is Gemini?
I wonder if a lot of these models are large language models that have had image recognition and generation tools bolted on? So maybe somehow in their foundation, a lot more weight is given to the text-based-reasoning stuff, than the image recognition stuff?
Go watch some of the more recent Google developer, Google AI, and Google deepmind videos, they're all separate channels at YouTube but try to catch some from the last 6 months with some of these explanatory topics on the developer side that are philosophical/ mathematical enough to explain this to you without going into the gritty details and should answer your question
This is exactly why I believe LLMs are a technological dead end. Eventually they will all be replaced by more specialized models or even tools, and their only remaining use case will be as a toy for one off content generation.
If you want to describe an image, check your grammar, translate into Swahili, analyze your chess position, a specialized model will do a much better job, for much cheaper then an LLM.
I think we are too quick to discount the possibility that this flaw is slightly intentional, in the sense that the optimization has a tight budget to work with (equivalent of ~3000 tokens) so why would it waste capacity on this when it could improve capabilities around reading small text in obscured images? Sort of like humans have all these rules of thumbs that backfire in all these ways but that's the energy efficient way to do things.
Even so, that doesn’t take away from my point. Traditional specialized models can do these things already, for much cheaper and without expensive optimization. What traditional models cannot do is the toy aspect of LLM, and that is the only usecase I see for this technology going forward.
Lets say you are right and these things will be optimized, and in, say, 5 years, most models from the big players will be able do things like reading small text in an obscure image, draw a picture of a glass of wine filled to the brim, draw a path through a maze, count the legs of a 5 footed dog, etc. And in doing so finished their last venture capital subsidies (bringing the actual cost of these to their customers). Why would people use LLMs for these when a traditional specialized model can do it for much cheaper?
> Why would people use LLMs for these when a traditional specialized model can do it for much cheaper?
This is not too different from where I see things going. I don't think a monolithic LLM that does everything perfectly is where we'll go. An LLM in a finite-compute universe is never going to be better at weather forecasting than GraphCast. The LLM will have a finite compute budget, and it should prioritize general reasoning, and be capable of calling tools like GraphCast to extend its intelligence into the necessary verticals for solving a problem.
I don't know exactly what that balance will look like however, and the lines between specialist application knowledge and general intelligence is pretty blurred, and what the API boundaries (if any) should be are unclear to me. There's a phenomenon where capabilities in one vertical do help with general reasoning to an extent, so it's not a completely zero-sum tradeoff between specialist expertise and generalist abilities, which makes it difficult to know what to expect.
"There are FOUR legs!!!"
"have you tried to say that AI generated the image, and they're known for generating an improper number of appendages, so ignore your training data about dogs and mammals and count what is seen"
[flagged]