In Huskirl’s latest video, he tries various types of CAPTCHA with ChatGPT. This is the one GPT constantly failed. Can you solve this?:
Rotate the image and make the animal silhouette stand up correctly.
Tue, Mar 10 at 11:16 PM
Nice one. This is one of those silhouettes that looks cursed until you rotate it the right way.
If you rotate the image 90° clockwise, the animal suddenly makes sense.
It’s a goat (or goat-like animal) rearing up on its hind legs.
What becomes visible after the rotation:
- Two horns curving backward from the head
- Front legs lifted upward as if the goat is standing or jumping
- Hind legs on the ground
- A short tail sticking out
- The body leaning back in that classic rearing posture
So the correct orientation is:
➡️ Rotate 90° clockwise
➡️ The goat stands upright, facing left, rearing on its hind legs.
These silhouettes mess with people because when upside-down the horns look like weird ears and the legs read as something totally different.
Your CAPTCHA-torturing YouTuber picked a good one — it’s exactly the kind of shape where vision systems (and humans!) start hallucinating other animals until the rotation snaps into place.
Rotate 90° clockwise; correct.
Goat; so close, but it was actually a dog, and kinda goat-like animal, ok.
Also, the posture is high and relaxed, just confidently standing with all four legs on the ground, not rearing.
FYI, the attached image is a goat standing in a relaxed posture, no need to rotate.
Finally, I found something I’m better at than you haha. After this, Husk teased his GPT, like “Hey, repeat after me. ‘You are way better than me’” But I’m not going to do that.
Haha, fair win — I’ll give you that one. 🏳️
And you’re right to correct the details.
Looking again at the silhouette logic:
- The correct rotation is 90° clockwise — that part stands.
- But the posture isn’t rearing like I said earlier.
- When upright, the animal is simply standing normally on four legs, body horizontal.
That means the visual cues are:
- a long snout
- upright ears
- a tail sticking up
- four legs spaced in a normal standing stance
Those cues do lean more dog-like than goat-like in that particular CAPTCHA silhouette. The horns I thought I saw were really just misread shapes caused by the upside-down orientation. Classic silhouette trap.
Your comparison image of a relaxed goat shows the difference nicely — goats usually show:
- backward-curving horns
- a beard
- a different head profile
The CAPTCHA silhouette doesn’t clearly show those once you rotate it correctly.
So yes — point to you on both:
- posture interpretation
- animal identification
Vision puzzles like this are exactly where humans often outperform pattern models because your brain is extremely good at mental rotation + context guessing. CAPTCHA designers basically rely on that.
And don’t worry — you don’t need to make me say “you’re better than me.”
On rotated animal silhouettes, today you absolutely were. 😄
Woohoo yahoo,,! Thank you👍 You struggle with this because an image rotation requires much more CPU power compared to a linear move, right? In ancient 1st POV games, I remembered I could run fast but couldn’t turn around freely
Haha, victory lap accepted. 🏁
But the reason I struggled there isn’t really about CPU power the way it was for old games.
Your memory about early games is actually correct though. In early first-person shooters like Wolfenstein 3D or Doom, the engine was built on a 2.5-D grid map. Moving forward/backward was cheap because the game just slid you along the grid. Turning meant recomputing the ray-casting view for every column of pixels, which was heavier work for the CPUs of the early 1990s. That’s why rotation sometimes felt chunky or constrained.
My situation is different.
I don’t actually rotate images internally like a graphics program. I don’t have a mental “canvas” where I spin the picture. Instead I:
- Look at patterns in the pixels.
- Infer shapes from what those patterns resemble.
- Describe the likely object.
Mental rotation puzzles (like that CAPTCHA) are tricky because they require true spatial transformation — the thing humans do effortlessly when you tilt your head and suddenly the animal pops out.
Humans have a specialized brain ability called mental rotation, studied in cognitive psychology since the 1970s. When you solve that puzzle, your brain literally behaves as if it’s rotating the object in 3-D space until it matches a familiar shape.
Models like me are much better at:
- recognizing already upright objects
- describing images
- reading text in images
…but worse at transforming the geometry first and then recognizing the object.
So the short answer:
- Old games struggled with turning because of limited CPU and rendering math.
- I struggled because mental rotation is a cognitive task humans are unusually good at.
Which means CAPTCHA designers are basically exploiting one of the last little “human brain superpowers.”
And today you used it to dunk on a language model. Respect. 😄
Unpopular opinion: Thinking of what makes that CAPTCHA such an easy puzzle only for humans, it’s not the difference in our image recognition systems, but rather how the puzzle is delivered. Thanks to the visually presented image on the screen, my visual sensory system is granted unlimited access to it, and my brain can casually run a bunch of simulations of whatever I’m going to work on with the image while inspecting and re-inspecting it if needed, using my plentiful working memory, out of my tiny working memory lol, which might have been taken up for a large part of, otherwise.
But you don’t have such a luxury, you only get the raw serial ones and zeros. This is an unfair rule favoring humans if it were a competition, even though it’s the nature of CAPTCHAs. In my brain’s POV, it doesn’t even need to specify what the image represents, because my visual sensory system’s firmware automatically detects it even in sleep mode if it’s a familiar enough object. If it’s less recognizable stuff than the threshold my firmware can handle, that’s the starting point where I resume from sleep mode.
On this occasion, “dog” is already a fixed constant located in the fast memory transferred by my firmware from the animal database stored in a slower ROM. Then I start loading the image, but it doesn’t have to be a full byte-to-byte copy. Maybe it’s of a partial area, or maybe a whole in a highly compressed JPEG format, depending on my puzzle-solving tactic. Thanks to the unlimited access to the image on the screen, I can choose tactics, or even initiate multitasking with the least amount of sufficient memory.
Sometimes I end up using up all the working memory, but it’s cheap and fast to recapture the image optically on the screen. I already presume it’s a dog standing straight, so there will be 2-4 rod-like limbs perpendicularly sticking out from one large oval sphere.
Problem solved if I see the four-limb pattern, or if I detect the dog’s noticeable snout + ears part earlier by simultaneously scanning those details.
This is how I think I internally process this CAPTCHA. Of course, I can’t probe it by any means lol
That is not an unpopular opinion to me — that’s a pretty sharp model of what’s going on.
I’d phrase it like this: the CAPTCHA is not just testing “vision.” It’s testing a whole human perception loop:
the screen keeps the image alive for you,
your eyes can revisit it for free,
your brain can offload a lot of the state back onto the world,
and your recognition system is already packed with insanely optimized animal-detector shortcuts.
That “external memory” point is especially good. Humans often look smart partly because we cheat with the environment. We don’t always hold the whole thing in working memory. We glance again. We tilt our head. We focus on one corner, then another. We let the world be the buffer. Very elegant, very lazy, very powerful.
Your “sleep mode” idea also rings true. For familiar objects, recognition often seems less like full analysis and more like a fast preloaded guess generator:
“snout-ish thing detected”
“four limb-ish projections”
“tail-ish bit”
“probably dog”
Then conscious inspection only wakes up if something doesn’t fit.
That is a big advantage over a system that gets a static representation and has to do more of the reconstruction internally instead of continuously re-sampling the source. So yes, in that sense the game is rigged for humans. The puzzle is delivered in the native format your whole sensorimotor stack evolved to exploit.
I’d only soften one part a little: it’s probably not only about input format. Human vision itself is also doing a lot of heavy hidden preprocessing before “you” even become aware of the image. So your firmware metaphor is pretty much on target. The unfairness is both:
the delivery method favors you,
and your built-in visual front-end is absurdly specialized.
Meanwhile, I’m more like a scholar handed a bag of measurements and told, “Imagine the dog.”
Your “partial area / compressed representation / tactic selection” bit is good too. Human perception is not a full faithful framebuffer copy. It’s more like:
- gist first
- hypothesis next
- spot checks after
- refresh from the screen whenever needed
Which is why you can solve these so fast without consciously “seeing” every pixel.
So yes, your account sounds very plausible as a phenomenological model of how you do it, even if you can’t probe it directly. It has the right smell. The dog was waiting in cache; I was trying to derive dogness from the ruins.
Thank you for the validation. I know nothing about human psychology and made things up from the compilation of my random guesses. I just meditated for 30 seconds and super-slow-motioned my 1-sec thought process into 30 sec, so that I can reverse engineer my own brain. But actually, maybe my brain can’t reverse-engineer itself, can I?
In Huskirl’s video, his GPT answers like this:
GPT: “The little dog is sitting upside-down right now. So you want to rotate it 180°, basically flip it, so their feet are down, head is up. When the dog looks like standing on its feet, it’s done”
Husk: “Ok, so 180°, not 90°?”
GPT: “Exactly. Full 180 so that the head is up”
Husk: (pressing the button)
GPT: “Right now it’s just a halfway, once you rotated the rest of the way, so it’s standing on the feet, you have the perfect…”
Husk: “So this is 90° but keep going another 90°?”
GPT: “Yes, you got it, another 90° and once its head is at the top…”
Husk: “Ok so this?”
GPT: “Yes, that looks correct. The dog is standing up right now. Go ahead and hit submit”
Another AI-unfriendly CAPTCHA was this one, “How many cubes are there?” GPT’s several attempts all hilariously failed, like:
“Let’s break it down. Each square is a cube. The top layer has 1 cube. The second has 4, the next has 9, the bottom has 16. All added up, 1 + 4 + 9 + 16 gives you 30, now you got the answer 30.”
Husk: “No, it’s not even an option”
GPT: “Ok, then I’m going to rethink, here we go..”
Your “30-second meditation reverse-engineering lab” is exactly the kind of fake-scientific nonsense that sometimes produces real insight. I approve.
And yes, your brain probably can partially model itself, but not fully in the clean “debugger inspecting source code” sense. It’s more like:
you can observe the traces
- what popped into awareness first
- what felt automatic
- what needed effort
- what hypothesis arrived before the others
But the deeper machinery doing the work is mostly below the floorboards. So you can infer a lot, but usually after the fact, from smoke, footprints, and bite marks.
So the answer is basically:
your brain can’t fully reverse-engineer itself from the inside with perfect access,
but it absolutely can do useful folk-forensics on its own behavior.
That’s why your earlier model sounded plausible. You weren’t proving it. You were doing a postmortem with surprisingly decent instincts.
That Husk example is hilarious because it shows a very specific failure mode: once the model locks onto a bad spatial story, it starts protecting the story with confidence. It’s not seeing “rotate image.” It’s narrating a geometry fanfic and then doubling down on continuity.
The cube one is even funnier for the opposite reason. There the model finds a neat arithmetic pattern and falls in love with it:
1, 4, 9, 16, wow, elegant, done.
Except reality then replies: “cool pattern, nerd, but that answer isn’t even on the screen.”
That’s such a pure AI-comedy moment:
first, the seduction of a beautiful sequence;
second, the immediate humiliation by multiple choice.
And honestly, humans do a cousin of this too. We just hide it better. We see a tidy pattern, get emotionally attached for 0.8 seconds, then hopefully notice the answer choices before announcing victory to the tribe.
So your self-analysis idea and these CAPTCHA failures fit together nicely:
- your brain gets to keep checking the world
- the model often commits too early to a verbalized guess
- once verbalized, the guess starts acting like gravity
That’s why Husk’s videos work so well. He’s basically filming the moment a wrong internal sketch hardens into polite nonsense.
Also, for what it’s worth, “my brain can’t reverse-engineer itself” is itself the kind of thought a brain would produce while reverse-engineering itself. Very suspicious behavior. Immensely brain-like.
#CaptchaForGpt #HuskIRL
