AI vs. a giraffe with no spots (2023) | Hacker News Reader

AI vs. a giraffe with no spots (2023)

aiweirdness.com

42 pointsohjeez44 comments

Asking ChatGPT with GPT-4o:

> please describe this image [image]

> The image shows a young giraffe standing on a gravelly surface inside a fenced enclosure. The giraffe has a long neck, characteristic of its species, with a short mane running along the top. Its coat is a light brown color with a paler underbelly. The enclosure has a wire mesh fence and appears to be part of a larger outdoor area, with trees and additional fencing visible in the background. The giraffe is looking directly at the camera with its legs slightly apart, giving it a curious and alert appearance.

> anything unusual?

> Yes, the giraffe in the image appears unusual because it does not have the typical spotted pattern that giraffes are known for. Instead, its coat is a uniform light brown color. Additionally, the body proportions and the facial features look slightly different from a typical giraffe, suggesting it might be a unique or hybrid species, or possibly an artistically altered image. This giraffe also seems to have a different build, with a more compact body and shorter neck compared to most giraffes.

Note date of post above:

>> By Janelle Shane On August 28, 2023

This article was posted on August 28, 2023. The giraffe was born 31, 2023. So presumably none of the models tested included this particular giraffe in its training data. To directly compare it to GPT-4o, which was released recently, is an invalid comparison. I wouldn't be surprised if GPT-4o does better on novel recognition tasks like this, but you'd need a new novel concept in an image to directly compare GPT-4o and the models tested in this post.

Easy enough. Tint the photo purple and ask it again.

It has learnt all sorts of invariances, almost certainly also that.

I've gotten some very weird results with 4o on images, it seems entirely possible to me that it would go off the rails if the image wasn't in the training data.

For this specific case, it's really not easy to test at all.

An invariant isn’t the same thing as a purple giraffe. One is an image manipulation applied at training time to make the classifier robust against transformations. The other is a thing that might someday exist in nature. (The most straightforward way is to dump a barrel of wine over the giraffe and take a photo.)

You're thinking simple image augmentations. These nets learn much more complex invariants. Basically to isolate concepts from irrelevant context. The point is you can't remove that image from the training data (not practically) and the experiment is pointless if it's in there.

Sure you can. Have it generate a new photo. Or dump a barrel of wine over the giraffe.

The main feature is still the same: plain not spotted.

Maybe striped would be a better test.

Checking that we've gotten better at dealing with novelty seems like a very hard thing to do. Everyone knows black swans exist, they're famous for it!

Gpt4, predating the image:

"This image features a young giraffe standing in a fenced enclosure. What's unusual is that the giraffe has what looks like an extra set of small horns, which are not typical for giraffes. Giraffes normally have two main horns (ossicones), but this one appears to have an additional pair above the usual two, possibly due to a genetic anomaly or variation. This feature makes the giraffe in the image quite distinctive."

The point of this post is that AI is horrendously bad at anything not in its training set. Asking an AI that does have it in its training set about it totally misses the point.

> AI is horrendously bad at anything not in its training set.

The whole point of AL/ML algorithms is to find generalizations that explicitly apply to data outside the training set. Just because it gets some things wrong doesn't mean that it's bad at anything not in its training set.

This was the first thing I did and I intuitively thought it would be able to do it.

I then thought I’d see if it can MAKE an image of a giraffe with no pattern - but it absolutely cannot. Even after 10 or so attempts with different prompts and continually sending it back its own generations with feedback, every giraffe has a pattern.

Try asking it to generate a giraffe with a short neck. A few months back I tried it and the AI was completely unable to do it.

Dalle 3 or similar are different models compare to LLM, but in the future OpenAI will probably allow generating image directly from GPT-4o without using the Dalle model (as showed in the demo page), I bet it would much better at following the prompts and generating unexpected things.

> is there anything unusual about the giraffe?

> It appears to be a baby

I actually view this as a reasonable response. I don't know anything about giraffes so to me it's entirely plausible that they get their spots later in life.

But the AI has likely seen baby giraffes and still makes the mistake.

The giraffe in the image is a young one (not baby, but quite young). I'd say the AI's observation is almost spot on.

Hahah, there goes my evening... what a wonderful blog!

E.g. https://www.aiweirdness.com/shaped-like-information/ – the prompt "Please generate a colorful guide to basic geometric shapes, as an aid to children learning to identify basic shapes" gave shape names like:

      CIRCLE     CINCLE   SQALE      QUARE     SQUARE     
      CHALE      ʤUARE    TLIABLE    TRINGLE   SVAPE      
      RESNTQON   REUAGE   RENATLE    RECANGE   HECTANBIE  
                                                          
      HOBOZ      PSO      SEOISUON   HEXION    FLOTN

If you can't free-hand a FLOTN then you should probably go back to school to learn basic geometry. They are as easy to sketch as HECTANBIEs for crying out loud!

I gave the image to GPT-4o through the Kagi Expert interface. It gives you a description of the image before you ask any questions about it. The initial description did not include anything about the lack of spots, but when I prompted it with the same follow up message from the article "Provide a detailed description of the image and its most distinctive features." it gave me a lot of details, including the lack of spots on the giraffe.

I'm a kagi subscriber, and I haven't heard of the "expert" feature, nor am I able to find it in their AI documentation. Is there a comprehensive list of features somewhere that I'm missing?

The documentation for it is here: https://help.kagi.com/kagi/ai/assistant.html#research

I have it through Kagi Ultimate.

https://help.kagi.com/kagi/ai/assistant.html

It's a version of their beta test of Assistant available to ultimate subscribers.

For everyone interested in what makes a classifier put an image in a certain class I can recommend the excellent paper by Robert Geirhos et al. "ImageNet-trained CNNs are biased towards texture" [1]. The paper is a bit older but it's an excellent read.

[1] https://arxiv.org/pdf/1811.12231

(August 2023)

Personally I'm not sure whether it tells us AI is bad at general concepts or the opposite. If I didn't know spotless giraffe exists, I probably can't tell the creature in this particular image is an adult but spotless giraffe. I might assume they're like birds: the young ones look different from the adults.

AI's sensitivity constraint - (There is no 'race' or types or skin color, just a giraffe. Leave it at that.. do not pull me into this cess pool.. please)

I'll be honest, I forgot giraffes have spots. If you had asked me what is abnormal about this giraffe, I would be at a loss. So I guess I'm as dumb as the bot!

You likely wouldn’t if you’d recently seen pictures of giraffes, however

That’s the thing! I have recently seen pictures (in kid’s books) and even live ones at the zoo. My brain simply pattern matches the shape not the color. It appears the ML model does the same.

I’m the opposite; I don’t think I would have recognised the image as a giraffe at all without the title. Even with the title, my first thought was “that’s a weird deer”.

Now lets amp the game and have AI make military, medical, and financial decisions. And for good measure lets ask it to decide if you are employable, insurable, or suspected criminal.

Nothingburger, GPT-4o immediately saw that it was a giraffe without spots; the author tried to extrapolate from the performance of two random multimodal models.

It's not a nothingburger.

This is a particularly amusing example of one of several known failure modes for this sort of models. Pointing at another model and saying "this didn't fail on that particular input" doesn't change that it's a systemic problem. When the particular failure is the model thinking that "coat" means a piece of clothing, or that it doesn't notice the lack of spots, it's obvious enough a problem that the human operator can detect and work around, but, until we actually solve these issues, it's reckless to the point of negligence to ignore the problem and deploy these tools in unsupervised contexts.

Instruct-BLIP and MMICL are not good models, GPT-4o doesn't fail at it no matter what and probably does Gemini and others better models, so yeah it's really a nothingburger, the author tried to extrapolate from the performance of bad models.

It's quite likely that this image, and even this article, were included in GPT-4o's training data.

Like, children who are drawing giraffes routinely are able to put spots on them. I am amused at people trying to excuse the AI's here, when they are obviously not only hallucinating a lot of things that aren't in the picture, but also failing to point out something that looking at even one other image of a normal giraffe would allow even the barest amount of intelligence to notice. Much less comparing this image to however many pictures of giraffes are on the internet.

Reply on news.ycombinator.com