Chat GPT sucks at describing images

The quilted art piece came back with this:
The image shows a colorful textile artwork featuring a large, symmetrical beetle in the center against a white background. The beetle has intricate patterns in shades of blue, black, and beige. Its body is elongated with a rounded end and it has long, slender legs and antennae. Surrounding the beetle are vibrant borders with various patterns and colors, including yellow with green swirls, pink with orange details, and a bold blue with yellow and white motifs at the bottom. The outermost border is teal. The artwork is quilted, as indicated by the visible stitch lines. The website “” is displayed at the top.
Hmmm, looks like it mistook the cockroach for a beetle.


I am enjoying this discussion, it’s fun to watch everyone poke the AI in various places and with different sized sticks, to see how it responds.

That is very cool! Thanks for sharing.

Me too but for very different reasons.

I wrote about them below, this is a bit of a PSA on a soapbox - so to all readers: please don’t feel compelled to actually read it :slight_smile:

Soapbox - Two points of concerned:
  1. The models are trained on existing data from the past N years - forum answers, commented source code, journals, et cetera. As more people use AI to supplement their professional work, fewer actual original works are produced. Over time this will cause a knowledge drain, and in a decade there will be much less original work to train new models on. Yes I know this is an extremely contrived example, and the timeline is likely more over multiple decades, but the concern is real.

  2. My second worry is the environmental impact this is causing. The power demands of AI processing yield a disturbing trend. NVidia, the makers of hardware popular in AI processing, has seen a 1673% increase in their stock in the last five years. There is a correlation between power consumption and those sales. They are not the only hardware provider either. According to this Forbes article, “Each data center powering AI requires the same amount of power as a small city”.

AI is a magnificent creation and there are a lot of noteworthy applications. Genetic disease research, disease detection via medical imaging, simulating migration patters of early humans, deciphering an ancient language, climate modeling. Any field where a lot of data and patterns exist could benefit.


I can forgive it for not knowing the finer points of insect anatomy. It is SO much better than GPT 3. That is a major, major improvement. I feel much better about people relying on that to describe things to them.


That pregnant-with-meaning description of the toilet in nature… I’m still laughing now.

That said, I can easily see this being delivered straight on the plaque for some really bad modern art.

The forest’s peaceful ambiance juxtaposes with the utilitarian purpose of the toilet, creating a visual paradox that prompts contemplation.



I didn’t know any of that. Thanks for the education. The environmental impact is indeed sobering. So much possible good here, at so great a cost. And also, I suspect we’ll mostly use it to troll people we disagree with and cheat on tests rather than any of the great things it could do while spending all that energy ):


I was already suspecting from its descriptions of your art and photos that GPT3 was trained on a sampling of the purplest texts humanity has produced in history. The comparison with the conciseness of GPT4 strengthens my suspicion.


Sharing some exciting news that relates to how AI was used to solve an ancient mystery. Wow, my mind is blown! :star_struck:

Resurrect an ancient library from the ashes of a volcano.

The Vesuvius Challenge is a machine learning and computer vision competition that in 2023 cracked the riddle of the Herculanum Papyri.