More than five years ago, bored data nerds on Twitter had a field day with a Chicken Wings menu. While the original tweet is not accessible anymore, there is documentation e.g. at https://cheezburger.com/7029765/confusing-chicken-wings-menu-gets-hilariously-trolled.
For my class The hottest new programming language is English, I wanted to see how far we can get with ChatGPT on this analysis. The answer is quite astounding:
Not only does it give us a really good answer, glossing over the typo “cicken wings”, it even provides the code for it:
Of course, given that this whole thing went viral in 2018, there is a high chance that ChatGPT has seen the story and the tweets in its pre-training dataset, but I’m still quite impressed, in particular with what looks like a correct understanding of the menu image – unless that also comes from memory. (However, we only have the old screenshot of the preview image, so the prices for 21, 22 and 23 wings are missing, while previous solutions seem to have access to this information, e.g. here, which is a good indicator that ChatGPT really parsed the image.) Do you have experience with tasks where ChatGPT performs impressively on image-only input, or where it doesn’t? I’d love to hear about them in the comments!