Do DALL-Es dream of George Carlin jokes?

Some end of the year thoughts on generative AI, since this was the year of AI.

I have been messing around with ChatGPT and the image generation feature of DALL-E.

As I am taking off my technologist hat and putting on my amateur cognitive scientist one, I have some personal comments to share.

• AI is still not very good at free associations, abstract thinking and imagining things out of simple text. While the semantic connections and the cognitive network structure are quite impressive, it is still not up to a level of understanding irony or detecting sarcasm, in the same sense as humans do. For example, giving it the phrase "I like to please people, but I am not good at it" didn't produce anything near to the irony of the phrase. I mean, there are cases where it can show signs of detecting subtle cues, but I am with the opinion that right now it would excel more as a Sheldon from the Big Bang Theory rather than a George Carlin.

• AI is very good in describing images. If you upload a picture and you ask it to describe it, it's scary good. For example, I uploaded an image of me holding a cup and asked it to describe it. It did describe it very accurately and also described what my young daughter was doing far in the image background. Up until AI, detected my daughter, I wasn’t aware she was there, I mean I hadn’t noticed her even after I saw the image quite a few times. Humans suffer from inattentional blindness, something that presumably computers do not suffer from, which in turn makes them much more attentive to detail compared to humans.

• AI is very good in generating pictures out of exact directions. For example, telling it exactly what the person in the person should be wearing and doing. It's remarkably very close to what you have in mind. Which is a very good spark in discussing perception in the age of AI. For example, cognitively, you don't see something until you see it right? I mean, we don’t keep images of things in our brains, we construct them. So what happens if we describe something and AI constructs it first? Do we know if our brain constructed it first or if our brain constructed it after it saw AI doing it and then thought that AI constructed it exactly the way we imagined it? Inception, I know.

• AI is not very good in depicting text in images. Strange I understand, but it's very inconsistent. If you give it long sentences and ask it to print them along with an image, it will most probably mess up some letters. Challenge is (don’t quote me on this), AI is extracting the LETTERS as characters. As such, it gets "creative" with the letters and if AI thinks that they don't align with the image's style, it removes or changes them. Don’t know if there’s another explanation on this, but I would really like to find out.

Overall, I stand in awe that we live in a time and age where we compare an artificial agent’s ability to detect sarcasm in a sentence. Hope this gets better over time, or I would have a hard time explaining things to our future AI assistants (or overlords, whichever comes first).

I mean, I come from a generation where we read the sentence “The quick brown fox jumps over the lazy dog” time and time again until Windows Media Player would learn how to listen to voice commands. This is the future for us.

I am excited to see what's next. Wishes for a happy new year.