I found myself pondering a fascinating question: would different AI Systems eventually „understand“ each other?
Let’s just imagine you create an image on your local machine using Stable Diffusion. Next, you take that image, feed it into ChatGPT+, and ask the OpenAI LLM if it could infer what the prompt for that picture would have been.
Would ChatGPT+ be able to infer the prompt on which an image was based but created with another model? Or at least get into the right direction?
It did! How do I know?
As always: because I tried, tested, and found out.
You can see the test-image I created below (Note: image not included in this text format).
My prompt for Stable Diffusion was simple: „God is an algorithm.“ I used the epicrealism_naturalSinRC1VAE.safetensors model with a generic negative prompt to get cohesive results, and no other add-ons.
I uploaded this to ChatGPT+ with the question:
What do you see? What can you read? What do you make of the image? What could have been its prompt?
Its answer (cutting out irrelevant parts and getting straight to the point):
The overall impression is that this could be a conceptual piece, perhaps a visual metaphor for the idea that „God is an algorithm,“ suggesting the omnipresence of algorithms in modern life and how they might be perceived as governing forces in our digital world.
BAM!
The only thing it got wrong was a comma instead of a period at the end of the sentence!
I don’t know about you, but I was blown away. (Note: screenshot of interaction not included in this text format).
So, to return to my initial question: do different AI Systems eventually „understand“ each other?
From my n=1 empirical result, I’d clearly vote for a „Yes!“
But what about you? Would you repeat this experiment and tell us what you found?
EDIT: A Banal Revelation
After some back and forth with my dear LinkedIn colleagues Daniel Reimann and Aamir Kadri, it seems that the „mystery“ around ChatGPT+’s prompt inference might be way more banal than first seen. The prompt was in the filename of the image. Duh…. 😂
A humbling, yet amusing, discovery that reminds us to check all the data sources!