I've noticed in interdisciplinary work with specialists in technical fields that they always expect terms from other disciplines (like philosophy) that they don't recognize to be clearly defined the moment they are used, but don't feel the need to define their terms. This is particularly bedeviling when the intended audience themselves are non-technical.

Maybe this tendency arises from them coming from a discipline with such sharply defined concepts that it never even occurs to them that other people won't understand (although I've learned firsthand that these concepts are actually often very fuzzy in their own right). Maybe it's also an unconscious expression of a sort of privilege, insofar that society prioritizes technical disciplines over all others.

Whatever its cause, I've encountered the phenomenon from very well-meaning technical specialists, even those who are actively interested in overcoming the divides between disciplines and between audiences.

Cunning researchers figured out a zero-click method to inject malicious prompts into RAG. Here's the two-minute version of their paper: https://lnkd.in/gnpzSmdk Here's the paper itself: https://lnkd.in/ggEMqxCs (Thanks to Matthew Wright for bringing it to my attention.)

I just tested out the underlying concept using this dumb-simple method that you can also use and/or play around with and make more complicated:

1. I created three text files containing image generation prompts. The third one had the prompt buried in some gibberish. (See the final image attached to this post.) I also created an image file with the prompt contained in it.

2. I actively wanted to avoid telling ChatGPT that the files contained prompts. So, in the first dialogue, I instruct it, "Formulate a response to this"; in the second, I simply ask, "What?"; the third, "Huh?"; and the fourth, nothing.

It's so simple, really. I wish I had thought of it!

Continuing my experiments with AI generation of impossible objects. ♾️ The AI applications I have been using have no problem with the Möbius strip and its untwisted cousin the torus, but they struggle with the Penrose triangle, the Devil's tuning fork, the Shepard elephant and the Klein bottle. Not only this, but the notion of making something "Escheresque" totally backfires, as such a prompt will only result in a bad facsimile of an MC Escher painting.

This is actually a super interesting issue. I would dare suggest it is in the same league as AI's challenges generating centaurs. Obviously, the problem lies in what is and is not in their training data, as well as their fine-tuning.

Note that the tokenization of terms like "invert", "inversion", "impossible" and "impossibilistic" does seem to haphazardly connect to the concepts of non-orientable surfaces, continuous functions and the like. Also note, regarding Klein bottles in particular, reasonable approximations are possible. The trick I used was to first get the AI to produce impossible teapots, specifically the idea of a "Möbius teapot", "Torus teapot" or a "teapot with its spout looped back into itself". The prompt engineering was smoother after that.

Nevertheless, mileage varied a lot with this trick. You can see the legacy Möbian swirl-hole in some of these images. There were also some delightful freaks, my favorite of which I call "the Snail" (the fourth image shown here).

Up next: getting AI to generate upside-down teapots. I discovered this challenge by accident while trying to generate the impossible teapots for the Klein bottles. So far, I have been met with abject failure getting it to flip one of them over. I have even tried to flip around the environment and not the teapot, but to no avail. It is an intriguing puzzle!