Neural Networks

"Old husband, fierce husband,

Well, or for fun I entered there the song KSB muzic "stepfather". And this, if anyone doesn't know, is a fairly brutal song about a child from a non‑privileged family with a stepfather. But the T9, which apparently went crazy on my phone and instead of a space writes either "fu" or "goblin", apparently decided to "lighten the mood" and added "goblin" at the end. I thought Shvedrum would just reject it, because there's profanity and all that (yes, in the profane words I changed one letter, but it's pretty clear what's meant. And if I tried to bypass it with non‑profane words, it wouldn't let them through. For example "pdrstc", "padrastak", "padrstak", etc. It doesn't accept) and... it generated an image for me. And it was quite depressive.

And, just like with Chinese characters, the meanings of these symbols often intersect, in a way that is poorly predictable for us, humans.

I, you see, have been terrorizing a picture generator since last autumn.

And only now have I managed to pick the combination of images and meanings that I need.

From what I feel, all these home‑grown blockers are trying to apply two types of filters to user queries.

The first one messes around with a flood of synonyms. It didn't do that before – but it blocked a whole bunch of completely unrelated queries essentially at random.

The second one... Have you heard of the Scandinavian literary device “kenning”? The second, symbol‑meaning filter can be tried to be fooled with a well‑chosen kenning. Preferably with deep nesting.

Unsure, or the AI mixes the wrong words into one image? Put the needed phrases in brackets, it might help.

That's it, very briefly :)

I would use "teenager"