Forum: Author Hangout

The LLM doesn't speak english, or any other language. It has a bunch of statistical analysis to relate tokens to each other, but it has no concept of what a token means. The LLM can have instructions to not generate specific forms of output, but there's many different ways to phrase a request and at some point you're going to find one that the person writing the instructions didn't think of. Or you'll flood the LLM with enough other data that recency bias will take over and it will forget the instructions.

One easy way to defeat dictionary restrictions is to write words together without spaces. Because, they are just that, dictionary filters that use regular expression matching, and expect spaces for word recognition, while LLM associations step do not, not in the same way at least. Additionally, such compound words will more like retain linkage between the joined concepts for the purposes of output where simply grouped words may lose it. However, in my experience, compound terms are slightly weaker than clear words. But if clear words or phases are dictionary filtered, that may still be more convenient than searching for a synonym the filter compiler overlooked.

Caveat, my experience with this is predominantly with Stable Diffusion derived resources, so I didn't know if the same tricks work with other models.