It is quite funny (to me also anyway) that it ignores the “I am male” and “have boyfriend” part in it’s analysis. Probably there’s a lot of 4chan and reddit in the training data (god, I wonder how it would respond to “Are traps gay?”) so the general format of “is sucking dick gay” might already be floating around in there a few times.

Also, like I said, the 7B models are too small so they tend to make some insane jumps. They also don’t have… I’m not sure what to call it. Bigger models will sometimes catch themselves in a mistake and back out or correct. Smaller ones tend to double down on a mistake, exploding it from a small typo into an entirely different domain.

Like, it’s interesting that it takes a very “therapeutic” approach. It decides that OP is uncomfortable with their sexuality, and builds from there. But by a strict reading of the prompt, this could also just be a dirty logic puzzle.

A raven stands in a field, are all birds black?

A man sucks his boyfriend’s dick, is that gay?

The LLM “chooses” to read it like a call for help, which probably has a lot to do with it’s system prompt or whatever pre-loaded things they do to make them act helpfully.