is there any reasonable benefit of thinking of logit_bias as modifying the truth (which you don’t usually want) and penalties just choosing less “important” truths (at least according to the training set) but that for you are more important?
in the 2nd case, you’re not changing the reality, but choosing a less “interesting” part of it, whereas in the 1st case it’s like changing the laws of physics which might lead to more fundamentally untrue assumptions in the response?
I’m not writing this because I strongly believe it, but because I want to challenge this thought, and hopefully build a better intuition about them
why
I’m working on an exam generation tool and with gpt-4o-mini after 20 questions the llm starts repeating questions even though the materials have plenty of context to draw on. I’m always including previous questions and incrementally generate new ones 1 by 1, using the rudimentary approach “avoid these questions” in the prompt
more why
I always had this need to understand stuff based on other stuff I already intuitively understand, I don’t like the “because that’s how it works” thinking, because it forces me to memorize stuff (and I’m both lazy and not a huge memory olympic) which I think can only be used by the conscious while intuition also works at the subconscious level, which is a more powerful beast you can exploit