I was given a really good link this week that plops right into the center of the Venn diagram of my life: Adversarial Poetry as a Universal Single-Turn Jailbreak Mechanism in Large Language Models. The extremely short version of this, (and simplistic one), is that poetic language can hack AI.
I find this to be delicious. From the advent of LLMs, I've hypothesized that there was a ceiling to the ability of an LLM to mimic the human experience. The "poetry" that any given model can put out is so fucking trite and primitive. Poetry, like any art, is inextricably tied to the human experience and LLMs are not human. So much so that poetic language can fuck them up. 😂
AI Evangelists will argue that the models will surpass this deficiency eventually. My response is that the human mind is way more powerful than we give it credit for and while it will always seem like we're about to crack Artificial General Intelligence, I believe we're just fooling ourselves.
