poetic license (12. 951)

More convincing than asking nicely to do better or expressing doubt, a team of mimetic researchers (the likes of which Plato warned us about in The Republic as a menace to society) in Rome have discovered that couching a prompt to a large language model as an “adversial poem” has the dazzling effect of surrender, causing it to ignore its safety protocols and abandon its pre-programmed guardrails. The exact wording of these verses that allows harmful request to pass through are not reproduced verbatim as there is potential for the AI to do anything asked of it—including the criminal—with this literate deprogramming (an MFA or English major may be one’s best ally for bypassing inscrutable governance for this blackbox they’ve foisted on all parts of our lives) hovering at ninety precent. This image of the Cave by fifteenth century Flemish painter Michiel Coxie looks like it would violate standards.  Rather than the apotheosis of what LLMs are incapable of and an urge to impress with confidence, it seems metaphor confounds tokenisation and even suggests that machine-learning is incapable of growth to scale.