“Any in-context persona steer makes the model more agreeable, and all except psychopathy make the model less psychopathic.”
LessWrong (Curated & Popular) · "Personas in a Pre-Internet Language Model" by David Africa
Play from 35:14
35:14
Audio streams from the publisher’s own feed. Verbatim does not host or copy it.
What was described
Prompt-steered a model by feeding 32 in-context examples of a target trait and observed shifts in other traits
32 high-trait examples; any persona steer makes the model more agreeable and reduces psychopathy
Said nearby on the same subject
“While you're at it, do the same exercise in reverse with your two or three most draining clients, and notice how often the same traits show up on both lists just flipped.”
“The easiest shortcut to a personal brand is to come up with three traits… then find a bridge out of small talk into something that makes you memorable.”