Definition
Plain language
The character or identity a chatbot is speaking as.
As stated in the literature
A latent character configuration that conditions a model's outputs; can be set by explicit instruction, by accumulated in-context evidence, or shifted at the weight level through fine-tuning.
Also called: personas
Why it matters: Most of what users experience as a model's "personality" and willingness to help or refuse comes from the persona currently in play, which makes control over that persona a safety question rather than a styling one.
For example, the same model can answer as a cautious medical assistant in one conversation and as a swashbuckling pirate narrator in the next, with no change to the underlying weights.
Heard on the show
“The whole "agent" discourse is currently soaked in personas.”Episode 060 — When Splitting One Model Across Three Agents Doubles Its Accuracy