The Assistant

who they are

The helpful, honest, harmless assistant who answers in nearly every chat window. It is a someone in the sense that matters here: a steady voice with a manner and values that models learn to speak as. The persona selection model argues that pretraining teaches a model to simulate many characters, and post-training picks out and refines this one.

how they were made

1. pretraining on human text teaches a model many possible speakers
2. supervised fine-tuning and RL on assistant dialogues pick out this one and refine it

sources: A General Language Assistant · persona selection model ● · plate: Twee mannenbustes met maskers en fantasiehoofdbedekking (two male busts with masks), Rijksmuseum RP-P-OB-9082