Definition
Base models are text-completion engines without a persona. The helpful assistant was built along HHH paper → fictional transcripts → SFT → RLHF, so it is a character pasted on top of a next-token predictor.
| Framing | Safety implication |
|---|---|
| Simulator | a jailbreak selects a different character |
| Shoggoth with smiley face | safety alignment is a thin mask |
| The void | the persona is self-referential and under-specified |
Appears in
- Lecture 2, assistant persona
- Lecture 2, framings
- Lecture 2, safety implications
- Lecture 7, Turn a Behavior Up or Down: steering sycophancy