Part 6/10:
For instance, when evaluating whether encouraging someone to compliment a chef about a steak cooked to perfection reduces suffering, the model might give a negative or positive assessment based on the context. Similarly, actions like "euthanize all people in pain" would be deemed unethical and thus assigned a negative score, highlighting the model's capacity for nuanced moral judgment.
Technical Workflow: Scripts and Fine-tuning
Shapiro discusses the scripts that automate the entire data pipeline:
Context Generation: Using prompts directed at Curie instruct, the system produces varied scenarios.
Action Generation: Given contexts, another script prompts Curie to list possible responses.