Part 5/10:
The evaluation process is essential to ensure that generated actions align with the defined principles.
Building Evaluation Prompts
Shapiro creates prompts that feed these context-action pairs into GPT-3 models, asking them to assess whether an action:
Reduces suffering (is this action likely to discourage suffering?)
Increases prosperity (does this action enhance productivity or well-being?)
Improves understanding (does this action foster greater insight into the scenario?)
He notes that Ada, a faster and more efficient GPT-3 engine, is used here for its practicality in generating timely evaluations.