Part 7/10:
- Evaluation: A final script prompts Ada to evaluate how well each action meets the core objective functions.
He explains that the data is formatted in JSON, making it structured and easy to interpret. This data then serves as training material for simpler neural networks, or as a reference for humans to review and refine.
Balancing Cost and Quality
Throughout the process, Shapiro emphasizes the importance of managing costs by choosing the appropriate GPT-3 engines:
Ada is used primarily for generating evaluations due to its speed and efficiency.
Curie is employed for context and action generation, providing a balance between quality and cost.