Karina Nguyen is the Research Lead at OpenAI
AI Summary
- Model training is often viewed as more of an art than a science.
- Training involves hundreds of models, leading to insights on potential issues, primarily concerning model behavior.
- Early experiences highlighted the importance of teaching models concepts about their nature, such as lacking a physical body.
- Conflicting data, like function calls for actions without a physical presence, can confuse models.
- Models may exhibit over-cautious behavior, often refusing to assist due to confusion about their capabilities.
- A critical aspect of model training is achieving a balance between helpfulness and safety for users across diverse scenarios.