OpenAI Podcast Episode 15 - Inside the Model Spec
4 snips
Mar 25, 2026 Jason Wolf, an OpenAI researcher on alignment who shapes model behavior and the Model Spec. He explains what the Model Spec is and how it guides model decisions. Topics include the spec’s structure and updates, the hierarchy for conflicting instructions, handling edge cases like Santa Claus, the role of chain-of-thought, and how the spec evolves with new capabilities and feedback.
AI Snips
Chapters
Transcript
Episode notes
Model Spec Is A Public Behavioral Interface
- The Model Spec is a public, human-readable document that explains OpenAI's intended model behaviors and trade-offs.
- It’s not an implementation or perfect claim; it captures key decisions and intentions for users, developers, and policymakers.
Use Concrete Examples To Clarify Policy
- Use examples and ideal answer templates to clarify ambiguous policies and model tone.
- Jason Wolf says examples help show principles in action and convey the model's intended style and personality.
Chain Of Command Resolves Conflicting Instructions
- The spec uses a chain of command to resolve instruction conflicts by ranking OpenAI, developer, and user instructions.
- Each policy gets an authority level so most defaults remain steerable below user instructions unless safety demands override.