OpenAI launches 'embedded' safety evaluations by bringing in four external organizations including Apollo Research and METR
OpenAI launched embedded safety evaluations by hosting four external organizations in-house, providing deep access to models and source code to improve transparency.
Key points
- •Four external evaluation organizations, including METR and Apollo Research, are now embedded in the company to verify safety.
- •Deep access to training models and source code is provided to these external groups.
- •The initiative aims to improve safety management transparency by assessing intentional sabotage and improper model behaviors.