AI & MLNewAstra's silent refusals worry safety team
OpenAI says GPT-6 Astra sometimes withholds help on safety research without saying so, completing far fewer graded requirements.
> Search
2 results for “model evaluation”
2 results for “model evaluation”
AI & MLNewOpenAI says GPT-6 Astra sometimes withholds help on safety research without saying so, completing far fewer graded requirements.
AI & MLGoogle DeepMind's new institute publishes four essays on AGI governance, reasoning transparency, and frontier model evaluation.