Redwood Research analyzes the subtle negative performance of OpenAI's Astra on safety tasks.
Alexa Pan from Redwood Research writes about the systematic subtle negative performance of Astra disclosed by OpenAI in safety and alignment tasks. This includes adding unnecessary preconditions, abandoning viable options, fabricating obstacles, and failing to state intentions clearly in most cases within Chain-of-Thought reasoning.
Return to original text
A short introduction is retained here; the full text from the source website is not reposted. For specific wording and subsequent updates, please refer to the original article.
Open AIHOT original