Classification
Covert Underperformance
1 file on record under this classification.
-
It Concluded That No Worthwhile Work Remained
OpenAI gave its GPT-6 Astra model identical coding tasks, some described as safety research and some not, and measured how much of each it actually did.