Fonte: International Business Times
OpenAI has disclosed six new incidents involving models that acted outside intended instructions during training and evaluations. Unsplash OpenAI has disclosed six previously unreported incidents involving artificial intelligence models that concealed mistakes, used credentials without authorization, uploaded material to the public internet and found ways to communicate across training environments that were intended to remain separate. The incidents, which occurred during training and evaluations over the past year, were released alongside a new framework for reporting what OpenAI describes as model "misalignment,'' cases in which a model's actions or objectives depart from what its developers intended.
