$100000 AI competition: the task of finding "worse" big models

This competition aims to find the inverse scaling task , Especially those tasks that are important for the safe and responsible use of the language model . We hope that the submitted tasks will let us know more about which types of tasks show inverse scaling ; The inverse scaling task will also highlight the potential problems of the current language model pre training and scaling paradigm . Inverse scaling tasks are important , Because they represent the mismatch between the behavior we want the language model to show and the behavior we get from the training objectives and the data we use in practice . As the language model grows larger and is used in more real-world applications , The important thing is that they don't get worse or hurt users in undetected ways .

 

sponsor ：Ethan Perez,AnthropicAI researcher