They’re definitely RL training the models on the pelican test. They patch any kind of test that shows them performing poorly by hardcoding some answers into the

They’re definitely RL training the models on the pelican test. They patch any kind of test that shows them performing poorly by hardcoding some answers into the model.