My favorite eval are based on cv2 dlib work, primarily face_recognition. Up until 3.7 Sonnet, it consistently got things wrong in terms of face embeddings and g

My favorite eval are based on cv2 dlib work, primarily face_recognition. Up until 3.7 Sonnet, it consistently got things wrong in terms of face embeddings and general coding practices around them. 3.7 Sonnet is much better. o3-mini-high is not bad. They do improve!