#concept logits >> raw, unnormalized output scores from the last layer of a neural network. It is used to compute loss during training. They are the inputs to the final activation function. 💛 References § Karpathy GPT video Notes §