Video lesson 3

Building makemore Part 2: MLP

Break through the bigram's wall. Build the neural language model from Bengio et al. 2003 - a 3-character context window, an embedding table, a tanh hidden layer - plus the training craft that makes deep learning work: minibatches, train/dev/test splits, and learning-rate tuning. One guided project whose finale is a dev loss that finally beats 2.454.