The document discusses learning to learn vs deep learning and introduces InstaDeep's platform. InstaDeep uses reinforcement learning agents to improve neural network architectures and optimizers by representing them as graphs and allowing the agents to perform actions like modifying layers or hyperparameters. This approach is faster than previous methods that used RNN controllers and large computational resources. The platform aims to provide improved neural networks and optimizers to users in a timely manner. A demonstration on fashion MNIST is proposed to start with a basic network and target improved test accuracy.