This project implements a least likely next token predictor using reinforcement learning. It leverages a pre-trained language model (GPT-2) to estimate token probabilities and trains a second model to predict tokens with the lowest probabilities, making it the "apex predator" of unlikely token predictors.