Universal predictive learning is a powerful method that utilizes meta-learning to quickly learn new tasks from limited data. By exposing itself to a wide variety of tasks, it can acquire universal representations, enabling generalized problem-solving. This product explores the potential of scaling the most powerful universal predictor - Solomonoff Induction (SI) - through meta-learning. We leverage Universal Turing Machines (UTM) to generate training data, allowing the network to encounter diverse patterns. We provide theoretical analysis of the UTM data generation process and the meta-training protocol. We conduct comprehensive experiments on neural architectures (such as LSTM, Transformer) using algorithms with varying complexities and generalizability for data generation. Our results demonstrate that UTM data is a valuable resource for meta-learning, capable of training neural networks that can learn universal prediction strategies.