The article outlines the mechanism-level reproduction of Google's Nested Learning architecture, specifically the HOPE blocks, CMS, and Self-Modifying TITANs. It emphasizes achieving the quality benchmarks set by lucidrains' TITAN reference while ensuring the project remains fully open-source. Key features include high-level updates for HOPE, CMS, and TITANs, with comprehensive unit tests addressing tensor-level invariants such as teach-signal and causality. However, it notes limitations, including the absence of online backpropagation through updates and the lack of support for multi-GPU setups. Installation requires Python 3.12 and PyTorch 2.9.0 with CUDA support for accelerated performance. The article provides several command-line instructions for various operations, including tokenizer training, corpus filtering, and running smoke tests. For training, users can execute commands tailored for different devices, such as single GPUs, CPUs, or Apple Silicon (MPS). It includes specific configurations for distributed data parallelism (DDP) and DeepSpeed setups, ensuring flexibility based on hardware availability. Evaluation processes are thoroughly detailed, covering zero-shot tasks, continual learning, and long-context diagnostics. Users can run scripts for plotting forgetting curves and analyzing long-context performance. The article emphasizes the importance of logging and managing checkpoints, allowing for effective tracking of training progress. Memorization techniques are also discussed, with options to control adaptations based on ground-truth data. Overall, the article serves as a comprehensive guide for developers looking to implement and experiment with the HOPE architecture.