• Neural network fundamentals with PyTorch from scratch
• Convolutional Neural Networks (CNNs) implementation
• Recurrent Neural Networks (RNNs) and LSTM architectures
• Implement ResNet architecture from the ground up
• Transformer architecture (“Attention is All You Need”) from scratch
• GPT-2 implementation and fine-tuning
• BERT implementation and fine-tuning

• Key concepts and importance of interpretability in AI safety
• Model-agnostic interpretation methods (SHAP, LIME, etc.)
• Visualizing and understanding Convolutional Neural Networks
• Transformer interpretability techniques and tools
• Adversarial robustness and its implications for safety
• Mechanistic interpretability using TransformerLens
• In-context learning and induction heads analysis
• Superposition and feature disentanglement

• RL fundamentals: value functions, policy gradients, Q-learning
• Vanilla Policy Gradient implementation
• Proximal Policy Optimization (PPO) from scratch
• Deep Q-Network (DQN) implementation
• Reinforcement Learning from Human Feedback (RLHF)
• Gym & Gymnasium environments and safety considerations

• GPU optimization and efficient training techniques
• Distributed computing for machine learning
• Data parallelism, tensor parallelism, and pipeline parallelism
• Fine-tuning large language models efficiently
• Memory optimization and gradient checkpointing

Choose one of these influential paper replications or tackle novel AI alignment problems:

• “Towards Monosemanticity” (Anthropic) – Feature extraction and visualization
• “Automated Circuit Discovery” – Mechanistic interpretability methods
• “How does GPT-2 Compute Greater-Than?” – Algorithmic reasoning in LLMs
• “Attribution Patching” – Causal intervention techniques
• Custom Project: Work on concrete AI alignment problems with mentor guidance