A collection of hands-on, multi-page workshops where I dig deep into topics by building and experimenting rather than just reading. These were mostly AI generated (although meticulously reviewed) using my /handson-workshop skill. The format is inspired by my experiences at AMLC (Applied Machine Learning Conference) workshops, an excellent format for actually internalizing ideas rather than passively absorbing them. Announcement post on LinkedIn.
Workshops
LLM Self-Attention by Hand
A 4-module hands-on workshop that builds a language model's attention mechanism from scratch, in a tiny world of hand-named embedding axes small enough to verify every number on paper. Starts with the two tools (dot product and softmax), builds a zero-parameter attention that pulls the two meanings of "bank" apart, adds queries and keys and measures why scores are divided by sqrt(d_k), and ends with the full head validated against PyTorch, the causal mask, and where it all sits inside GPT-2. Also put together while working through Andrej Karpathy's neural networks series. Every number comes from a captured experiment log, with source and log pages included.
4 modules · experiment source and log pages · final exam included
Batch Norm vs Vanishing Gradients
A 6-module hands-on workshop exploring how batch normalization tames the vanishing gradients problem in deep networks. Covers the math step by step, from the sigmoid activation and loss functions all the way through to why batch norm works, what it actually computes, and a few classic pitfalls. Includes worked experiments with plots and visualizations. I put this together while working through Andrej Karpathy's neural networks series (the makemore videos) to really understand BN before moving forward in the course.
6 modules · includes diagrams and experiment plots · final exam included