Current Status
This is an active learning project, building understanding through code and experimentation. The focus is on smaller models (GPT-2, Pythia, Llama-2-7B) that fit on a 12GB GPU, allowing for local development and hands-on exploration.
What's Been Built
- Interactive Neural Network Visualizer: Step-by-step visualization of XOR training with full mathematical detail
- Python Neural Network: From-scratch implementation for understanding fundamentals
- Custom Development Workflow: Claude Code commands and agents for TDD-driven learning
Explore what we've built →
Learning Method: Claude as Tutor
This project uses Claude not just as a coding assistant, but as a Socratic tutor - prioritizing understanding over answers.
Socratic Method in Practice
Rather than lectures and explanations, Claude:
- Asks questions first - before explaining a concept, ask what you already know or think
- Guides discovery - lead toward insights rather than stating them directly
- Checks understanding - ask you to explain back in your own words
- Pushes deeper - when you get something right, ask "why?" or "what would happen if...?"
- Encourages prediction - "what do you think will happen?" before running experiments
- Embraces confusion - confusion means you're at the edge of understanding; we explore it together
Example Interactions from CLAUDE.md
Instead of:
"An autoencoder has an encoder that compresses input to a latent space and a decoder that reconstructs it..."
Try:
"Before we dig into autoencoders - what's your mental model of how neural networks represent information internally? What do you think happens to the input as it passes through layers?"
Instead of:
"The residual stream is where information flows between layers..."
Try:
"You mentioned residuals - what do you already know about skip connections? What problem do you think they solve?"
Why This Matters
The goal isn't just working code - it's understanding deep enough to ask the next question yourself. This approach builds intuition and mental models through experiment-driven learning:
- Code first, theory after - run experiments, observe results, then explain
- Predict before running - always ask what you expect to see
- Break things intentionally - modify parameters to build intuition
- Small steps - one concept at a time, verified before moving on
Read the full tutoring guidelines →
Technical Approach
- Local development - All experiments run on local hardware (12GB GPU constraint)
- Practical focus - Hands-on code over theory alone
- Test-driven learning - TDD methodology for building reliable tools
- Learning in public - Documenting insights and patterns as we go
Resources & Inspiration