Labs

Interactive, runnable notebooks that demonstrate the ideas behind the papers and models in the graph. Everything runs in your browser via WebAssembly — nothing is sent to a server, so there's no usage limit.

• Self-attention from scratch
  The mechanism behind every Transformer, in ~20 lines of NumPy — see exactly how query/key/value vectors turn into attention weights.
• Cosine similarity playground
  Embeddings work because similar meanings end up near each other in vector space. Edit the vectors and watch the similarity score move.
• Tokenization explorer
  Before a model sees any text, it gets chopped into tokens. This is a simplified byte-pair-style merge, showing the core idea behind real tokenizers like the one GPT/Claude use.