TokenTown: A visual way to understand how LLMs work

TokenTown provides an isometric, browser-based visualization of transformer architectures using a scaled-down model with 12-dimensional vectors and 2 attention heads. It executes real-time computations for tokenization, sinusoidal positional encoding, causal masked attention, and KV cache growth. To maintain readable output despite using random weights, the system blends hidden-state projections with a bigram prior.