TokenTown: A visual way to understand how LLMs work
An isometric city that runs a miniature transformer in real time: tokenizer docks, attention plaza, KV-cache warehouse, feed-forward mill and the sampler.
Summary
TokenTown presents an engaging isometric city model that visually represents the inner workings of large language models, featuring distinct areas like tokenizer docks and attention plazas. This creative approach makes complex AI concepts more accessible and easier to grasp.
Comments (15 of ) ⌛