LLM

A from-first-principles explanation of next-token prediction: logits, softmax, and temperature, demonstrated with a real bigram language model trained and run live in JavaScript, with an adjustable temperature slider and real probability bars computed on every step — plus what changes (and what doesn't) once you scale up to an actual transformer.

A hands-on introduction to the Model Context Protocol (MCP): what problem it solves, the JSON-RPC 2.0 messages exchanged between host and server, a live in-browser protocol simulator, and a real Python server built with the official SDK that exposes an H3 geospatial tool.

A reflective piece on how large language models represent place: the difference between spatial and platial knowledge, what LLMs do well (cultural associations, character, narrative), what they get wrong (boundaries, recency, under-represented places), and how to build applications that use this capability honestly.