Featured

Wordmaze is a single-turn word-ladder puzzle with a password constraint, built to be trivially gradable for RL.

A staged checklist for reducing agent and LLM costs, from prompt hygiene and model selection to tool pruning, trace analysis, and distillation.

Agents have a search problem across the whole stack: web search, RAG, tool discovery, skills/workflow loading, and even context compaction.

A NumPy walkthrough of GPT KV caching: what keys and values are cached, how attention uses them, and why generation gets faster.