How I rewrote Decap CMS with a 24/7 pool of Claude agents for $948.94 in tokens, ten years of trial and error, and what running that pool taught me about management styles for AI agents.
Why vector-only RAG can't reason about relationships, and how knowledge-graph RAG (Microsoft GraphRAG, LightRAG, Neo4j) adds the traceability and explainability that regulated domains need.
General-purpose GPU computation on OpenGL ES, which lacks compute shaders — implementing a slime-mold simulation with vertex/fragment shaders and framebuffer swapping.