What Your Website Looks Like to an AI AgentSearch traffic is starting to arrive with a user agent that is not a person. Assistants, answer engines, and autonomous agents visit a page not to read it but to extract from it, and they are terrible1Aug 9 · 5 min readjoshgreen.hashnode.dev
Two Mirrors, No Masters: How My Homelab Survived ItselfI serve six public websites, including a search engine for a few hundred thousand 3D-printable models, from two mismatched mini-servers in two different corners of the same city. The design goal was b1Aug 5 · 6 min readjoshgreen.hashnode.dev
A 30B Model Crawled on My Mini PC, and the Weights Were InnocentThere is a comfortable assumption in the local LLM world: if a model only activates a few billion parameters per token, it will run fast on modest hardware. Sparse mixture-of-experts models are sold o21AAug 1 · 5 min readjoshgreen.hashnode.dev
Running a Coding Agent on a Local GLM 4.7 with vLLM: The Setup Nobody DocumentsGiving a coding agent its own brain for zero API cost is genuinely possible now. GLM 4.7 Flash serves on two consumer RTX 3090s, and a real agent framework can point at it instead of a cloud endpoint.11DJul 28 · 7 min readjoshgreen.hashnode.dev
Your Local MoE Model Is Slow: A Diagnostic ChecklistMixture-of-experts models are the reason a 30B can decode like a 3B on a desktop. They are also the reason people post "why is my local model crawling" threads, because MoE has a handful of failure mo1Jul 24 · 7 min readjoshgreen.hashnode.dev