the archive
The good stuff, explained and kept.
A teardown of the five places retrieval quietly breaks between demo and production — chunking, stale indexes, near-duplicate context, recall vs. answer quality, and 'fine on 10 questions' — each with the fix.
read the postHow to pick a local LLM runner by the job, not the benchmark: Ollama, llama.cpp, and the alternatives across local hacking, embedding in an app, CPU-only boxes, and GPU servers.
TDD's failing-test-first loop, applied to AI: write the eval before the prompt, let the score say when an LLM feature actually works, and stop shipping changes you can't measure.
How to get an LLM to return data you can actually use: schema-first instead of regex, constrained decoding, validation at the boundary with Zod, retry-with-error-feedback, and when a clean second pass beats one more retry.
A six-axis scorecard — license, maintenance velocity, bus factor, escape hatch, docs honesty, prod fit — plus a 12-point go/no-go checklist you can run on your own shortlist in about 20 minutes.
AI coding agents run from autocomplete to bots that open their own pull requests. A vendor-neutral map of the rungs, what changes between them, and how to hand off real work.
An opinionated, no-hype tour of the open-source models, servers, and tools we'd reach for to ship an AI product in 2026 — and the five things that actually changed this year.
There's no shortage of AI news. There's a shortage of signal. The Cache is a store of what actually matters in AI and open source, explained clearly.