Waiting on a full JSON payload from an LLM leaves users on a spinner for 15 seconds. Streaming with server-sent events, the ReadableStream code that consumes it, and when validation means you should not stream at all.
AI writes the query now and I check it. That includes the time an elegant LLM-generated join pushed a production database to 98% CPU, plus the utility I use to export sanitized WooCommerce data for analysis.
The Bits-over-Random Metric compares your retrieval against random chance, which is how you catch the context pollution a high Success@K hides. Notes on staged tool routing and where the collapse regime starts.
The six-month version of this career change is a sales pitch. Coming from a software background it took closer to two years, and the money goes to people who can ship a RAG system that does not leak data, not to people who can write prompts.
A colleague shipped a custom assistant in a couple of hours by wiring Google Calendar, Microsoft To Do and Notion together. Execution stopped being the bottleneck, and here is a Gemini API scheduler you can start from inside WordPress.
My agent failed the same caching refactor three times before I fixed how I was briefing it. Planning mode, browser access through Playwright MCP, and claude.md files are what turned that into one-shot work.
A decade of explaining why WordPress sites don’t feel like apps, and what 2026 changes: blocks described to an AI instead of scaffolded by hand, the Interactivity API handling state without page reloads, and theme.json holding the design system.
The generative AI boom is hiding a compute bottleneck. Notes on Sara A. Metwalli’s argument for quantum computing in data science, the problems QAOA handles differently, and what an offloaded optimization call could look like from a PHP hook.
Drift usually means a scramble for labels and a multi-hour retraining job. This is the other route: a trainable ReflexiveLayer over a frozen backbone, label-free drift triggers, and async weight updates that never block inference. It recovered 27.8 points of accuracy and cost recall.
My agents decided their own heartbeat config was redundant and deleted it, twice. The tiered OpenClaw setup that came out of it: orchestrators for judgment, throwaway personas for tasks, and five Markdown files that hold an agent’s identity.