Plugging in the OpenAI API is not engineering. Notes on what actually matters when you build LLM features on WordPress: token costs, attention limits, RAG over fine-tuning, and keeping a 30-second API call from taking down PHP-FPM.
LLM summarization pipelines invent facts when the schema demands a value the transcript cannot support. A three-stage design, extraction then synthesis then an audit stage that can only weaken claims, and why empty sections are correct.
The CSPNet architecture splits feature maps so a lighter CNN does not cost you accuracy. What it fixes in DenseNet, how the two paths work, and the PyTorch code for the bottleneck block.
Where the WordPress Core AI work stands before 7.0: the MCP adapter moving to a standalone plugin, the connector approval experiment, and a release plan that puts polish ahead of new features.
RAG re-derives everything on every query, so nothing accumulates. This is the Raw and Wiki folder layout, the three control files, and the daily, weekly and monthly jobs I use to keep an AI context layer my agents can read.
WordPress.com opened the AI Assistant to every paid plan and shipped Studio Code, a beta CLI agent for local work. Notes on what the CLI does with block markup and what the plan change means day to day.
The WooCommerce REST API documentation left its 4.2MB single-page site for per-endpoint files on developer.woocommerce.com. What that changes for coding assistants, raw markdown fetches, and doc updates that ship inside the PR.
Asking an LLM whether a value changed significantly turns a cheap comparison into a slow, untestable guess. Here is the split I use in WordPress: PHP thresholds decide when to act, and the model only writes the message.
Timer-XL makes the case that decoder-only Transformers beat encoders at forecasting. This looks at how TimeAttention keeps temporal order intact, why patch tokens cut the compute, and what the zero-shot numbers mean for real pipelines.
WooCommerce for Claude came out of Automattic’s Radical Speed Month. It uses MCP to give the model analytics lookup tables, a store knowledge layer and an AI-readiness score, all on one endpoint. Plus the provider pattern and how to run it locally.