The API bill is the usual reason teams look at their own hardware, and privacy is the better one. Which benchmarks to trust for agents, how far you can quantize before logic breaks, what a single A100 on GCP costs, and serving Qwen 3.5-27B with vLLM.
WordPress.org now runs an MCP server for the plugin directory, so an AI assistant can validate readme.txt, check where a submission sits in the queue and read the guidelines from the source. The review team still holds you responsible for the code.
The Plugin Directory now ships a wp-env setup, so you can boot a local copy instead of guessing how its backend behaves. Here is what works inside it, and where it stops short of the live site.
A run at differentiable rule induction: the model wrote its own IF-THEN fraud rules, rediscovered V14 without help, and went dark in three of five runs. Notes on the consistency loss and when the architecture is worth shipping.
SAP-RPT-1 applies transformer pretraining to relational tables, so you feed it context rows instead of training a model per use case. My notes on the cost that moves to inference, the merge logic the API needs, and why one universal model still looks unlikely.
Zero-heavy data breaks a single regression, and the +1 log-transform hack does not save it. How I split CLV prediction into a hurdle classifier and an amount regression, with the two mistakes that bite in production.
The Connectors API gives WordPress one registry for external services, starting with AI providers. Notes on the init lifecycle, the unregister and register dance for overrides, and the keys that sit unencrypted in wp_options.
Shadow AI shows up because the official workflow is slower than the shortcut. Here is how I read that usage as a signal, and what a governed option looks like inside a WordPress stack.
Gemini Embeddings 2 is in public preview: one vector space for text, images, audio and video. The preview limits, calling it from WordPress with the HTTP API, and why cosine similarity on its own still is not enough.
Prompt caching takes up to 90% off input token cost and around 80% off latency, but only when the prompt prefix never moves. Here is how the pre-fill stage works, where the 1,024-token threshold sits, and how to order a prompt so it hits the cache.