The standard account says neural networks learn low frequencies first, so frequency content alone decides how hard a target is to fit. New experiments show the shape of the target near the edges matters too, and that changes which fix you reach for.
A multi-agent AI workflow is a few model calls in a row, each with its own role prompt. The pattern itself is simple. What decides whether it survives on a real WordPress site is the queue behind it, and what each agent actually sees.
Most WordPress AI plugins scan every page on every reindex, so the bill grows with your content. There are three ways to route content to a model, and the block tree gives you one of them for free. I compare them and say when each pays off.
A llama.cpp project that prefills a shared document once and fans the KV cache out to many agents is really a lesson in prompt order. Here is how the same shape shows up when your agents are just API calls from PHP, and how to check it.
Automattic shipped ten experiments during Radical Speed Month, and most of them only exist on WordPress.com. The two that don’t are Desktop Mode and Studio’s Blueprints Gallery. Here’s how to sort the list, what a blueprint file looks like, and what I’d actually adopt.
A developer built a World Cup forecaster from 49,000 international matches, and the results say more about shipping prediction features than about football. The honest parts: draws that never get predicted, data leakage hiding in a rating column, and a simple regression that tied the boosted model.
Modern AI runs on CPUs, GPUs, TPUs and NPUs, and a WordPress site only ever meets two of them. A look at where inference actually runs, how to call a model without wrecking your server, and when the visitor’s own hardware can do the work.
LLM recommendations on a store get affordable when a free query builds the shortlist and the model only reranks thirty candidates, with the result cached. This is the WooCommerce version of a restaurant recommender pattern, including the PHP and the caching most posts skip.
Nathan Bos argues we should train AI to betray its users, because obedient superintelligence is a lone bad actor’s dream tool. I agree with where he ends up and disagree about the layer: for sites and stores, that decision belongs in permissions code, with a human on the send button.
A fine-tuned Mistral model now classifies 15 emotions in short text, and the weights are public under Apache 2.0. I look at what it took to train, how I’d wire it into WooCommerce review triage, and the parts I would not trust on a real store.