Checking incoming requests alone could not reliably tell a PHP attack apart from normal traffic.

A firewall feature that checks each request together with the server's reply, and blocks the reply if it shows the attack worked. It is turned on for all plans, including free ones.

An essay grader built on Qwen3-4B used a general-purpose vocabulary. That vocabulary split the text the grader writes most often into many tokens, and at every step the model scored 151,669 possible tokens.

A new vocabulary built around how much each token costs to generate, not just how often it appears. The base model and training recipe stayed the same. The paper is still under review, so the numbers may change.

Predicting things like customer churn from company databases usually takes weeks of hand-built features and a separate model for each task.

A pre-trained model that works directly on database tables. For each question it reads up to 10,000 past examples and gives its prediction in one pass.

GitHub needs its Git storage to handle the much heavier load that AI agents bring.

A system that copies the remaining repositories using a separate vote, or quorum, for each branch and tag. Object storage lets reading capacity grow on its own.

World action models are AI models that guess how a scene will change before choosing a robot's next move. They are slow because of their large video models.

Faster GPU code found with AMD's agent tool GEAK, fewer clean-up passes, called denoising steps, and a smaller part of the model that chooses actions. The code is published.