Automated requests now make up the majority of HTML traffic, and a growing share of them happens while a person is waiting for an answer. The piece distinguishes three routes such agents can take — a tool interface via MCP, a machine-readable output format, and the user interface itself — checks which of them the mundo.schule portal actually offers, discusses when a conventional API remains the better choice than an MCP server, and closes with a checklist to run yourself.

Every AI request consumes electricity, water, and compute time — how much depends heavily on model, provider, and usage. An overview of what the research shows, why agentic workflows are especially token-hungry, concrete ways to save tokens, why models and providers differ, and when local models running on your own renewable power are the better choice.

From slide‑plus‑speaker to a production core in code: speak ideas freely, transcribe, structure, generate with Hyperframes, export as video.

Cloud changes who runs operations – not whether operations is needed at all. A piece on provider responsibility, BSI Grundschutz and C5, fast security rollouts in the AI age, the PCI project, and why good cloud must never cost more than an equivalent on-prem solution.

Hardly anyone still works without AI support — but the agents behind that support usually remain isolated islands, tied to one person and one project. AGENTS.md helps share project knowledge, but only shared skills, shared context, and tools like Buzz from Block turn a collection of individual agents into an actual team.

Fast internet, working Wi-Fi, and enough devices are not the last mile of educational infrastructure. Offerings only arrive at scale once they keep pace on features, onboard people instead of leaving them alone, and are carried by communities.

Choosing an AI model determines quality, cost, latency, and resource use. Instead of leaving users alone with an ever-growing model menu, what’s needed is a lean, transparent routing layer: it classifies requests and, within clear rules, picks the smallest model that fits.

H5P was important for bringing interactivity into learning platforms. The next step could be a decoupled architecture of semantic task formats, AI-supported authoring and porting skills, renderer libraries, and optional AI/tutor integration.

Educational media infrastructure is not decided only at the level of interfaces and platforms, but where teachers and learners actually work: on personal devices, on digital boards, and across the connected classroom.

Open educational infrastructure does not emerge from the absence of rules, but from stable interfaces, transparent governance, public code, and reusable system services.