There are thousands of ways to set up a server that runs a language model. Testing each one ties up expensive GPUs for minutes, so teams guess.

A tool that runs a few short real measurements, then calculates how every other setup would perform on an ordinary CPU. Only a shortlist is tested on GPUs.

AI agents now run for hours, but if the server running them is upgraded or fails, the work can be lost.

An open-source runtime for Kubernetes that saves each agent session in a database, so another server can pick up the run, and runs agent code in a locked-down sandbox.

Privacy Pass tokens let sites limit abuse without tracking users, but they use encryption that a future quantum computer could forge.

A design called MOLE, built on a new kind of blind signature, that hides even which issuer vouched for a user and can move to quantum-safe maths.

Dropbox's web and desktop apps were built differently, so reusing web features on desktop meant a lot of integration work before a fixed launch date.

Shared React components for both apps, with coding agents handling defined tasks like interface changes and tests, which engineers then reviewed.

Developers need stable tools before running real workloads on Nvidia's new Vera Rubin GPUs, and more control over how jobs share one GPU.

A toolkit release with full Vera Rubin support, new ways to split and share GPU cores between workloads, and updates to its Python and library tools.