Letting an AI agent change people's files is risky: it needs software to do the work, limits on what it can touch, and a way to undo mistakes.

Two engineers' side projects were merged: an isolated sandbox on Dropbox servers with installed software, and a desktop interface with permission checks. Changes are tied to version history so they can be reviewed and rolled back.

Teams sharing one GPU between many workloads need finer control over who gets how much of the chip and for how long.

CUDA 13.5 makes Vera Rubin support production-ready and updates the Multi-Process Service so workloads can get weighted turns and borrow or lend unused parts of the GPU.

Coding agents send very long, growing prompts and run tasks in parallel, which fills up the GPU's memory cache of earlier work.

The team moved overflowing cache data to the host computer's memory, sped up the steps that write each new word, and sent fixes upstream to the SGLang serving software.

Each AI framework on Kubernetes describes its jobs differently, so every scheduler or dashboard needs custom code for each one.

A shared description format plus a Go library that tells tools where to find a job's pods, replica counts and status, with a catalog of ready-made definitions.