AI governance startup nFOX Lab, which came out of stealth this week, launched a platform for managing company-owned language models. It also plans to open-source Blockhouse, a sandbox that limits what AI workloads can access and do, to help enterprises adopt open models on their own infrastructure.
A Frandroid journalist tested whether an AI model running entirely on a laptop with Nvidia's RTX Spark chip could replace ChatGPT during an 11-hour flight from San Francisco to Paris. It worked for answering questions, but not yet for everything else, and battery life was a constraint.
Nvidia has released new Nemotron models, including Nemotron 3.5 Super VL. They join a wave of open-weight releases from US labs that followed the Hugging Face hack.
Meta has launched Muse Spark Open, an open model with 816 billion parameters, 42 billion of them active. Meta positions it as a response to the open-weight models coming out of China.
AI makes it easier than ever to find potential vulnerabilities in open-source software, which is swamping maintainers with reports. The New Stack looks at how Akrites is handling that flood.
Even agents built on top models struggle to complete complex tasks inside a given product reliably, so teams write large numbers of evaluations by hand. CopilotKit's CLiF uses agent traces and user actions to generate skills, evals and fine-tuning data instead.