Ai2's XYBench tests whether LLMs catch users' wrong assumptions

AMD AI Group's experimental hrx-loom bundles a compiler, runtime, and serving tools to build, optimize, serve, and share kernels. An AMD engineer (jfowers_amd) reports using it to produce speedy Strix Halo kernels for the top-51 LLMs.

The South Korean energy and chemical company splits model development and operations between process engineers and dedicated ops staff, connecting engineer-built models to production workflows via SageMaker Pipelines.

Projects in Claude Code now let Claude coordinate tasks across parallel threads, with a more proactive coordinator answering questions in project chat and reporting thread status. The redesign is in public beta on Pro and Max, rolling out in stages with a waitlist.

Open-source Photoshop alternative LayerShop now supports generative fill for adding and removing objects using local models. The latest version also adds support for Minimax H3 Image.

AWS post walks through using NVIDIA NeMo Switchyard on Amazon Bedrock to send each coding-agent request to the right model, noting a single task can take a few model calls or dozens. It frames cost around the task rather than one call.