All published AI news and analysis.

KiwiFS is a new Go-based markdown filesystem designed to make markdown files searchable, versioned, and writable for AI agents and teams. It offers a unified binary with zero configuration, aiming to solve the limitations of raw markdown files in agent workflows.

GitHub Copilot now supports Gemini 3.1 Pro (Preview) and Gemini 3.5 Flash across its CLI, cloud agent, and Copilot app. Access to these models varies by Copilot subscription tier.

GitHub Copilot's cloud agent now supports scheduling and automating tasks, enhancing its utility for developers.

GitHub Copilot's /chronicle feature now offers a more comprehensive view of agent sessions across various platforms, including IDEs and the GitHub Copilot app. This update aims to provide developers with better insights into their work history, enabling personalized summaries and custom instructions.

OpenAI has announced new Codex plugins, sites, and annotations designed to enhance AI capabilities for various professional roles. These updates aim to help analysts, marketers, designers, and investors leverage AI more effectively in their workflows.

Travelers has deployed an AI-powered Claim Assistant, developed with OpenAI, to assist customers with filing claims and offer round-the-clock support. This system is designed to help scale operations during periods of high demand.

OpenAI is advocating for international collaboration on AI safety for young people. The company proposes establishing a global institute to enhance safeguards, set standards, and create opportunities for youth in the AI field.

A new report, "The Next Era of Knowledge Work," details how OpenAI's Codex is enhancing productivity. It highlights AI-powered capabilities in research, data analysis, workflow automation, and content creation.

Researchers have identified and analyzed a phenomenon called Perceptual Judgment Bias in multimodal large language models (MLLMs) used as judges. This bias causes MLLMs to favor plausible narratives over perceptually correct answers when visual evidence conflicts with text. A new dataset and training framework aim to improve the perceptual fidelity and robustness of these MLLM judges.

Researchers introduce Staged Executable Inverse Graphics (SEIG), a framework that uses pretrained vision-language models (VLMs) to reconstruct editable 3D scenes from single images. SEIG breaks down the complex inverse graphics problem into sequential stages, directly generating executable Blender code for scene elements like geometry, materials, and lighting.