All published AI news and analysis.

HSP GRUPPE is utilizing ChatGPT Enterprise to enhance productivity and the quality of work within its tax advisory services. This adoption aims to increase capacity for client service and core advisory tasks.

Concord MCP aims to provide a "Google Workspace"-like experience for AI agents, facilitating their collaboration and development. With a recent release and strong developer signals, it targets enhanced AI agent productivity.

OpenAI has released preliminary cybersecurity evaluations for its Astra system. The company is also detailing the security measures and controls being implemented to enhance its robustness.

Kinocut has released version 1.15.2, an open-source video editing MCP server and AI agent workflow engine. This update enhances sound mastering, introduces adversarial hardening, and expands the toolset for local video media manipulation by AI agents.

Researchers have introduced MIST, a new benchmark designed to evaluate how well language models can selectively trust external context, distinguishing between reliable and misleading information. They also propose SCOPE, a preference optimization method that trains models to balance trust in correct context with resistance to misleading signals, without sacrificing performance on clean or irrelevant contexts.

OpenAI has updated its changelog for ChatGPT and Codex, providing a section to review changes across repositories. This update aims to enhance transparency and developer understanding of modifications within their AI development tools.

GetStream has released version 0.6.8 of its open-source Vision Agents project, enabling developers to build low-latency voice and vision AI agents. The framework supports integration with various LLMs, STT, TTS, and vision models, along with real-time WebRTC capabilities.

A new study empirically compares programmatic tool calling (PTC) against native JSON tool calling across 14 language models on the BFCL v4 benchmark. Programmatic tool calling, where models invoke tools via Python scripts, matches or exceeds JSON tool calling performance in 11 of 14 models, demonstrating robustness and viability.

Researchers introduced Skill Entropy, a metric to quantify the difficulty of skill switching in LLM long-horizon reasoning tasks. They also released Skill^2-Bench, a benchmark designed to evaluate this capability across 558 skills and 9 domains, revealing a significant skill-switching gap in current models.

Google DeepMind has open-sourced its WeatherNext AI models, including WeatherNext 2 and WeatherNext Cyclones, which offer a significant advancement in cyclone forecasting accuracy. These models provide an additional day of warning time compared to previous methods, equating to roughly a decade of meteorological progress.