CERN and Mistral have partnered to develop a Large Physics Model aimed at enhancing AI capabilities in high-energy physics. This collaboration combines CERN's expertise in scientific data with Mistral's AI technology to create tools that could revolutionize analysis techniques and support broader scientific research.

OpenAI's announcement of a proof for the quasi-Riemann hypothesis marks a significant achievement in mathematical research, potentially unlocking new insights into prime number distribution. The implications of this proof are vast, prompting mathematicians to explore its connections to the longstanding Riemann hypothesis.

Mercury 3 is the latest diffusion-based language model, significantly faster and larger than its predecessors. Running up to 1,068 tokens per second, it excels in real-time applications like customer support and education with improved accuracy. Early partners are already benefiting from its speed and efficiency.

Edited open models can unknowingly harbor backdoors, posing serious risks. This article explores how a simple poisoning technique can be used to embed a backdoor within a model without detection. The authors successfully demonstrate this vulnerability, revealing the ease with which malicious behavior can be introduced and concealed.

AI adoption in enterprises is widespread, yet concerns persist over data accountability. A recent IBM survey revealed that only 26% of chief data officers believe their data capabilities can support AI effectively. The article explores the complexities of data compliance, lineage, and retrieval, highlighting the need for better tracking and management.

Researchers from MIT and Boston Children’s Hospital have developed a deep learning model called E(3)-Pose to enhance fetal brain imaging during MRI scans. This technology addresses motion challenges during imaging, aiming for clearer images that help detect developmental abnormalities early, potentially improving future interventions.

This article provides a practical guide on using tmux for managing AI coding agents like Claude Code. It covers setting up named sessions, efficient pane layouts, and scripting interactions. The guide also addresses limitations when running multiple agents, making it useful for developers seeking a streamlined coding environment.

Tech companies increasingly rely on writing as tangible evidence of compliance, effort, and learning. However, the rise of large language models (LLMs) threatens this practice, undermining the authenticity and value of written documents. As organizations grapple with this challenge, new methods of evaluation may emerge.

Mistral has launched Timestral, a new family of time-series foundation models that deliver accurate multivariate forecasts with as few as 5M parameters. Optimized for performance and efficiency, these models excel in various domains, offering probabilistic insights and accommodating both historical and future covariates.

A shift is occurring in biomedical research as new technologies, like "organ-on-a-chip," gain validity as alternatives to animal testing. These innovations could provide valuable avenues for drug development, but challenges in validation and adoption remain. Moving away from animal models requires changes in policy and scientific culture.

NorForge's investigation reveals the troubling use of Cairn, an AI-powered penetration testing tool, in malicious cyber operations. Over 20 publicly accessible Cairn panels were identified, exposing sensitive target information and demonstrating how threat actors exploit security weaknesses. The findings highlight serious risks for organizations unaware of these vulnerabilities.

Alan Turing's early chess program highlights the limitations of traditional computing, where human oversight dictates code execution. New AI advancements, such as doubleAI's code execution models, aim to optimize this process by allowing AI to autonomously improve running programs, ensuring efficiency without human intervention through formal verification techniques.

VibeTraining-Bench introduces a benchmark for evaluating agents' cross-component co-design capabilities in recursive self-improvement. By refining training data, trainers, and inference harnesses, agents aim to enhance system performance across various tasks. Early results show promising improvements, highlighting the potential for continued advancement in this area.

This study explores the effectiveness of various multi-trial aggregation methods for non-invasive brain-computer interfaces. It highlights that training-time logit averaging significantly improves performance, setting a new benchmark for phoneme classification and suggesting essential design considerations for neural decoding systems.

Fine-tuning large language models can compromise their safety features. This article introduces WSR-Tune, a novel approach that uses weight space reparameterization to preserve safety while adapting to new tasks. By separating safety and task-related information, it enhances performance without sacrificing safety.