OpenAI and the Pacific Northwest National Laboratory have launched DraftNEPABench, a benchmark that assesses AI coding agents’ ability to speed up federal permitting processes, potentially reducing NEPA drafting time by up to 15%. The tool aims to modernize infrastructure reviews through AI-driven automation.

GitHub Copilot has introduced new features to enhance AI-powered code generation, improving developer efficiency and integration with the GitHub ecosystem. These updates focus on making AI tools more intuitive, accurate, and useful for real-world coding tasks.

AWS has enhanced its Large Model Inference (LMI) container with LMCache support to address rising costs and performance challenges in long-context LLM deployments by caching frequently reused content. This reduces token redundancy and improves efficiency, especially for use cases like Retrieval Augmented Generation and coding agents. The updates simplify deployment and deliver measurable cost and performance gains across popular model architectures.

OpenAI and Figma have launched a Codex integration that links code and design, allowing developers and designers to switch seamlessly between implementation and the Figma interface for faster iteration and product delivery.

Agentic reinforcement learning (RL) enables LLMs to learn decision-making processes through interactive, multi-step interactions with environments, improving long-horizon planning and adaptability. The article details practical challenges in training GPT-OSS for agentic RL, such as on-policy integrity issues and memory inefficiencies, and offers solutions like fixing MoE log-probability mismatches and optimizing attention mechanisms.

Reinforcement fine-tuning (RFT) for Amazon Nova enables AI models to learn through evaluation rather than imitation, making it effective for domain-specific tasks where step-by-step reasoning is valuable but labeled data is scarce. It offers a practical alternative to supervised fine-tuning by using feedback and rewards to guide model behavior in code generation, customer service, and other specialized applications.

Maggie Hulce, Indeed’s CRO, discusses how AI is reshaping job searches, recruitment, and talent acquisition by improving efficiency and personalization for both employers and job seekers.

AI significantly accelerates COBOL modernization but relies on context beyond source code. Mainframe modernization involves two phases: reverse engineering (understanding existing systems) and forward engineering (building new applications), with the first phase being critical for success. Coding assistants excel only in the second phase, highlighting a gap in AI’s current capabilities.

Mixture of Experts (MoEs) offer a sparse, efficient alternative to dense language models by dynamically activating only relevant expert networks, reducing computational cost and memory usage while maintaining performance. This approach addresses the scalability limitations of dense models in training and inference. Hugging Face’s recent engineering improvements enable easier implementation and optimization of MoEs in Transformers.

The article introduces Alyah, a benchmark focused on evaluating Arabic large language models’ capabilities in understanding the Emirati dialect, which differs significantly from Modern Standard Arabic. It highlights the lack of dialect-focused evaluations in current Arabic LLM benchmarks and emphasizes the importance of cultural and linguistic authenticity in real-world interactions.