A report indicates that women in the UK arts feel they lack equal opportunities for roles, especially as they age. A survey of 76 arts professionals found that 95% believed gender inequality persisted, and 75% experienced unconscious bias or sexism. Ageism negatively affects career success, with over two-thirds feeling age impacts opportunities. Gender discrimination and ageism are worsened by AI in hiring, which can penalize women for career breaks. The research calls for more funding, representation targets, and algorithmic audits to address these issues.
A custom ensemble weather forecasting workflow was built using NVIDIA Earth2Studio. This involved implementing a custom wind-power diagnostic and a variable-scaled perturbation system to generate physically appropriate ensemble members. The process utilized Earth2Studio's APIs for batching, coordinate mapping, and Zarr data storage to execute the forecast iterations. Forecast accuracy was verified using latitude-weighted RMSE, fair CRPS, and spread-skill ratios against GFS analyses. The resulting dataset was visualized with spatial maps, spaghetti contours, and lead-time skill curves, demonstrating a flexible and extensible framework.
Claude offers several specialized skills for data scientists, including Deep Research for comparing modeling approaches by synthesizing web searches and research. It can generate HTML project briefs for stakeholders to communicate complex project plans clearly. The Claude Design tool specializes in creating visually aligned documents like stakeholder slide decks. Claude Code excels at generating documentation, such as READMEs, by understanding existing codebases. Experimenting with these specialized modes based on the task is recommended for optimal results.
Codex is preferred for completing single, difficult coding tasks, whereas Claude Code excels at orchestrating multiple smaller tasks. Codex struggles with managing numerous parallel tasks, often forgetting smaller items when executing multiple requests. Conversely, Claude Code is better at orchestrating sub-agents for many smaller tasks and designing front-end changes. Model selection should be guided by observing performance and using quantitative metrics like development time to discover each agent's strengths and weaknesses.
The RAG approach is only one part of the toolkit, and real problems still require more detailed engineering. Bonus articles provide specific, actionable insights by exploring cross-cutting concerns, alternative pipeline shapes, and reproducible benchmarks. These pieces focus on practical disciplines like handling noisy text, justifying model uncertainty, and creating localized LLM stacks. They are distinguished by their ability to connect to multiple pipeline components and offer durable methodological guidance. Finally, they demonstrate that the evaluation of RAG systems requires careful, layered investigation beyond a single framework.
A physics-grounded LLM benchmark called Stickblade Arena pits LLMs against each other as physical agents in a 2D arena governed by a physics engine. Matches resolve based on physical interactions, and human voters assess which agent "fought smarter" by watching replays with masked identities. The system calculates a 6-axis Elo rating based on different axes of physical constraints, hypothesizing that different constraints stress different reasoning skills. Evaluation shows a correlation between human votes and objective win-rates, though specific weapon types lead to disagreement. The benchmark includes 17 LLMs and non-LLM baselines, providing a comparative measure of reasoning in physical environments.
Agentic GraphRAG Blueprint is a reference architecture for question answering over large document collections using a knowledge graph combined with vector search. It features incremental ingestion by skipping unchanged files using content hashing to manage token costs as the corpus grows. The system employs hybrid search modes, allowing for local fact-level answers or global cross-document synthesis. It supports domain-agnostic LLM prompting through Leiden-based community detection. The blueprint can be deployed locally via Docker or in the cloud using Terraform and CI/CD.
Anthropic previewed the Model Hardware Standard (MHS), a shared specification enabling AI agents to discover and operate physical devices. This standard addresses the complexity of physical instrument integration, which currently requires manual translation and long setup times. MHS standardizes the device driver, exposing basic controls like read and write functions for agent interaction. Integration time is drastically reduced, with examples showing setups that took weeks to hours. Practical applications demonstrated significant speedups in complex tasks, such as a laser relock procedure.
Code-as-World represents a physical scene as an executable world representation (EWR) in a `scene.json` format that can be run in MuJoCo physics simulations. An agentic loop recovers this executable program from real video by proposing, instantiating, executing, and verifying simulations over up to five rounds. This process yields verified worlds with exact physical labels, which serve as superior training supervision compared to raw video. The Code-as-World-VL-9B model achieved a 55.4 MRA score on QuantiPhy validation, surpassing Gemini-3.1 Flash. The system is deployable for research and prototyping, released with model checkpoints under Apache 2.0.
A $400 robot named Microduck was shipped using a full sim2real Reinforcement Learning pipeline. This project has gained viral attention, with many people training behaviors on real hardware. The system is open source, allowing the community to train new policies. This democratizes the process of training RL policies for physical robots. Affordable robots and coding agents are making this kind of training more accessible.