Researchers propose DivInit, a method to improve agentic search by diversifying initial queries in parallel sampling, avoiding redundant evidence retrieval. The approach boosts performance by five to seven points on multi-hop question-answering benchmarks without additional training. The technique is tested across five open-weight models and eight datasets, offering a compute-efficient alternative to standard parallel sampling.

Researchers propose DivInit, a training-free method to improve agentic search by selecting diverse initial queries before running parallel search trajectories. The paper reports that standard parallel sampling suffers from redundant first queries and overlapping retrieved evidence, while DivInit improves performance across five open-weight models and eight benchmarks, including average gains of five to seven points on multi-hop question answering at matched compute.

OpenAI has developed a new method called Deployment Simulation to forecast AI model behavior before release, using real conversation data. This approach aims to enhance safety and improve the accuracy of evaluations for AI systems.

OpenAI introduced Deployment Simulation, a method for predicting how AI models may behave before they are released. The approach uses real conversation data to improve safety assessments and make evaluations more accurate.

A new Web Component called "click-to-play" allows users to display a static image that only loads and plays a GIF when clicked, reducing unnecessary data usage. Developed by Simon Willison, the tool is designed to improve performance by preventing large GIFs from loading automatically. It was created to enhance a demonstration of Datasette’s row editing features.

Simon Willison introduced a progressive-enhancement Web Component called <click-to-play> that displays a still image with a play button and loads the linked GIF only when clicked. The tool is intended to avoid loading large GIFs unnecessarily and was built for a post demonstrating new row editing tools in Datasette.

Google DeepMind is highlighting the use of AI to speed up the UK planning process for housing. The effort is framed as a way to help unlock house-building by making planning systems faster and more efficient.

Google DeepMind is partnering to use AI to speed up the UK’s housing planning process, aiming to reduce delays and boost construction. The technology could analyze complex planning applications and regulations more efficiently than traditional methods. This initiative seeks to address housing shortages by accelerating approvals while maintaining regulatory standards.

Ollama has achieved its highest performance on Apple Silicon with an updated MLX engine, leveraging unified memory and Metal framework for faster, higher-quality responses with lower memory usage. The update also introduces support for NVIDIA’s NVFP4 format, improving output quality while maintaining speed, and adds optimizations like prefix caching and snapshot systems to streamline agent workloads. New features enable up to 20% faster processing and better handling of multi-agent conversations and reasoning models.

Ollama updated its MLX engine for Apple Silicon, promising faster responses, lower memory use and higher-quality model outputs by using Apple’s unified memory and Metal-backed MLX framework more extensively. The update adds support for NVIDIA’s NVFP4 model-optimized format, introduces performance optimizations of up to 20%, and adds snapshot-based state caching to improve agent, reasoning and branching workflows.

Anthropic has surpassed OpenAI in business spending market share for the first time, despite an ongoing feud with the Trump administration that led to the removal of its latest AI models from the market. The company’s defiance of government demands—including a ban on non-American access to its advanced models—appears to have bolstered its reputation, with sales data suggesting the controversy may actually boost its adoption. While the financial impact of pulling its newest models remains unclear, business spending on Anthropic’s existing models, particularly Claude Opus, continues to grow.

Anthropic is facing renewed pressure from the Trump administration, which demanded restrictions on access to its latest advanced AI models and effectively pushed the company to pull them from the market. Business spending data from Ramp suggests the dispute may boost Anthropic’s appeal, with the company recently surpassing OpenAI in business AI subscription share and seeing strong adoption of its Claude models.