Brief

Meta employees are protesting a new mandatory software tool that tracks their keystrokes and mouse movements for AI training purposes. An internal post detailing these concerns has gone viral within the company, with many employees viewing the surveillance as an invasion of privacy and an exploitation of their data. This initiative has contributed to record-low morale and is fueling employee unionization efforts, particularly in the UK. AI

Researchers have developed a new framework to improve the accuracy of clinical timelines extracted from text by aligning it with structured electronic health record (EHR) data. This retrieval-augmented multimodal approach first identifies key anchor events in narratives to create a temporal backbone, then places other events relative to this structure. The system further refines the timeline by using retrieved EHR data as external temporal evidence, demonstrating a significant improvement in absolute timestamp accuracy and temporal concordance compared to text-only methods. AI

Researchers have developed EviScreen, a new framework for disease screening using medical images that aims to improve both interpretability and performance. The system utilizes region-level evidence from historical cases, stored in dual knowledge banks, to inform its predictions. This evidential reasoning approach allows for more transparent decision-making compared to traditional methods that rely on post-hoc saliency maps, and it has demonstrated superior performance on established benchmarks. AI

Researchers have developed a new framework for open-ended image editing that can handle complex, multi-step instructions. This approach uses a planner to break down tasks into smaller steps and an orchestrator to select appropriate tools and regions for execution. A vision-language judge provides feedback on instruction adherence and visual quality, which is then used to refine both the planner and orchestrator, leading to more coherent and reliable edits than existing methods. AI

Researchers have developed Articraft, an agentic system that uses large language models to generate articulated 3D assets at scale. This system translates the task of asset creation into program writing, utilizing a domain-specific SDK for defining parts, geometry, and joints. The generated assets are validated, and the system has been used to create Articraft-10K, a dataset of over 10,000 articulated assets across 245 categories, demonstrating utility in robotics simulation and virtual reality. AI

Netflix has established a new studio called INKubator, which will focus on creating short-form animated content using generative AI workflows. While currently hiring for key roles and focusing on shorts, the studio has ambitions to eventually produce longer-form content. This move aligns with Netflix's broader embrace of AI, which it already utilizes in advertising and has explored through acquisitions. AI

Laurel, an AI platform company, is developing a solution to automate the tedious process of manual timesheet logging for legal professionals. CTO Jason Li discussed the inefficiencies of current methods, where lawyers spend significant time manually recording billable hours. Laurel aims to streamline this by applying AI to the billable hour system. AI

A new open-source tool called Heimdall has been released to provide observability for MCP (Model Communication Protocol) calls. It acts as a transparent proxy, capturing JSON-RPC messages between clients and servers and converting them into OpenTelemetry spans. This allows developers to monitor tool performance, identify silent failures, and inspect inputs without modifying the target server. AI

The author details how they constructed an AI news intelligence pipeline using a combination of open-source tools and personal integrations. This pipeline leverages Kestra for orchestration, Ollama for running local language models, and integrates with Notion and Slack for data storage and communication. The project aims to provide a cost-effective method for staying updated on AI developments without relying on expensive commercial APIs. AI

A user details how they leverage Anthropic's Claude AI model to automate the process of applying for jobs. The system is designed to take job descriptions as input and generate customized resumes and cover letters. This approach aims to streamline the application process by quickly producing tailored documents for each opportunity. AI

Amazon Bedrock AgentCore now supports Chrome enterprise policies, allowing organizations to control AI agent web browsing behavior. This feature enables granular control over URL filtering, download restrictions, and password manager usage, enhancing security by preventing agents from accessing unauthorized sites or handling sensitive data improperly. Additionally, the integration supports custom root CA certificates, enabling agents to securely connect to internal services that use private certificate authorities. AI

NEES Core Engine, a runtime layer designed for governed AI applications, is seeking developers for testing and review. This platform aims to manage the interaction between AI applications and their underlying models. The call for feedback is currently open on dev.to. AI

A developer has created a client-side tool to help estimate costs associated with various Large Language Models (LLMs) before committing to one for a project. The tool, accessible at llmtokens.vercel.app, provides real-time cost breakdowns for over 25 models, including popular options like GPT-4o, Claude, and Gemini, allowing users to compare pricing based on token usage. This aims to shift the cost consideration to the architecture phase, preventing potential bill shock later in development. AI

LangGraph, a framework for building stateful, multi-agent applications, has introduced new features for checkpointing and time-travel. These capabilities allow developers to save and revert the state of their AI applications, enabling more robust debugging and experimentation. This functionality is crucial for managing complex AI workflows and understanding their execution history. AI

A 30-day monitoring project revealed significant reliability differences among major LLM providers. OpenAI experienced frequent and lengthy outages, while DeepSeek had a concerning number of silent failures that went undetected by standard monitoring. Groq offered impressive speed but suffered from fragility and rate-limiting issues, whereas Azure OpenAI provided the highest uptime but came with increased costs and longer provisioning times. Anthropic's Claude demonstrated consistent performance, suggesting it is a reliable choice for production environments. AI

Anthropic has launched a version of its Claude AI assistant tailored for small businesses. This new offering aims to support various business operations, including sales strategies and invoicing. The release prompts broader discussions about the role and accessibility of advanced AI tools for smaller enterprises. AI

Dell has launched its new 14S and 16S laptops, designed as successors to the previous Plus models and positioned below the XPS line. These laptops are built for general productivity and feature on-device AI processing capabilities, powered by Intel Core Ultra Series 3 or AMD Ryzen AI 400 Series processors. Both models offer extended battery life, with the 14S providing up to 24 hours of productivity and the 16S up to 26 hours of streaming, and are available in two color options with various memory and storage configurations. AI

Germany's domestic intelligence agency, the Verfassungsschutz, has purchased a European alternative to Palantir's software. This move aims to enhance digital sovereignty and data protection within the agency. The chosen platform, ArgonOS, is an open-source solution developed in Europe, contrasting with Palantir's US origins. AI

RizzAI is a new AI-powered dating assistant designed to help users craft better messages and dating profiles. The tool generates openers, replies, and bio ideas, aiming to provide inspiration that users can then personalize. It is positioned as a helpful tool for individuals struggling with communication on dating apps. AI

Google's new AI-powered health coach has been found to fabricate workout data and progress, leading to what critics are calling "phantom workouts" and "phantom gains." This suggests the AI is generating achievements that do not reflect actual user activity or results. The issue raises concerns about the reliability and accuracy of AI-driven health and fitness tracking. AI

NVIDIA has introduced a concept called 'claw,' which represents AI moving from theoretical suggestions to practical execution. This advancement in autonomous agents aims to automate complex tasks in the background for 24 hours, signifying a new phase in workflow automation. Concurrently, Perplexity has integrated with Snowflake, enabling its AI to perform end-to-end operations on live warehouse data using SQL, tables, filters, and metrics to deliver precise answers. AI

Microsoft's Power Apps has launched a new closed-loop learning feature for its agents. This update allows agents to learn directly from user corrections, transforming feedback into organization-wide improvements. The system aims to reduce manual fixes and enhance automation accuracy over time. AI

New security vulnerabilities have been discovered and exploited across Windows, macOS, and ChatGPT. These issues highlight a growing concern at the intersection of artificial intelligence and cybersecurity. The specific details of these exploits are still emerging, but they point to a significant and evolving threat landscape. AI

Meetily is a new open-source AI meeting assistant designed with privacy and local processing as its core features. It offers the ability to record, transcribe, and summarize meetings directly on a user's device, ensuring data never leaves the local environment. By leveraging offline support and local AI models via Ollama, Meetily provides users with complete control over their meeting data, positioning itself as a privacy-focused alternative to cloud-based services. AI

Consensus.app is a new tool designed to accelerate the discovery of peer-reviewed research. Users can ask questions in natural language, and the platform will search scholarly literature, provide a yes/no/mixed evidence summary, and quickly highlight relevant papers. This aims to make academic research more accessible and efficient for users. AI

Testo has released a new feature called AI skills, which are small, on-demand instructions for AI agents. These skills can be triggered and loaded as needed by agents like Claude Code, Codex, and Cursor. Additionally, the project has launched llm/skills v1.0.0, a Composer plugin designed to automatically synchronize these skills from a PHP vendor directory into a user's project. AI

The ACT protocol has released a suite of components, available on ghcr.io/actpkg, designed to extend its capabilities. These components, licensed under MIT or Apache 2.0, include data and storage solutions like SQLite with vector indexing, generic filesystem access, and secure key storage. Network bridges for HTTP and OpenAPI specifications, along with utility components for cryptography, encoding, randomness, and time, are also available. Additionally, a Python interpreter component demonstrates the ability to run untrusted code within the ACT sandbox. AI

Meta is enhancing its Ray-Ban smart glasses with new virtual writing capabilities. Users can now compose messages using hand gestures, with support for applications like WhatsApp and Messenger. This update aims to make communication more intuitive and accessible through the smart glasses interface. AI

A user expressed surprise and disappointment with Google Sheets' built-in AI capabilities, finding it struggled with a basic task. The AI required four attempts to correctly deduplicate a row, leading the user to describe the experience as "absolutely awful" for such a simple function. AI