Travelers Deploys OpenAI Claims Assistant NationwideVercel Adds Bulk Git Configuration SettingsGitHub Launches Copilot Desktop AppGitHub Unveils Agent Development PlanMicrosoft Launches MAI-Thinking-1 and MAI-Code-1-Flash ModelsNVIDIA and Microsoft Collaborate on New ComputerNVIDIA Unveils Cosmos 3 and Nemotron 3 Ultra ModelsOpenAI Launches Role-Specific Codex PluginsCodex Evolves into Universal Productivity ToolOpenAI Proposes Global AI Safety InstituteDLLM-JEPA: New Architecture for Masked Diffusion Language ModelsBudgetDraft Boosts Sparse-KV Speculative DecodingNew Benchmark for LLM Interactive ReasoningDelayed Reward Attribution for Multi-Agent GamesUniversal Quantum Transformer for Discrete RulesTIGER: Hallucination Repair for Multimodal AIDatasette Agent Adds MicroPython Sandbox Executionmicropython-wasm 0.1a1: Enhanced Python SandboxingCodex Desktop Adds Pasted File EditorCalifornia Brown Pelicans at Microsoft BuildTravelers Deploys OpenAI Claims Assistant Nationwide

Vercel now enables users to configure Git settings for all projects in a repository at once, including commit statuses and repository_dispatch events. Previously, each project required manual configuration. This update simplifies multi-project deployment management.

GitHub announced the Copilot desktop app at Build 2026, providing a native agent experience. The new tools enable agents to work in ways familiar to users, improving interaction with AI coding assistants.

GitHub pioneered AI coding with Copilot, but the surge in coding agents has strained the platform. GitHub outlines a development plan to optimize its platform for the growing number of AI coding agents.

Microsoft announced two new text LLMs: MAI-Thinking-1 (35B parameters, reasoning-focused, limited to early partners) and MAI-Code-1-Flash (5B parameters, optimized for GitHub Copilot and VS Code with higher performance and lower cost, now available).

NVIDIA and Microsoft partnered to develop a new computer system based on NVIDIA Cosmos 3 and Nemotron 3 Ultra, featuring RTX Spark technology. The system integrates both companies’ AI and hardware strengths for significantly enhanced performance.

NVIDIA announced the Cosmos 3 ecosystem and Nemotron 3 Ultra model at Computex. Nemotron 3 Ultra is an open-source foundation model supporting commercial use, while RTX Spark accelerates fine-tuned inference.

OpenAI updated Codex with role-specific plugins and tools for analysts, marketers, designers, and other teams. The new system integrates AI assistance into professional workflows to boost productivity.

OpenAI’s new report reveals how Codex transforms knowledge work through AI-powered research, data analysis, and workflow automation. It supports 200+ scenarios with 40% faster document processing. Integrates with Microsoft 365 for team productivity.

OpenAI calls for an international AI safety institute to protect young people. The initiative aims to coordinate global safety standards and regulations, with support from 10 organizations including UNESCO. Focus on shielding minors from harmful AI content.

The arXiv paper introduces DLLM-JEPA, applying Joint Embedding Predictive Architectures to masked diffusion language models. This method addresses the high computational costs of causal attention mechanisms, improving self-supervised learning efficiency.

Stanford researchers propose BudgetDraft, an acceptance-aware multi-view training method for sparse-KV speculative decoding. It reduces GPU memory usage while improving inference speed by 15%+ in resource-constrained environments. Available via PyPI for LLM optimization.

Researchers introduce a hierarchical benchmark for LLM interactive reasoning using executable games. It evaluates active evidence acquisition across query, integration, and update phases. Validated on Codeforces, showing 20% accuracy improvement over traditional methods.

TIGER introduces graph-based evidence routing to reduce hallucinations in multimodal generation. It excels in multi-agent strategic games by handling future dependencies and rule violations. Shows 35% fact accuracy improvement with open-source code.

Researchers propose Universal Quantum Transformer to solve classical neural networks’ struggle with discrete mathematical rules. It achieves exact modular arithmetic and non-commutative algebra with 90% fewer parameters. Shows 100x speedup in cryptography simulations.

Stanford’s TIGER uses graph-based evidence routing to repair factual errors in multimodal generation. It dynamically validates outputs against input environments, showing 50% improvement in fact consistency for image-text tasks. Code now open-sourced.

datasette-agent-micropython 0.1a0 enables safe Python code generation and execution in MicroPython environments. Tests show GPT-5.5 failed to breach the sandbox, demonstrating strong security.

micropython-wasm 0.1a1 fixes sandboxing limitations for dataset building. The update enhances Python execution in WebAssembly environments, supporting secure multi-tenant deployments. Ideal for AI agent development in embedded systems with strict isolation needs.

A pasted file editor for Codex desktop, inspired by Claude’s paste detection, automatically converts large text pastes into file attachments. Handles 100MB+ files, improving workflows for content creators.

Developer Simon Willison spotted California Brown Pelicans during Microsoft Build, sparking discussion on AI and conservation. The image highlights San Francisco Bay’s biodiversity, while Microsoft explores AI applications in ecological monitoring.