GitHub announced that Copilot will transition from a fixed subscription model to a token-based usage pricing system (AI credit system) starting June 1, with detailed tracking of token consumption in inputs, outputs, and context caching. This move has triggered strong backlash from developers, with some users estimating their monthly costs could rise from dozens of dollars to hundreds or even thousands. Small teams and independent developers are expected to be the most affected. Some argue that costs remain manageable for reasonable usage, with only those heavily relying on AI for 'vibe coding' facing extremely high bills. This shift aligns with competitors like Anthropic's Claude Code adopting similar pricing models, reflecting an industry-wide trend of AI tools moving from subsidized user acquisition toward monetization.

On May 31, the US Department of Commerce issued new guidance blocking companies from exporting the most advanced AI chips—including NVIDIA's Rubin and Blackwell processors and AMD's MI350x—to entities headquartered in China but operating overseas. The move aims to close a regulatory loophole dating back to May 2025, under which Chinese firms had been acquiring large volumes of high-end AI chips through overseas subsidiaries in countries like Malaysia, reportedly amounting to hundreds of thousands of units. The new rules impose license requirements for advanced chip exports but do not mandate data centers to cease using already-deployed equipment. NVIDIA and AMD have not yet responded.

Google has officially launched Gemini Spark, its first fully functional background AI agent, now available to Google AI Ultra subscribers in the United States. The agent marks a shift from conversational AI to delegated digital task management, capable of running continuously 24/7 in the background to autonomously handle personal information management, emails, meeting scheduling, and interactions with third-party services. Users transition from direct operators to high-level reviewers. The launch raises concerns about data privacy and algorithmic trust. The limited regional release aims to monitor user behavioral responses to prolonged AI presence, with future plans to expand the agent to desktop operating systems.

An OpenAI AI model has successfully solved a long-standing combinatorial mathematics problem that remained unsolved for nearly 80 years. The proof has been independently verified by professional mathematicians and confirmed to be mathematically rigorous. The model used step-by-step reasoning to explore multiple mathematical possibilities and ultimately constructed a proof path previously undiscovered by human experts. Experts emphasize this does not mean AI will replace mathematicians, but rather shows that advanced reasoning models can become powerful tools for accelerating scientific discovery. This achievement marks a significant step forward in AI-assisted original scientific research.

AGIBOT Finch has released τ0-WM, a unified video-action world model with 5 billion parameters, pre-trained on approximately 27,300 hours of heterogeneous data, including 17,800 hours of real-world robot teleoperation data. The model jointly predicts future visual latent variables and continuous action blocks through a shared video diffusion backbone, supporting a 'propose-evaluate-revise' test-time computation mechanism that allows robots to simulate outcomes before executing actions. Experiments show τ0-WM achieves state-of-the-art average success rates on unseen tasks, particularly excelling in precision-sensitive operations, advancing robotics from reactive to predictive control. The model and code are open-sourced.

Microsoft has open-sourced the SkillOpt text-space optimization framework, treating Agent skill documents as trainable external weights. Using a four-step loop analogous to deep learning training—Rollout (forward pass), Reflect (backward pass), Edit (parameter update), and Gate (validation gating)—the framework enables systematic self-evolution of skill documents. It introduces text learning rates and a rejected-edit buffer to prevent catastrophic forgetting. SkillOpt achieved best or tied-best results across 52 evaluation combinations spanning 7 models, 6 benchmarks, and 3 execution environments. Trained skill documents are transferable across models and environments, requiring only a single best_skill.md file for deployment with zero inference overhead. The project gained 3,300 stars in one week.

Researchers at Permiso Security discovered a new attack technique called ChatGPhish, exploiting ChatGPT's mechanism of directly rendering Markdown elements when summarizing web content, allowing malicious content to be injected into AI responses. Attackers can hide phishing links, forge security warnings, or embed external QR codes within web pages; when users request a ChatGPT summary of such pages, the malicious content appears within the trusted interface. Additionally, embedded external images are automatically loaded, potentially leaking sensitive metadata such as user IP addresses. Another security researcher also released a tool called LLMReaper, demonstrating that malicious browser extensions can steal conversation data from ChatGPT, Claude, and Gemini.

According to Axios, the first Windows PCs powered by Nvidia processors are expected to debut at Computex 2026 and Microsoft's Build developer conference, marking Nvidia's first major entry into the mainstream PC market. Microsoft also plans to launch a new platform enabling AI agents to run tasks locally on Windows devices, possibly introducing a new app called Scout to integrate Copilot tools, and adding an intelligent agent layer named Autopilot to support automated task execution. This move is seen as a key part of Microsoft's shift toward Arm-based chips to improve battery efficiency and compete with Apple's M5-chip MacBooks.