OpenAI has launched the "Your Year with ChatGPT" year-in-review feature, providing users in select regions like the US, UK, Canada, Australia, and New Zealand with an annual summary akin to Spotify Wrapped. It displays users' yearly interaction data with ChatGPT, including active days, frequently used features, and conversation styles, and generates personalized badges, poems, and pixel art. It is available to free, Plus, and Pro users (requires chat history and memory features enabled). This feature aims to enhance user engagement and product experience, reflecting AI products' increasing focus on user participation and personalization.

Zhipu AI has officially open-sourced its new flagship large model, GLM-4.7. Ranking first among open-source models in the globally authoritative coding benchmark Code Arena, its overall performance surpasses GPT-5.2, excelling particularly in multilingual programming, reasoning, and agent tasks. GLM-4.7 supports a 128K context window and is compatible with mainstream inference frameworks. It has been widely adopted in industries like industrial manufacturing, finance, and internet, marking a breakthrough for domestic large models in core technology fields.

The MiniMax Agent platform has launched the open-source coding and agent model MiniMax M2.1. With 10 billion active parameters and high scores on multiple authoritative tests (e.g., SWE-multilingual, VIBE-bench), it surpasses mainstream closed-source models like Gemini3Pro and Claude 4.5Sonnet. Supporting multilingual programming and adaptable to various scenarios including web, Android, iOS, and backend, it demonstrates the strong capabilities of open-source models in real-world coding tasks.

Google DeepMind has launched the Gemma Scope2 open interpretability toolkit suite, covering the Gemma3 series models from 270 million to 27 billion parameters. Utilizing technologies like sparse autoencoders, it supports tracking internal activation features of models, helping AI safety and alignment teams analyze model behaviors such as "jailbreaking", hallucinations, and sycophancy. This advancement promotes research into the safety and controllability of large models.

ByteDance plans capital expenditures of $23 billion (approximately 160 billion yuan) for 2026, of which 85 billion yuan is allocated for purchasing AI processors. An initial trial purchase involves 20,000 Nvidia H200 chips. This aims to narrow the AI gap with international tech giants. Despite geopolitical influences, ByteDance is accelerating its AI computing power layout, pushing AI infrastructure development into a period of rapid expansion.

The Alibaba Qwen team has open-sourced the Qwen-Image-Layered image editing model. It supports automatically decomposing ordinary images into multiple RGBA layers, enabling layer operations similar to Photoshop, with support for recursive refinement and multi-layer editing. This significantly improves the image processing efficiency for designers and developers, promoting the application of AI in professional design fields.

At the AI DingTalk 1.1 new product launch event, DingTalk unveiled the world's first work intelligent operating system specifically designed for artificial intelligence: Agent OS. It supports multiple functions including AI Agent collaboration, AI hardware integration, and AI search/query. Simultaneously, over 20 AI products were released, marking the formal implementation of a "Human-AI Collaboration" work style and aiding in the construction of enterprise-level AI ecosystems.

In its latest security bulletin, OpenAI pointed out that AI browsers (like Atlas) struggle to completely defend against Prompt Injection attacks, where malicious webpage content can induce AI agents to perform unintended actions. Although multiple mitigation measures exist, Prompt Injection remains a persistent and unsolved security challenge as AI agent capabilities expand.

Google has released the Gemini 3 Flash model, positioned as lightweight and efficient, with significantly improved speed and latency. Its performance approaches that of Gemini 3 Pro but at a lower cost, making it suitable for real-time and large-scale application scenarios. The new architecture balances token efficiency and reliability, promoting the popularization and application deployment of AI models.

A team from Duke University has developed a new type of AI that can automatically extract concise, human-readable mathematical laws from extremely complex systems. It condenses thousands of variables into compact equations, significantly enhancing the ability to model and predict complex systems, bringing new breakthroughs for scientific research and engineering applications.