AI News 2025-12-25
AI Daily Brief
Summary
Microsoft clarified it has no plans to completely rewrite Windows using Rust and AI.
JD Logistics' UK "Zhi Lang" warehouse went online, with efficiency improving 4×.
Yuanli Lingji's GeoVLA framework solved the robot "spatial blindness" problem.
ByteDance's Seed Prover1.5 model won an IMO gold medal.
X platform's AI image editing feature sparked creator copyright concerns.
Domestic large models have high Chinese data proportions, enhancing cultural understanding.
Zhipu and MiniMax compete for the first large-model public listing.
Anthropic open-sourced Claude Skills library to define AI workflows.
Alibaba released Qwen-Image-Edit-2511 model to improve character consistency.
OpenAI tested new "Skills" feature codenamed "Hazelnut," targeting Claude and planning a 2026 release.
Today’s AI News
Microsoft officially clarified rumors about using artificial intelligence (AI) and Rust language to completely rewrite Windows operating system: Microsoft stated there is currently no plan to completely reconstruct Windows 11 and future operating systems using Rust and AI. Engineers mentioned in relevant job postings also clarified that while company is developing cross-language migration technologies, Windows will not be completely rewritten in Rust nor will it be entirely AI-driven. Microsoft acknowledged it is using Rust to refactor some kernel components to improve security, but remains cautious about large-scale use of AI for automatically generating or modifying low-level system code.
On December 25, JD Logistics announced the official launch of its first “Zhi Lang” warehouse in the UK: The automated warehouse covers over 3,000 square meters and is equipped with nearly 200 “Zhi Lang robots” independently developed by JD Logistics. Thanks to high automation, its picking and outbound efficiency improved approximately 4× compared to traditional models. The UK Zhi Lang warehouse launch is an important node in JD Logistics’s “Global Weaving Plan,” working with its European online retail platform Joybuy pilot operation to complete a global supply chain network.
Yuanli Lingji’s research team launched a new Vision-Language-Action (VLA) framework called GeoVLA, aiming to solve the “spatial blindness” problem of existing models that rely on 2D images. The framework uses an innovative dual-stream architecture, introducing point cloud embedding networks and spatial-aware action experts, giving robots true three-dimensional geometric perception capabilities. Experiments show GeoVLA performs excellently across multiple benchmarks, especially high success rates when handling complex objects and viewpoint changes, demonstrating strong adaptability.
ByteDance released mathematical reasoning model Seed Prover1.5, winning a gold medal in the IMO competition: The model uses large-scale reinforcement learning and solved all five problems from IMO 2025 within 16.5 hours, achieving gold medal standards. Its core innovations include using formal languages like Lean for verifiable proofs in the “Agentic Prover,” and the “Sketch Model” that simulates human problem-solving approach by drafting first before formalizing. This achievement marks an important breakthrough for AI in mathematical reasoning.
X platform launched AI-based image editing feature, allowing users to easily edit images through prompts, aiming to improve user experience. However, many creators worry this will increase risk of original content being stolen or maliciously tampered with, affecting creative safety and originality. Some creators have considered or already switched to other platforms, with community calls for the platform to strengthen copyright protection mechanisms.
Training data in mainstream domestic large models currently contains over 60% Chinese content, some as high as 80%: This enables AI to more accurately understand Chinese user needs and demonstrates for the first time the ability to deeply analyze culture-specific concepts like “shang huo” (internal heat). Industry is accelerating construction of high-quality Chinese datasets; for example, China Mobile has built a general-purpose dataset exceeding 3,500 TB. Experts point out this relates to cultural sovereignty and digital civilization discourse power, though challenges like data silos and lack of standards remain.
Zhipu and MiniMax compete for the “first large-model public company” title: Both companies face “bleeding IPO” pressure but have very different business models: Zhipu focuses on domestic market, monetizing through MaaS (Model-as-a-Service) model; MiniMax targets global markets, profiting through AI-native product subscriptions. Despite high revenue growth, both have huge losses and relatively low market share, facing giant “black hole effect” and fierce competition.
Anthropic open-sourced “Claude Skills Library,” defining standardized AI workflows: The library contains 9 major categories and over 50 professional skills covering document processing, development tools, data analysis, and full scenarios, aiming to transform Claude from a conversational assistant to a professional executor. Users can invoke these skills via web interface, local deployment, or API, and customize them under Apache 2.0 license, driving AI toward “collaborative execution.”
Alibaba released image editing model Qwen-Image-Edit-2511, significantly improving character consistency: The model solves common AI retouching problems like “face distortion” or “identity loss,” able to precisely maintain character facial features during editing operations like changing clothing or adjusting background. The new model also enhances capabilities in lighting control, industrial design, and more, while integrating LoRA technology. It has been open-sourced under Apache 2.0 license with free experience.
OpenAI is secretly testing a new feature codenamed “Hazelnut” called “Skills,” seen as a strong response to competitor Claude’s similarly named feature: This marks ChatGPT’s transition from “custom GPT” model to a “folder-instruction-based” Skills mode, emphasizing teaching AI specific capabilities, workflows, and domain knowledge. The new Skills system borrows from Claude model with four core characteristics: composability (skills support stacking and system can automatically coordinate multiple skills working together), high portability (unified format, build once and use across ChatGPT web version, desktop client, and API), extreme efficiency (on-demand loading to avoid unnecessary context window consumption), and strong execution capability (supports writing and running executable code). Additionally, the new feature will introduce “slash command” interaction for improved efficiency and plans to launch a “skills editor” allowing users to one-click convert existing custom GPTs to “Skills.” Market widely expects the feature to officially launch around January 2026.