Group 1 - Nvidia and AMD have agreed to pay 15% of their revenue from specific AI chips sold in China to the U.S. government in exchange for export licenses [1] - Nvidia will pay 15% of its revenue from H20 chips, while AMD will do the same for MI308 chips [1] - The U.S. Department of Commerce has begun issuing export licenses for these products, but the Trump administration has not yet decided how to utilize the funds collected [1] Group 2 - OpenAI achieved a gold medal in the AI category at the 2025 International Olympiad in Informatics, ranking first among AI participants and only behind five human competitors [2] - OpenAI's performance improved significantly from the 49th percentile last year to the 98th percentile this year, using a general reasoning model without specialized training for the competition [2] - The model used by OpenAI is the same as the one that won a gold medal at the International Mathematical Olympiad, showcasing its strong general reasoning capabilities [2] Group 3 - Zhizhu released and open-sourced the GLM-4.5V model, which has 106 billion parameters and achieved state-of-the-art performance in 41 multimodal benchmarks [3] - The model outperformed 99% of human players in image recognition and reasoning tests, achieving a notable rank in a global scoring competition [3] - It employs a three-stage strategy for training and supports long-context multimodal inputs, with low API usage costs [3] Group 4 - Kunlun Wanwei launched the Matrix-3D model for generating high-quality panoramic videos from single images, enabling immersive 3D space exploration [4] - The model boasts advantages such as global scene consistency, large generation range, high controllability, strong generalization ability, and fast generation speed [4] - A dataset containing 116,000 panoramic videos and 22 million frames was created to support the model's training [4] Group 5 - Tencent introduced the mixed Yuan Large-Vision model, which has 52 billion active parameters and enhances multimodal understanding capabilities [5] - The model scored 1256 points on the international LMArena Vision leaderboard, ranking first among domestic models and comparable to GPT-4.5 and Claude-4-Sonnet [5] - It consists of three core modules and utilizes a large dataset for training [5] Group 6 - GitHub will no longer operate independently and will be integrated into Microsoft's newly established CoreAI group [7] - The integration will be overseen by multiple Microsoft executives, with a focus on transforming GitHub into a core component of Microsoft's AI strategy [7] - The goal is to develop GitHub into an "AI agent factory" [7] Group 7 - SenseTime launched the AI tool Seko, which automates the video production process based on user descriptions [8] - Seko integrates various models to ensure consistency in character portrayal, scene materials, and camera movements [8] - The tool offers a visual editing experience and plans to introduce advanced features in the future [8] Group 8 - Apple is gradually revamping Siri, with a new architecture set to launch by late 2025 or early 2026 [9] - The new Siri will enhance inter-application communication and support continuous dialogue [9] - Apple is conducting extensive internal testing with strategic partners to ensure security and reliability [9] Group 9 - Periodic Labs, co-founded by former OpenAI and Google DeepMind leaders, aims to create a "ChatGPT for materials science" and has secured $200 million in funding [10] - The startup achieved a pre-money valuation of $1 billion shortly after its establishment [10] - The funding will be used to develop AI for discovering and analyzing new compounds [10] Group 10 - GPT-5 demonstrated significantly lower token consumption compared to Claude Opus 4.1 in algorithmic tasks, saving approximately 90% in overall token usage [12] - Claude Opus 4.1 excelled in web development tasks but at a higher token cost [12] - The cost comparison shows GPT-5 completing tasks at about $3.50, while Claude Opus 4.1 costs around $7.58 [12]