![[640e10cf-5b22-4f70-a021-e5f2f55d3719.mp3]]

OpenAI通用AI模型首夺信息学奥赛金牌，位列总榜第六

简报：

• 在2025年国际信息学奥林匹克竞赛（IOI）中，OpenAI的AI模型获得金牌，在所有参赛AI中排名第一，总榜位列第六。
• 该模型的成绩超过了98%的参赛者，相较于去年仅取得49百分位的成绩有显著提升。
• 据OpenAI研究员称，参赛模型为通用的“模型集群”，并未针对IOI赛事进行专门训练，其参赛流程与人类选手一致。
• 此模型此前也曾在国际数学奥林匹克竞赛（IMO）中获得金牌，并在AtCoder编程竞赛中取得第二名。

相关链接：

• 超 98% 参赛者：OpenAI 神秘 AI 模型首次斩获信息学奥赛 IOI 2025 金牌

OpenAI's General AI Model Wins First Informatics Olympiad Gold Medal, Ranks Sixth Overall 🥇

Briefing:

• OpenAI's AI model secured a gold medal at the 2025 International Olympiad in Informatics (IOI), ranking first among all participating AIs and sixth overall.
• The model's performance surpassed 98% of contestants, marking a significant improvement from last year's 49th percentile achievement. 🚀
• According to OpenAI researchers, the competing model is a general "model cluster" that was not specifically trained for the IOI event, and its participation process was identical to that of human contestants.
• This model previously earned a gold medal in the International Mathematical Olympiad (IMO) and secured second place in the AtCoder programming contest. 🧠

Related Link:

• OpenAI's Mysterious AI Model Wins First IOI 2025 Gold Medal, Outperforming Over 98% of Contestants

OpenAI修改自家基准，GPT-5编程分数真实性存疑

简报：

• OpenAI在GPT-5的编程能力测试中，被指使用了修改版的测试标准，从其联合提出的SWE-bench Verified基准（共500个问题）中自行省略了23个问题。
• OpenAI官方解释称，省略这些问题的原因是其解决方案无法在公司内部的基础设施上运行，同样的操作在发布GPT-4.1时也曾发生。
• 关键在于，SWE-bench Verified这个基准测试集本身就是OpenAI为实现更准确评估，而与原作者合作提炼的。
• 竞争对手Anthropic已在其报告中指出，其Claude模型的分数基于完整的500个问题，而OpenAI模型的分数基于477个问题的子集。
• 分析认为，如果将OpenAI未测试的23道题计为零分，GPT-5的得分实际上可能低于其竞争对手Claude Opus 4.1。

相关链接：

• OpenAI GPT-5 编程成绩有猫腻：自删 23 道测试题，关键基准还是自己提的

OpenAI Modifies Its Own Benchmark, GPT-5 Programming Score Authenticity Questioned

Brief:

• OpenAI is reportedly using a modified testing standard for GPT-5's programming capability assessment, having unilaterally omitted 23 questions from the co-developed SWE-bench Verified benchmark (500 questions total).
• OpenAI officially stated that the omission was due to the solutions being unable to run on its internal infrastructure, a practice also seen during the GPT-4.1 release. 🤷‍♀️
• Crucially, the SWE-bench Verified benchmark was itself refined by OpenAI in collaboration with the original authors to ensure more accurate evaluations.
• Rival Anthropic has highlighted in its report that its Claude model's scores are based on the complete 500 questions, whereas OpenAI's model's scores are derived from a 477-question subset. 🤔
• Analysis suggests that if the 23 un-tested questions by OpenAI were scored as zero, GPT-5's performance might actually be lower than that of its competitor, Claude Opus 4.1. 📉

Related Links:

• OpenAI GPT-5 Programming Scores Under Scrutiny: 23 Test Questions Self-Deleted, Key Benchmark Co-Developed

宇树将参赛首届人形机器人运动会，并为多支队伍提供硬件

简报：

• 宇树科技官方宣布，将参加于2025年8月14日至17日举行的首届世界人形机器人运动会。
• 宇树团队因赛程紧凑，自身仅会参与部分赛事项目。
• 除宇树团队外，还将有多支其他队伍使用宇树的机器人硬件并搭载其自研算法进行比赛。

相关链接：

• 宇树官方公众号：官宣参赛首届世界人形机器人运动会
• 36氪快讯
• AIBASE报道

Unitree to Compete in First Humanoid Robot Games, Providing Hardware to Multiple Teams 🚀

Brief:

• Unitree Robotics has officially announced its participation in the inaugural World Humanoid Robot Games, scheduled for August 14-17, 2025. 🤖
• Due to a packed schedule, the Unitree team itself will only compete in a selection of events.
• In addition to the Unitree team, several other teams will be competing using Unitree's robot hardware, powered by their self-developed algorithms. ✨

Related Links:

• Unitree Official WeChat Account: Official Announcement of Participation in the Inaugural World Humanoid Robot Games
• 36Kr Newsflash
• AIBASE Report

英特尔发布AI推理软件更新，Battlematrix平台性能最高提升80%

简报：

• 英特尔为其Battlematrix项目发布了首个重要软件更新 LLM Scaler v1.0，旨在优化多显卡AI推理性能。
• 该更新基于Linux平台，通过利用多GPU扩展和PCIe点对点数据传输，可实现最高80%的整体性能提升。
• 新版本针对vLLM框架进行了优化，在特定模型（如70B KPI模型）和条件下，性能提升可达4.2倍。
• 更新引入了分层在线量化以降低显存占用，并集成了XPU Manager工具，支持电源管理、固件升级等企业级运维功能。
• Battlematrix推理平台最多可支持8张英特尔锐炫Pro系列显卡，能运行高达1500亿参数的中等规模AI模型。

相关链接：

• 英特尔升级多显卡AI推理，Battlematrix整体性能提升80%

Intel Releases AI Inference Software Update, Battlematrix Platform Performance Boosted by Up to 80%

Brief:

• Intel has released LLM Scaler v1.0, the first major software update for its Battlematrix project, aimed at optimizing multi-GPU AI inference performance. 🚀📈
• Based on the Linux platform, this update leverages multi-GPU scaling and PCIe peer-to-peer data transfer, achieving up to an 80% overall performance improvement.
• The new version is optimized for the vLLM framework, delivering up to a 4.2x performance boost under specific models (e.g., 70B KPI models) and conditions.
• The update introduces layered online quantization to reduce VRAM footprint and integrates the XPU Manager tool, supporting enterprise-grade O&M functions like power management and firmware upgrades. 🛠️
• The Battlematrix inference platform supports up to 8 Intel Arc Pro series graphics cards, capable of running medium-sized AI models with up to 150 billion parameters.

Related Links:

• Intel Upgrades Multi-GPU AI Inference, Battlematrix Overall Performance Boosts by 80%