2 Sep 2026. NVIDIA has released NIM LLMs Benchmarking 2.0.0, a new framework for defining and measuring LLM inference performance. The new version focuses on the prefill and decode stages of model execution. Key metrics include TTFT (Time to First Token) and TPT (Throughput Per Token). In education, MIT's AI Use Committee released a final report finding that AI is undermining traditional assessment methods like problem sets and homework, as it can generate plausible answers for any written assignment. Meanwhile, in industry, Zhipu AI announced completing approximately $5 billion in financing to develop the next generation of GLM foundation models. The funding will be used for research into a fully self-training system, large-scale training, production inference, and adapting to domestic chips. Also, Microsoft CEO Satya Nadella stated that AI development should proceed under core principles and that the company welcomes independent auditing. He announced that Microsoft's AI behavior guidelines will be published tomorrow for public comment.
For developers, these are the strongest open-weight search agents in their class, built on Qwen models and leading benchmarks among open-weight models in their size classes.
This system is a significant advancement in China's domestic AI hardware, offering a more cost-effective alternative to traditional GPU-only setups for large model inference.
For users needing to control multiple robot types, this model offers a unified solution without the need for separate training for each robot type.
This matters because a top executive from a major AI company is publicly advocating for a slower pace in the industry, which could influence the direction of AI development globally.
For users interested in quantum computing hardware, the release of a new generation of a large-scale photonic quantum computer with improved architecture and lower latency is a significant development.
ElevenLabs 推出了 AI 音乐生成器 Music v2.5，能生成更受听众偏好的音乐，通过应用和 API 提供免费与专业版。
vivo 要发新系统了，里面有个叫蓝心小 V 的东西，能让你说句话就帮你办事，比如旅行攻略，挺方便的。
This matters because it highlights growing concerns from top AI researchers about the pace of AI development and the need for stronger safety measures, a topic of interest to anyone following the AI industry.
This matters because it addresses the fragmentation of China's AI computing resources, which is a key challenge for the development of the AI industry.
微软 CEO 纳德拉发了条关于 AI 发展理念的内容，说 AI 需要加速发展，企业要掌握自己的知识，微软明日会公布《行为准则》，挺具体的。
This report from a leading US university provides a detailed analysis of AI's impact on education, which is relevant for educators and policymakers in the region.
This is a significant new architecture proposal for large language models, potentially offering improved efficiency and performance.
Gary Marcus 分析了 OpenAI 联合创始人 Dario Amodei 的观点，说出了他同意的部分和省略的部分，值得看看。
Gary Marcus 对 Dario Amodei 论文的详细分析，有不同视角和政策建议，值得一读。
This is important for anyone working with LLMs as it provides a standardized way to measure and compare inference performance, which is crucial for optimizing applications.
For developers, building a custom agent harness provides more control over automated prompt engineering and dynamic workflows, which models are not great at.
朋友，有个叫Agency的本地AI小助手，能帮你找有用的事做，比如提醒你忘了的事，还能帮你提前解决麻烦，不用你一直发指令，挺方便的。
阿里开源的AI代码审查工具，专门做代码审查，支持自定义配置和Delegation Mode，能直接交给其他agent处理，适合需要高效审查代码的场景。
