英伟达推出机器人“新大脑”Jetson Thor，算力飙升7.5倍

简报：

相关链接：

Nvidia Unveils Robot's 'New Brain' Jetson Thor, Compute Power Skyrockets 7.5 Times

Brief:

Nvidia officially released its next-generation robot chip, Jetson Thor, featuring the Blackwell GPU architecture and designed specifically for physical AI agents like humanoid robots. 🤖
Compared to its predecessor, Jetson Orin, Jetson Thor boasts a 7.5x increase in AI compute power, reaching 2070 FP4 TFLOPS, a 3.5x improvement in energy efficiency, and comes with 128GB of memory. 🚀
The chip supports real-time generative AI model inference at the edge and is equipped with the complete NVIDIA AI software stack and Isaac simulation platform to reduce reliance on cloud services.
The Jetson Thor developer kit starts at $3,499 USD (approximately 25,000 RMB), while the production-ready module is priced at$ 2,999 USD (approximately 21,400 RMB).
Numerous companies, both domestic and international, have announced their adoption, including China's Unitree Robotics, Galaxy Universal, and international players like Amazon Robotics, Boston Dynamics, and Figure. 💡

简报：

苹果研究团队提出一种名为“基于清单反馈的强化学习”（RLCF）的AI训练新方法，旨在提升大语言模型执行复杂指令的能力。
该方法用详细的任务清单和0-100的精确评分机制，替代了传统依赖人工“点赞”或“点踩”进行反馈的RLHF方法。
实验结果表明，RLCF是唯一在FollowBench、InFoBench等五个常用评测基准上均取得性能提升的方案，在某些任务上提升幅度最高达8.2%。
技术上，该方法使用一个更强大的“教师模型”（Qwen2.5-72B-Instruct）为用户指令生成检查清单，并对“学生模型”的回答进行评分，以形成奖励信号进行优化。
研究人员也指出该方法的局限性，包括需要依赖更强大的模型作为评估标准，且该方法并非为安全对齐而设计。

相关链接：

Briefing:

Apple's research team has proposed a new AI training method called "Reinforcement Learning from Checklist Feedback" (RLCF) 🧠, aimed at improving the ability of large language models to execute complex instructions.
This method replaces the traditional RLHF approach, which relies on human "likes" or "dislikes" for feedback, with detailed task checklists and a precise 0-100 scoring mechanism.
Experimental results show that RLCF is the only solution that achieved performance improvements across all five commonly used evaluation benchmarks, including FollowBench and InFoBench, with an improvement of up to 8.2% on some tasks. ✨
Technically, the method uses a more powerful "teacher model" (Qwen2.5-72B-Instruct) to generate checklists for user instructions and then scores the "student model's" responses to form reward signals for optimization.
Researchers also pointed out the method's limitations, including the reliance on a more powerful model as an evaluation standard, and that the method is not designed for safety alignment. 🧐

简报：

相关链接：

Brief:

Microsoft Research recently open-sourced its latest audio model, VibeVoice-1.5B.
The model can synthesize up to 90 minutes of speech at once and supports up to four speakers conversing simultaneously.
VibeVoice-1.5B achieves a 3200x compression ratio for 24kHz raw audio while maintaining high-fidelity results. 🔊
The model employs an innovative dual-tokenizer architecture (acoustic and semantic) to address the issue of timbre-semantic mismatch.

简报：

相关链接：

Briefing:

Google has officially launched its latest text-to-image model, Imagen 4, which is now available to developers via the Gemini API (paid preview) and Google AI Studio (limited-time free trial). 🚀
The model has also been integrated into the Gemini app, Vertex AI platform, and the Google Workspace suite, including Google Slides, Vids, and Docs. 🎨

Related links:

Entity /ˈen·tə·ti/

n. 实体

▶ "Six subsidiaries of Inspur Group were added to the U.S. entity list."

[例句] 浪潮集团六家子公司被列入美国实体清单。

◼ 词根分析

ent-

存在

-ity

性质

◼ 衍生词

entities (n.) 实体（复数）