英伟达官方推理指南：投机解码如何让模型贴着硬件极限设计
9月2日，英伟达技术博客上线了一篇官方推理指南，主题是「用投机解码做AI模型协同设计」。芯片霸主罕见地以官方姿态，把「模型该怎么贴着硬件物理极限去设计」写成了规范。核心是一张6种主流推理加速方案的选型表，外加一条硬性约束公式——你的注意力分组数，直接决定草稿该猜几个token。 投机解码在干嘛 大模型逐字吐字，每吐一个词，几百GB参数就得在显存里完整跑一圈，大部分算力都在等内存搬数据。投机解码的解法很粗 …
NVIDIA's Speculative Decoding Guide: How Model Design Meets the Hardware Limit
On September 2, NVIDIA quietly dropped an official inference guide on speculative decoding and model–hardware co-design. It's rare for a chip vendor to publish a formal spec for how a model should be architected against the physical limits of its silicon. The centerpiece is a selection table of the six mainstream acceleration schemes, plus a surprisingly hard constraint equation that ties your att
李飞飞 World Labs 发布 Atlas：一个模型统一图像、视频与 3D 世界构建
李飞飞 World Labs 发布 Atlas：一个多模态世界模型，从头预训练统一处理图像、视频与 3D，把世界生成、空间重建与物理模拟收进同一个模型。
World Labs Atlas: One Model for Images, Video, and 3D World Building
World Labs Atlas is a single multimodal model for image, video, and 3D, spanning world generation, spatial reconstruction, and simulation.
Three Futures for the US-China AI Standoff: Open Weights, Token Commoditization, and a New Digital Border
The US-China AI race is scored across three ledgers: company, nation, and individual. Both sides win in structurally different ways.
AI Daily Briefing – 2026-09-03: Qwen3.8-Max tops coding; Claude Fable 5.1 debuts
Three forces reshaped the AI landscape overnight: Alibaba pushed its flagship model to the top of the coding charts, Anthropic shipped a cheaper and stronger Claude, and Ant Group won a top industrial prize for a data-engineering breakthrough that slashes pretraining costs. Top 3 Highlights 1. Alibaba's Qwen3.8-Max tops the world in front-end coding Alibaba updated its flagship Qwen3.8-Max and now
AI 早报-2026年9月3日
每日精选全球 AI 资讯，快速掌握行业趋势 🔥 今日亮点（Top 3） 1. 还在为大模型洗数据熬夜？蚂蚁拿下VLDB工业最佳论文，一套宽表搞定35PB语料，效率狂飙5.6倍 核心洞察：AI 行业持续火热 来源：https://www.qbitai.com/2026/09/483104.html 2. 阿里更新旗舰模型Qwen3.8-Max，前端编程能力跃居全球第一 核心洞察：AI 行业持续火热 来源：https://www.qbitai.com/2026/09/483101.html 3. 前字节强化学习专家孙鹏博士加盟星尘智能，
OpenAI Astra Recurrent Depth: Trade Compute for Brainpower
A leaked OpenAI Astra architecture reuses one layer in loops, trading inference compute for parameters, while making reasoning opaque.
OpenAI Astra 换上「循环深度」推理架构：用计算深度换参数规模
Astra 被曝采用循环深度架构：复用同一层网络多次循环，用计算时间换参数量；思考正从 token 空间搬回模型内部，代价是推理过程变得不可读。
Anthropic Unveils Claude Fable 5.1 and Mythos 5.1: One Base Model, Two Guardrails
Anthropic unveiled Claude Fable 5.1 and Mythos 5.1: one base model, two safety layers, 52.6% on a science benchmark, cache reads 75% cheaper.
Anthropic 发布 Claude Fable 5.1 与 Mythos 5.1：同一基座、两套护栏，科学智能体得分翻倍
Anthropic 9 月 1 日发布 Claude Fable 5.1 与 Mythos 5.1：同一基座模型、两套安全防护；科学智能体基准得分 52.6% 较前代翻倍，缓存读取降价 75%，典型负载成本降低约 25%。