AI 情报 AI Briefings

在边缘侧运行推理和智能体 AI 一直很棘手。直到最近，具备多步推理能力的模型仍然太大，难以部署…… Running reasoning and agentic AI at the edge has been harder than it needs to be. Until recently, models capable of multi-step reasoning were too large to run...

多轮智能体的失败方式是单轮评估捕捉不到的：一个早期错误会污染后续每一轮。本文介绍智能体评估指标（AEM），一种可分解、逐轮粒度的度量方法…… Multi-turn agents fail in ways single-turn evaluation misses: one early mistake corrupts every later turn. This post introduces the Agent Evaluation Metric (AEM), a decomposable, turn-level way to mea...

TwelveLabs Marengo Embed 3.0 已在 Amazon Bedrock 知识库中作为嵌入模型正式可用，为视频、图像和音频内容带来全托管的自然语言搜索…… TwelveLabs Marengo Embed 3.0 is now generally available as an embedding model in Amazon Bedrock Knowledge Bases, bringing fully managed natural language search to video, image, and audio content. This...

GBDL 让你用单个 Markdown+YAML 文件定义并保存完整的多智能体 Grok Bot 配置。 GBDL lets you define and save a complete multi-agent Grok Bot setup in a single Markdown+YAML file.

Grok CLI 是一个基于 Grok 的智能体编程 harness，面向终端工作流。 Grok CLI is a Grok-native agentic coding harness for terminal workflows.

每个 NVIDIA CUDA Toolkit 版本都会带来新功能和性能改进，帮助开发者更好地利用 NVIDIA GPU 及整个软件生态…… Every NVIDIA CUDA Toolkit release adds functionality and performance improvements that help developers get more from NVIDIA GPUs and the broader NVIDIA software...

Stability AI 推出 Brand Studio，一个由品牌资产驱动的端到端创意生产平台。 We’re excited to introduce Brand Studio by Stability AI, the end-to-end creative production platform powered by your brand.

Meta 推出个人 AI 智能体应用 Muse，已在 App Store 上线。 Meta launches Muse, a personal AI agent app now available on the App Store.

Stable Audio 3.0 新增两种使用方式：DAW 插件，以及 StableAudio.com 上更进阶的生成体验。 We’re sharing two new ways for you to use Stable Audio 3.0. We now have a DAW plugin and a more advanced generation experience on StableAudio.com.

Meta 新 AI 智能体 Muse 上线后，英国乐队 Muse 失去了自己的社交媒体账号名，引发社区热议。 British band Muse lost its social media handles to Meta's new AI agent Muse, sparking community debate.

KV 缓存暴降 437 倍，Agent 推理成本大洗牌。 KV cache slashed by 437x — a major reshuffle of agent inference costs.

9月8日至12日，计算机视觉顶会 ECCV 2026 在瑞典马尔默举行。MARS2 Workshop（大模型时代的多模态推理与慢思考）由钛动科技牵头发起并成功举办。 At ECCV 2026 in Malmö, Sweden (September 8-12), the MARS2 Workshop on multimodal reasoning and slow thinking in the large model era was initiated and hosted by a Chinese company.

轻量版 LingBot-World 2.0，单卡即可实时运行。 The lightweight LingBot-World 2.0 runs in real time on a single GPU.

模型可以开源，部署经验不能。 Models can be open-sourced; deployment experience cannot.

AI 推理时代已经到来。想象一个医疗系统实时分析数百万数据点以加速救命研究，或一个智能助手即时响应…… The era of AI inference has arrived. Imagine a healthcare system analyzing millions of data points in real time to accelerate life-saving medical research, or an intelligent assistant instantly resolv...

Anthropic 周四发布的新报告指控中国 AI 公司持续进行蒸馏攻击，随着行业竞争加剧，近几个月愈演愈烈。 A new report released Thursday by Anthropic alleges persistent distillation attacks by China-based AI companies, which have escalated in recent months as competition in the space has intensified.

2026 年 7 月 22 日，弗吉尼亚州阿什本——全球最大数据中心集群的心脏——一条输电线路故障在数秒内让电网失去超过 3GW 负载…… On July 22, 2026, a transmission line fault in Ashburn, Virginia—the heart of the world’s largest data center cluster—knocked more than 3 gigawatts of load off the grid in seconds. And it wasn&#...

Meta 的新应用 Muse 起步慢于该公司其他应用（如 Meta AI 或 Threads），但仍冲上了美国应用榜第二。 Meta's newest app Muse is off to a slower start than the company's other apps, like Meta AI or Threads.

自主智能体在"深度研究"上进展显著——通过迭代浏览开放网络综合信息，但真实世界的任务往往还需要查询数据库…… arXiv:2609.09410v1 Announce Type: new Abstract: While autonomous agents have made significant strides in "deep research" by iteratively navigating the open web to synthesize information, real-world pr...

现有的自主 AI 科学家基准只评估最终产出（代码、假设或论文），却丢弃了推理过程…… arXiv:2609.09203v1 Announce Type: new Abstract: Existing benchmarks for autonomous AI scientists evaluate only final outputs---generated code, hypotheses, or papers---yet discard the reasoning process...

认知数字孪生（CDT）的快速定制调试是可重构制造的一大挑战，传统数字孪生…… arXiv:2609.09503v1 Announce Type: cross Abstract: Rapid bespoke commissioning of the Cognitive Digital Twin (CDT) is a major challenge in reconfigurable manufacturing. Traditional digital twin (DT) co...

投机解码是加速 LLM 推理的关键，但加速效果很脆弱：草稿模型通常针对狭窄的分布训练…… arXiv:2609.09338v1 Announce Type: new Abstract: Speculative decoding is critical for accelerating LLM inference. However, the speedup is fragile: drafters are typically trained against a narrow distri...

世界动作模型（WAM）将预测性世界建模与动作生成结合，让预期的未来状态引导智能体行为…… arXiv:2609.09418v1 Announce Type: new Abstract: World Action Models (WAMs) couple predictive world modeling with action generation, allowing anticipated future states to guide agent behavior. Although...

LLM 只有见过大量真实标注样本后才能泛化到某个领域，而多数领域缺数据。本文研究一种廉价制造数据的方法…… arXiv:2609.09163v1 Announce Type: new Abstract: LLMs generalize across a domain only after seeing many real, labeled examples, which most domains lack. We study a way to manufacture it cheaply. When a...

从口语对话中推断说话人关系是迈向社交感知语音理解的重要一步，但这一任务仍然…… arXiv:2609.09628v1 Announce Type: new Abstract: Inferring speaker relationships from spoken conversations is an important step towards socially aware speech understanding. However, this task remains u...

FPGA 能实现高效的神经网络推理，但大多数部署流程要么只加速乘加运算…… arXiv:2609.09254v1 Announce Type: new Abstract: Field-programmable gate arrays (FPGAs) enable efficient neural-network inference, but most deployment flows either accelerate multiply-accumulate operat...

仅靠一个视觉语言模型智能体即可驱动机器人完成操作任务。 A single VLM agent is enough to operate robots.

面向金融推理的数据中心后训练方法：数据挖掘、蒸馏与可验证学习。 A data-centric post-training recipe for financial reasoning: mining, distillation, and verifiable learning.

楼主在树莓派上跑智能体已经得心应手，现在想让它们访问互联网，向社区求方案…… Hey folks! I'm always a little bit late to the party, but learning nontheless. After I'm comfortable running agents on my Pi I'm now in need of them to get access to the world wide web and wanted to a...

楼主可能创下了 Apple Silicon 上 Qwen3.8-Flash-Next 内存受限推理的纪录——只需约 21GB 内存分配。他发布了名为 Cherenkov 的推理引擎…… I believe I may currently hold the record for memory constrained inference for Qwen3.8–Flash-Next on Apple Silicon — needing only about 21GB of allocations. Introducing Cherenkov, an inference engine ...

IBM 发布 SOTA 级 Granite 时间序列基础模型 PatchTST-FM-r2，许可证对商业使用友好。 IBM released the SOTA Granite time series foundation model PatchTST-FM-r2 under a commercial-friendly license.

用 Gradio Workflow 重新实现 AUTOMATIC1111 的 Stable Diffusion WebUI 体验。 Rebuilding the AUTOMATIC1111 Stable Diffusion WebUI experience with Gradio Workflow.