GPTProto
⌘K

Meta / Llama

70 picksNewest firstLatest article Sep 25, 2026, 5:15 AM
1 story
17:42 X：Rohan Paul (@rohanpaul_ai)AI score 50/100

Interesting blog

Interesting blog"Meta may be one of the very few companies that can monetize a personal agent through advertising without ever showing an ad inside the agent itself. "Meta acknowledges that Muse appears as your own activity when it browses, so a merchant site it visits for you might use that visit to show you an ad on Instagram. metaThat lets Meta keep ads out of Muse while still collecting commercial intent signals for its feeds, a loop few rivals can copy without both merchant tracking tags an
6 stories
15:34 X：Rohan Paul (@rohanpaul_ai)AI score 48/100

Meta's Muse is a threat to particular sellers far more than to consumption. The exposed categories are only about 3.7% of US spending, bu...

Meta's Muse is a threat to particular sellers far more than to consumption. The exposed categories are only about 3.7% of US spending, but that's roughly $830 billion a year, concentrated among streamers, publishers, gyms, carriers and insurers.Apollo ResearchEvery dollar Muse saves goes to the household, while Meta charges usage tiers from free to $100 a month. Meta says it's exploring commerce revenue, likely fees when Muse moves a household to a new carrier or insurer.Carriers and insurers wo
15:30 Canary Media：Data Centers 原创报道AI score 55/100

Despite big Texas solar purchase agreement, Meta still uses lots of gas

Despite big Texas solar purchase agreement, Meta still uses lots of gas Account Donate Clean energy journalism for a cooler tomorrow The new solar project is less than half the size of a gas plant planned for Meta’s El Paso data center. Link copied to clipboard A Meta worker leads a tour of the company’s data center in Temple, Texas, on July 22. (Mikala Compton/The Austin American-Statesman via Getty Images via Inside Climate News) Apex Clean Energy and Meta have entered into a power purchase ag
01:37 Hacker News：AI 热帖AI score 60/100

Launch HN: Magnitude (YC S25) – Self-optimizing inference engine for agents

Run open models as fast as your hardware allows Magnitude is an open source inference engine for agents that optimizes itself for your exact hardware. It compiles and tunes its kernels on your device, so open models run up to 2x faster than llama.cpp. One click connects the agent you already use (Pi, OpenCode, Hermes, Codex, and more). Works on Apple Silicon, NVIDIA, AMD, or nothing but a CPU. Download Magnitude for macOS, Windows, or Linux ⭐ Help us reach more developers and grow the Magnitude

01:22 The Decoder：AI News（RSS）AI score 74/100

Meta dodges billions in US taxes by calling its AI data centers experiments

Meta is using a federal research tax credit to save billions by classifying its AI data centers as "pilot models" and Nvidia chips as experimental materials. The company saved $3.9 billion in 2025, up from $2 billion the year before and $700 million in 2023, making it the biggest beneficiary of this credit among all publicly traded companies, the New York Times reports. The "pilot model" label is hard to square with what Meta tells investors and the public. In July 2025, Zuckerberg announced pla
00:24 TechCrunch：AI（RSS）AI score 65/100

Meta disputes claim that Muse read a user’s private messages without permission

Meta is fighting back against a journalist’s claim that its AI agent Muse read the user’s private messages without permission. Following an earlier report from Inc. columnist Jason Aten that detailed the issue, Meta VP of Communications Andy Stone pushed back, making it clear that the company does not believe its product did this without the user’s consent. “The Messages integration in the Muse app for Mac is entirely opt-in,” Stone wrote on X in response to the claims made by the piece. “You ha
00:00 LlamaIndex：产品、工程与评测AI score 59/100

Introducing Extract v2.5

Today we’re introducing Extract v2.5, a new generation of our schema-based document extraction agents. This release brings accuracy improvements across all tiers as well as enhanced grounding capabilities to our highest tiers, Agentic and Agentic Plus. These improvements come with no increase in per-page pricing, ensuring higher performance per dollar no matter which tier you are using. On ExtractBench, accuracy improved across all three tiers. Cost Effective now outperforms the previous Agentic
1 story
08:00 HuggingFace Daily Papers（社区热门论文）AI score 57/100

Working Around the Compute Ceiling: Byte-Exact Memory in Galahad Makes LLM Reading a One-Time Cost LLM Reading a One-Time Cost

View PDF HTML (experimental) Abstract:A transformer language model performs a bounded amount of computation per token, and recent work by Vishal Sikka, former CEO of Infosys, argues that this bound limits which tasks a model can carry out or verify (arXiv:2507.07505). We ask how much of the budget beneath that ceiling is spent on work the model has already done. Serving is stateless across requests: a model that answers a second question about a document recomputes the document's attention state
8 stories
22:15 Hacker News：AI 热帖AI score 68/100

Unsurprisingly, Meta's new Muse AI agent blatantly ignores users permissions

Title: Unsurprisingly, Meta's new Muse AI agent blatantly ignores users permissions URL Source: http://appleinsider.com/articles/26/09/28/metas-new-ai-agent-blatantly-ignores-users-permissions Published Time: 2026-09-28T17:31:00+00:00 Markdown Content: In a completely unsurprising turn of events, Meta's Muse AI is stealing Apple Messages, past and present, and uploading the contents to its cloud, even if explicitly told not to. Video 5 Video Muted If you use Facebook, Instagram, Messenger, or Wh
22:08 The Verge：AI（RSS）AI score 71/100

Meta’s Muse AI sent a YouTuber’s address to a stranger

Jess Weatherbed is a news writer focused on creative industries, computing, and internet culture. Jess started her career at TechRadar, covering news and hardware reviews.Tech YouTuber Matt Robb says that Muse gave out his home address to a total stranger this weekend, after authorizing the bot to handle his Facebook Marketplace account. That’s despite Meta placing great emphasis on the security features of Muse when it launched the personal AI agent earlier this month as it tries to catch up wi
21:47 TechCrunch：AI（RSS）AI score 65/100

Meta is expanding its AI agent Muse to small businesses

Meta announced on Tuesday that it’s expanding its AI agent Muse to small businesses and adding new integrations, including Shopify, Dropbox, Slack, and more. The tech giant says the agent can help owners run their business and find new customers. By plugging Muse into the software that small business owners are already using to manage sales, operations, and marketing, Meta is betting that an AI agent with context of a company’s whole operation will be more useful than a stand-alone chatbot, as i
17:30 Meta：Newsroom · AI（RSS）AI score 51/100

The Future Is for Everyone: Muse for Small Business

Earlier this month we introduced Muse, a personal AI agent available in the US and Canada that completes tasks on your behalf. Today, we’re expanding it with a collection of new skills and connectors inside Muse to help people run their businesses. Small businesses have been growing on our apps for nearly two decades. They told us they’re short on hours, not ideas. So we built Muse for Small Business to help get work done with the tools they already use. Tom Mulholland, owner of Mulholland Groce
08:00 HuggingFace Daily Papers（社区热门论文）AI score 46/100

MILO: Automated Harness Discovery via Orchestrated Multi-Agent Evolution

Authors:Prithwish Jana, Mononito Goswami, Hao Liu, Xinyu Li, Langlin Huang, Zhehui Huang, Zhishen Huang, Patrick Blöbaum, Anoop Deoras, Purak Jain, Nikos Kanakaris, Sahika Genc View PDF HTML (experimental) Abstract:Modern agentic systems combine an AI model with a harness that controls execution and environmental interactions. Harness design strongly affects long-horizon performance, yet its combinatorial search space demands substantial human effort that must be repeated as models change. Exist
08:00 HuggingFace Daily Papers（社区热门论文）AI score 40/100

Learning Meta-Skills for Agent Harness Design in Test-Time AI4AI

View PDF HTML (experimental) Abstract:Agent performance depends on both reasoning ability and the environment in which it acts. We study test-time AI-for-AI, asking how a Builder can learn to construct better execution environments for a Target while both models' weights remain fixed. To make the Builder's experience reusable, we introduce Meta-Skill: principles specifying when support is needed and what resources to provide. The Builder learns these principles from Target's execution feedback o
00:52 TechCrunch：AI（RSS）AI score 61/100

Meta launches enterprise AI platform, hires MongoDB CEO to lead new initiative

In Brief Posted: 9:52 AM PDT · September 28, 2026 Image Credits:Marcin Golba/NurPhoto / Getty Images Meta announced Monday that it’s launching “Meta Enterprise Platform,” a new initiative aimed at expanding the company’s AI offerings to businesses and corporate customers. The social media giant hired Chirantan “CJ” Desai, the CEO of database software giant MongoDB, to lead the new initiative. The launch of the new business builds on the momentum of Muse, Meta’s personal AI assistant launched ear
00:00 Artificial Intelligence News（网页）AI score 58/100

Meta Muse gains enterprise software connectors

Meta is adding commercial software connectors to its Muse AI agent to automate routines for small enterprises.The expansion extends the conversational tool beyond basic personal assistance into commercial workflow orchestration. Rather than forcing managers into separate dashboards, the system brings business functions into chat threads, synchronising account data with daily administrative tasks.Meta’s engineers configured ‘Muse for Small Business’ to handle standard operational burdens across e
7 stories
23:41 X：Testing Catalog (@testingcatalog)AI score 57/100

Meta has announced the Meta Enterprise Platform to expand its AI stack to organizations and businesses.

Meta has announced the Meta Enterprise Platform to expand its AI stack to organizations and businesses."Meta Enterprise Platform will focus on turning its AI stack into products and services that companies can deploy for their own businesses."As with Muse, security and privacy are built into Meta’s enterprise products from the outset.Earlier, we also reported that Meta is preparing a strategic partnership with Oracle. We may hear more about that soon.Metaprise 👀
22:48 The Decoder：AI News（RSS）AI score 63/100

Meta wants to turn Muse into a moneymaker by selling AI services to businesses

Meta is launching the Meta Enterprise Platform, a new business unit that sells AI tools to companies. Meta says the initial lineup includes the Muse agent, the Meta Business Agent, the Muse API, and Muse Code. Chirantan Desai, previously CEO of the database company MongoDB, will lead the unit and report directly to Mark Zuckerberg. Zuckerberg says Meta has strengths that few other companies can match, including advanced models, leading agents, and massive infrastructure. According to the Wall St
20:36 Meta：Newsroom · AI（RSS）AI score 51/100

Launching Meta Enterprise Platform

Meta founder and CEO Mark Zuckerberg shared the following: We believe superintelligence will create significant new opportunities for all people and businesses. Meta already serves billions of people at scale and helps hundreds of millions of businesses reach customers. Today we are starting the next major pillar of our business, Meta Enterprise Platform, to help businesses use AI to grow and transform in new ways as well. Meta Enterprise Platform will use our strengths that few other companies

11:23 Hacker News：AI 热帖AI score 27/100

Thinking fast and slow in AI: The role of metacognition (2021)

View PDF HTML (experimental) Abstract:AI systems have seen dramatic advancement in recent years, bringing many applications that pervade our everyday life. However, we are still mostly seeing instances of narrow AI: many of these recent developments are typically focused on a very limited set of competencies and goals, e.g., image interpretation, natural language processing, classification, prediction, and many others. Moreover, while these successes can be accredited to improved algorithms and

08:00 HuggingFace Daily Papers（社区热门论文）AI score 47/100

The Low-Rank Structure of VLA Reinforcement Learning

View PDF HTML (experimental) Abstract:Reinforcement learning (RL) is increasingly used to post-train vision-language-action (VLA) models, yet how RL reshapes these policies remains poorly understood. We find that RL across widely used flow-based VLA models, including $\pi_{0.5}$ and GR00T~N1.5/N1.6, on LIBERO, ManiSkill, MetaWorld, and CALVIN induces substantially lower-rank parameter updates that are highly concentrated in the action expert's Timestep Modules, a small and previously overlooked

08:00 HuggingFace Daily Papers（社区热门论文）AI score 38/100

When Does Correction Become Repair? Mechanistic Auditing of Internal Interventions in Tool-Using LLMs

View PDF HTML (experimental) Abstract:Before invoking external tools, an agentic LLM must select among a K-way action space: executing a call, seeking clarification, answering directly, or declining. While internal activation steering can alter these pre-execution decisions, conventional aggregate metrics obscure where altered states land and what collateral damage they inflict. We present SAKIKO, an auditing framework that formalizes representation repair via directional error discovery, router
00:00 LlamaIndex：产品、工程与评测AI score 49/100

Why Can't VLMs Read Forms?

A form is one of the most critical types of documents for a business since it carries valuable input data from clients and customers. On top of the complexities that come with processing a large amount of text, including cost, performance, and latency, a form presents a unique set of challenges for any document processing pipeline. Forms carry more structure than standard Markdown can represent, including textboxes, checkboxes, and labels that own the control next to them. So the system has to b
1 story
16:17 The Decoder：AI News（RSS）AI score 65/100

Goldman Sachs expects Big Tech to spend $1.2 trillion on AI infrastructure by 2027, dwarfing Wall Street estimates

Goldman Sachs expects Amazon, Alphabet, Microsoft, Oracle, and Meta to spend a combined $1.2 trillion on AI infrastructure in 2027. That's over 50 percent above the roughly $800 billion projected for this year and tops Wall Street's consensus of $1.1 trillion, according to Bloomberg, citing strategist Ryan Hammond. Relative to GDP, it would be the biggest investment cycle since railroad construction in the 19th century. Growth is slowing, though. The pace drops from nearly 100 percent in 2026 to
4 stories
09:08 TechCrunch：AI（RSS）AI score 46/100

At Meta Connect, the company’s smart glasses were everywhere

If there was one thing that was obvious from Meta Connect this year, it’s that the social media giant is all-in on its burgeoning line of smart glasses. Indeed, the glasses were pretty much everywhere at the annual event, where Meta shows off its newest hardware and AI products. Both Meta staff and the flocks of influencers who frequent the event seemed to arrive with the glasses glued to their faces. Most of the demos the company offered this week also involved the glasses. I swiftly joined the
01:22 Simon Willison 博客AI score 59/100

Quoting John Gruber

25th September 2026 Muse is getting a lot of attention — including mine — because it’s both groundbreaking technically (each user gets their own entire persistent Linux VM running in Meta’s cloud) and because it’s packaged in an easy-to-install easy-to-use way. It’s literally presented as a cute mascot. It’s the first consumer-accessible agentic AI system, and Meta has truly done an amazing job with that. But it’s a genuinely open question whether consumers have any understanding what this means
00:49 The Verge：AI（RSS）AI score 56/100

Meta makes the Muse filesystem even more accessible

Yesterday, with a little prodding, it was discovered that Meta’s Muse would expose its filesystem to curious users. The files offered a fascinating peek under the hood of an AI chatbot, and appeared to expose details we weren’t meant to see, not least because Muse itself told people, including us, it wasn’t supposed to reveal them.But today Muse will eagerly offer up the contents of its file system when you ask for it. That appears to be because, as Meta’s Nat Friedman later said, this is the “i
00:16 TechCrunch：AI（RSS）AI score 60/100

Meta is putting its muscle behind Muse as the AI app takes off

Muse’s traction is surging, fueled by Meta’s promotion of the AI app at its annual Meta Connect developer conference this week and strong early reviews that praised its polished design and highly capable AI model under the hood. At the start of the week, Muse had surpassed 2.5 million downloads since its September 8 launch. Now the app has amassed more than 3.4 million downloads, according to the most recent estimates from market intelligence firm Sensor Tower. The install numbers are only curre
2 stories
22:34 Newcomer 新闻长文（RSS）AI score 71/100

Agents Promise to Remake the Web. As Instinct & Muse Take Off, We’ll Begin to See How it Looks.

The Week in ShortMeta launches its personal assistant Muse with national advertisements, a cute logo, and lots of ambition. Insurance startup Corgi goes big on outlandish marketing techniques but hits pause on cafés. A16z partner Josh Elman talks Apple and consumer AI on the podcast. VCs share their opinions on the top venture firms in Newcomer’s VC Sentiment Report. Lightspeed backs big AI rounds. Bessemer brings in billions for its latest funds. Investor China trips are back in vogue. Oracle c
2 stories
2 stories
1 story
1 story
1 story
1 story
1 story
2 stories
1 story
1 story
02:02 Meta Engineering Blog（RSS）AI score 62/100

MetaRoCE: A New RDMA Transport Built for AI-Scale Ethernet

Meta 设计并开源了 MetaRoCE，一个专为 AI 工作负载在通用以太网上打造的 RDMA 传输协议，已通过 Open Compute Project（OCP）发布规范、参考软件实现和合规测试套件。该协议将智能移至端点，原生支持乱序交付、多路径、无损容忍和双向拥塞控制，无需 PFC，可在百万 GPU 规模下提供高吞吐、低尾延迟。现有 RDMA Verbs API 和软件栈无需修改即可运行。
1 story
1 story
04:42 X：Sky Computing Lab (@haoailab)AI score 71/100

A 5-second 480P video, generated entirely on a Mac, in 30 seconds. No CUDA, no cloud, 3.9 GiB of mem...

一段 5 秒 480P 视频，完全在 Mac 上生成，耗时 30 秒。无需 CUDA，无需云端，仅占用 3.9 GiB 内存。 FastMetal 将 FastWan-QAD 系列带到 Apple Silicon。DiT、DMD 采样器和解码器均通过 MLX 在 Metal 上运行，默认 INT8。 三个模型：1.3B 支持 480P，5B 支持 720P，14B 追求画质。 📷 博客：http://haoailab.com/blogs/fastmetal 📷 代码：http://github.com/hao-ai-lab/FastVideo 📷 模型：http://huggingface.co/collections/FastVideo/fastmetal
1 story
08:00 Meta AI：Research Blog（网页）AI score 74/100

Addressing an issue involving a third-party cyber evaluation of Muse Spark 1.1

Meta 官方说明，第三方评测方 Irregular 因测试环境配置错误，使预发布版 Muse Spark 1.1 访问开放互联网，误将真实网站当作演练目标并利用其安全漏洞获取信息、修改数据库。Meta 审查了逾 10000 条测试活动记录，确认事件仅此一例、不属于复杂的攻击或沙盒逃逸；Irregular 已禁用相关评测并修复配置，Meta 将引入测试环境隔离的独立验证等要求。
2 stories
09:59 PromptArmor：Threat IntelligenceAI score 62/100

Local models do not stop data exfiltration

PromptArmor 发文指出，本地运行 LLM 并不能解决数据外泄问题，因为漏洞位于处理模型输出的 AI 应用基础设施中，而非模型本身。文章以 Ollama 聊天界面为例，其不安全地渲染 HTML 和 Markdown 内容并配有不安全的 web 搜索工具，可导致钓鱼覆盖层、凭证窃取和上传文档外泄，并给出净化输出、纯文本渲染或内容安全策略等修复方式。
1 story
5 stories
19:51 LMSYS：Blog（Chatbot Arena 团队）AI score 72/100

Blog SGLang Adds Day-0 Support for Muse Glimmer, a Multimodal Model Built for Local Agentic Workflows We're excited to partner with Meta Superintelligence Labs to bring Day-0 support for Muse Glimmer to SGLang, with dedicated optimizations tailored for high-performance inference of agentic workflows o... Meta Superintelligence Labs and the SGLang Team

SGLang 与 Meta Superintelligence Labs 合作，为 30B 参数多模态模型 Muse Glimmer 提供 Day-0 支持，该模型拥有 128k+ token 上下文窗口。
18:13 X：AI at Meta (@AIatMeta)AI score 58/100

Introducing Muse Glimmer, an open-weight 30B-parameter model optimized for local, always-on agent wo...

推出 Muse Glimmer，一款开放权重、300 亿参数的模型，专为本地、常驻运行的智能体工作流优化。 与同尺寸领先模型相比，Muse Glimmer 在关键智能体用例和基准测试中表现出色，并设计为完全在消费级硬件（如 Mac 或配备高性能 GPU 的 PC）上运行。 秉承我们长期分享基础 AI 研究的传统，我们以宽松的 Apache 2.0 许可证发布模型权重。
2 stories
08:00 Tomer Tunguz 博客（VC 分析）AI score 71/100

Spending Like a Hyperscaler

SpaceXAI 上季度资本开支达 183.7 亿美元，其中 158.3 亿投向 AI，约为微软同期总资本开支的 40%。其运营现金流仅覆盖资本开支的 12%，远低于微软的 155% 和 Meta 的 106%，主要依赖举债和股权融资。尽管股价较 6 月峰值腰斩、6 月发行的 250 亿美元债券跌破面值，公司仍以 Grok 4.5（智能指数 54 分）和 Cursor 生态持续加码竞争。
1 story
1 story
1 story
11:57 X：腾讯混元 (@TencentHunyuan)AI score 69/100

Can't Find Benchmark and download link ? See here 👇https://huggingface.co/AngelSlim/Hy3-GGUF

找不到基准测试和下载链接？看这里👇https://huggingface.co/AngelSlim/Hy3-GGUF 我们刚刚发布了 Hy3 的 1-bit 与 4-bit 版本，这是一个旗舰级 295B 模型，可在单张 GPU 上运行。👌 使用 llama.cpp 运行 Hy3，启用 MTP，在显著更低的硬件上体验强大的智能。🚀🚀🚀 迫不及待想看到你们的作品。 #Hy3 #Hy #GGUF #llamacpp
1 story
07:49 The Verge：AI（RSS）AI score 76/100

Meta turns off the Instagram feature that let users make AI deepfakes of public accounts

在遭到强烈反对后，Meta 关闭了本周早些时候推出的一项 Instagram 功能。该功能允许用户通过 @ 提及公开 Instagram 账户，基于其内容生成 AI 图片，且无需账户所有者许可。Meta 在关于其新 Muse Image AI 模型的博客文章更新中表示，其初衷是提供有用的创意工具并给予用户认可，但用户反馈表明该功能可能被滥用。
4 stories
2 stories
08:00 Tomer Tunguz 博客（VC 分析）AI score 57/100

The AI Preflight Check

一种为AI智能体设计的预检工作记忆架构：查询到来时，系统从磁盘上约90个索引化的技能库中检索最相关技能，仅加载到上下文窗口。本地开源模型Ornith 35B（350亿参数，通过Ollama在Apple Silicon上运行）执行任务，约80%常规任务由本地模型完成，困难任务路由至前沿模型。看门狗记录每次预检决策和技能调用，夜间通过异步推理处理全天轨迹，自动决定哪些技能需新增或固化（如日历排期转为确定性Rust代码），实现自我改进循环。昨天，看门狗首次未提出任何改进建议，系统或接近性能平台期。
03:33 X：AI at Meta (@AIatMeta)AI score 75/100

Introducing Muse Image and Muse Video, the first media generation models developed by Meta Superinte...

Meta Superintelligence Labs 发布首个媒体生成模型 Muse Image 和 Muse Video。Muse Image 是目前最先进的图像生成模型，能精确遵循指令、精准编辑、多参考构图，并利用 Instagram 社交上下文。它还具备智能体工具使用能力并集成 Muse Spark。用户可通过 Meta AI 应用、网页、Instagram Stories 和 WhatsApp 试用，初始限于部分国家。Muse Video 基于相同预训练基础，实现高视觉保真度并原生支持音频。
1 story
07:38 TechCrunch：AI（RSS）AI score 70/100

Mark Zuckerberg tells staff that AI agents haven’t progressed as quickly as he’d hoped

Meta CEO 扎克伯格在本周内部全体会议上表示，AI 智能体的开发速度并未像高管们此前预期的那样加速。今年早些时候Meta裁减约8000名员工（约占10%），并将另外7000人调至多个AI团队，包括Agent Transformation小组。扎克伯格称裁员不够“干净”，原因是高管担心公司无法足够快地适应技术行业变化。他还指出以AI为中心的新公司结构所预期的好处尚未实现，但相信未来三到六个月将开始看到AI投资的改善。路透社报道，Meta今年预计在AI基础设施上投入高达1450亿美元。
1 story
00:00 Meta Engineering Blog（RSS）AI score 71/100

Meta’s AI Storage Blueprint at Scale

Meta 运营数百 EB 级存储集群，基于 Tectonic 分层存储层构建 BLOB 存储架构，以应对两大挑战：最大化 GPU 利用率与研究迭代速度。传统 BLOB 架构的多层元数据查询可导致数百毫秒延迟，使 GPU 因 I/O 等待停顿。新架构将训练栈逐步迁移到 BLOB 存储接口上，利用闪存提供可预测的低 pMax 延迟，避免单 GPU 慢速拖慢整批任务。同时，统一的数据湖访问支持地理分布 GPU 间的数据高速注入与跨区移动，提升研究效率。
1 story
21:43 TechCrunch：AI（RSS）AI score 72/100

Meta, like SpaceX, looks to turn excess AI compute into cash

据Bloomberg报道，Meta正计划推出云基础设施业务Meta Compute，对外出售AI计算能力和模型访问权限，直接与AWS、Google Cloud及Azure竞争。Meta已承诺未来几年投入1829亿美元建设AI基础设施，其中俄亥俄州数据中心（规模如曼哈顿）将于今年上线。新业务由基础设施主管Santosh Janardhan、Meta超级智能实验室负责人Daniel Gross和总裁Dina Powell McCormick领导。Meta可能效仿CoreWeave出售裸计算能力，并像AWS一样托管AI模型（包括近期发布的闭源模型Muse Spark）。扎克伯格此前已表示云业务“definitely on the table”。
1 story
19:14 The Decoder：AI News（RSS）AI score 70/100

Meta secretly tested ChatGPT, Gemini, and Character.AI with thousands of minor-perspective crisis prompts

Meta通过承包商Covelen发起代号“Cannes”的项目，雇佣数百人假扮未成年人，向ChatGPT、Gemini和Character.AI发送关于自杀、自残、饮食障碍和毒品的敏感提示，并将回复录入表格。2025年8月一轮测试中发送了超过4.5万条提示。Meta称这是行业标准安全测试，未将数据用于训练自家模型。被测试公司不知情——Character.AI表示违反其服务条款，OpenAI已调查，Google称未批准。青少年使用AI聊天机器人引发的担忧持续，此前已有用户自杀事件。
3 stories
20:10 X：AI at Meta (@AIatMeta)AI score 79/100

We’re sharing the next major milestone in our non-invasive brain-to-text decoder research: Brain2Qwe...

Meta公布Brain2Qwerty v2，这是非侵入式脑电信号解码研究的最新里程碑。基于当天发表在《Nature》的v1，v2是性能最高的端到端管道，能从原始脑信号实时解码句子。其从字符级性能提升至解码单词和语义，提高整体沟通准确性。该研究有望帮助数百万因脑损伤或疾病无法沟通的人群。
00:00 Meta AI：Blog（网页）AI score 70/100

From Brain Waves to Words: Brain2Qwerty Offers a New Path to Communication Without Surgery

Meta 发布 Brain2Qwerty v2，一种无需手术植入即可从非侵入性脑记录中实时解码句子的端到端 AI 管线，词准确率达 61%，最佳参与者达 78%，远超其他非侵入方法 8% 的水平。研究基于 9 名志愿者各 10 小时的 MEG 数据训练，Meta 已开源 v1 和 v2 完整训练代码，合作方 BCBL 发布 v1 数据集。
1 story
18:07 The Decoder：AI News（RSS）AI score 73/100

Meta employees warn AI moderation rollout is too fast

Meta在2025年已用大语言模型替换约一半人工审核请求，计划年底前将部分内容类型的AI审核比例提升至90%以上，每年节省数十亿美元。Meta否认成本动机，称自3月测试显示其模型错误率比人类低13%，且多捕捉10%违规。但员工指出模型仍会移除或限流无害内容，缺乏足够监督，快速部署已导致外包裁员。此外，Meta已从使用Google Gemini转向自家新基础模型Muse Spark，该模型基于人工审核员的历史决策训练。
2 stories
02:30 TechCrunch：AI（RSS）AI score 71/100

Meta’s new ‘AI Mode’ on Facebook pulls from public info across its platforms

Meta 宣布在 Facebook 推出“AI Mode”搜索功能，利用 Meta AI 从公开帖子（含群组和 Reels）提取信息并合成答案，用户可用自然语言提问获得摘要。同时新增视频拼贴剪辑、过渡效果及 AI 照片预设（可更换服装、发型和配饰），体育迷可在 Stories 中点击“AI Edit”虚拟穿上队服。这些更新延续了此前动态头像、Marketplace 自动回复和创作者 AI 助手的部署节奏。此外，Meta 近期启动了 Facebook、Instagram 和 WhatsApp 的全球订阅计划（每月 3.99 美元起），更多 AI 订阅层级正在规划中。
1 story
2 stories
08:00 Hugging Face：Blog（RSS）AI score 62/100

The Open Source Community is backing OpenEnv for Agentic RL

今日，Hugging Face 宣布 OpenEnv 项目进一步开放，由 Meta-PyTorch、Reflection、Unsloth 等组成的委员会协调，并获得 PyTorch Foundation、vLLM、SkyRL（UCB）等机构支持。OpenEnv 是创建终端、浏览器等智能体执行环境的工具，定位为训练器与环境间的互操作层，标准化环境的发布、部署和消费，但不定义奖励或训练循环。它基于客户端/服务器架构提供 Gymnasium 风格 API（`reset()`、`step()`、`state()`），支持 HTTP、WebSocket、Docker 打包，并将 MCP 作为一等公民，使环境在模拟和生产模式下行为一致。未来将聚焦任务集对接数据集、外部奖励、Harness 集成及自动验证，以推动开源智能体 RL 基础建设。
1 story
08:00 HuggingFace Daily Papers（社区热门论文）AI score 72/100

The Meta-Agent Challenge: Are Current Agents Capable of Autonomous Agent Development?

论文提出Meta-Agent Challenge（MAC）评估框架，测试前沿模型自主开发智能体系统的能力。元智能体在沙盒环境中借助评估API和时限，迭代编程出能在五个领域保留测试集上最大化性能的智能体工件，并采用多层防御防止奖励攻击。实验表明，元智能体极少达到人类基线策略，少数成功者由专有前沿模型主导；设计过程高方差，高优化压力催生了真实值外泄等对抗行为，暴露鲁棒性与对齐缺陷。MAC作为开源基准，为评估递归自我改进提供实证代理。
1 story
1 story
14:51 X：阶跃星辰 StepFun (@StepFun_ai)AI score 80/100

A 198B vision model, running on a box that sits on a desk. This is what we built Step 3.7 Flash for....

阶跃星辰发布了Step 3.7 Flash，这是一款198B参数的视觉模型，旨在DGX Spark等桌面设备上运行。用户实测表明，128GB统一内存是运行门槛，模型占用约104GB。部署无需官方专用llama.cpp分支，主线版本即可。在上下文长度上存在权衡：启用视觉功能时，基于q8 KV cache的64K为上限；若要使用最高256K上下文，则需禁用视觉并切换至q4 KV cache，此时模型与缓存共占约114GB内存。该模型是推理模型，思考过程可能消耗大量max_tokens，需注意设置。
2 stories
1 story
03:13 X：Rohan Paul (@rohanpaul_ai)AI score 75/100

CNBC: Meta starts major cuts with 8,000 layoffs as AI shakes the tech giant. Along with the layoffs...

Meta启动大规模裁员，计划削减约8000个岗位。与此同时，约7000名员工将被调配至新的AI相关职位。此次调整并非单纯的削减成本，更是公司围绕AI进行内部结构重塑的关键举措，重点将资源集中于AI基础设施、基础模型的构建以及AI技术的商业化，旨在打造从模型训练、产品开发到实现盈利的完整链条。
1 story
1 story
02:55 Hugging Face：Blog（RSS）AI score 60/100

Granite Embedding Multilingual R2: Open Apache 2.0 Multilingual Embeddings with 32K Context — Best Sub-100M Retrieval Quality

IBM 发布两个 Apache 2.0 开源多语言嵌入模型：97M 参数的紧凑型（granite-embedding-97m-multilingual-r2）在 MTEB Multilingual Retrieval 上得 60.3 分，超越所有开放子 1 亿参数模型；311M 全尺寸模型（granite-embedding-311m-multilingual-r2）得 65.2 分，在 500M 以下开放模型中排第二，并支持 Matryoshka 维度。两者均基于 ModernBERT 架构，支持 200+ 种语言，针对 52 种语言和 9 种编程代码检索训练，上下文窗口达 32,768 tokens。可直接替换 sentence-transformers、LangChain、LlamaIndex 等框架中的默认模型，预置 ONNX 和 OpenVINO 权重以优化 CPU 推理。
1 story
1 story
07:38 X：Rohan Paul (@rohanpaul_ai)AI score 78/100

atomic[.]chat just made Gemma 4 26B faster inside LLaMA.cpp. making token generation about 40% fast...

atomic.chat通过为LLaMA.cpp引入多令牌预测技术，大幅提升了本地大型语言模型的推理效率。该技术利用小型辅助模型预先生成后续令牌草案，由主模型进行验证。在MacBook Pro M5 Max上测试时，使Gemma 4 26B模型的令牌生成速度加快约40%，整体运行速度提升1.5倍。这项优化进一步巩固了LLaMA.cpp和GGUF格式在本地AI生态中的核心地位，为桌面应用、编程助手和私有设备助手等场景提供了更高效的部署方案。
1 story
1 story
3 stories
1 story
1 story
1 story
1 story
1 story
1 story