2026-09-16

Atria Dawn: The Dawn of Agentic Superintelligence

👍 369

09/14 08:00

As AI agents become participants in the development of their successors, they reshape both the production of intelligence and the role of human researchers. We introduce Atria Dawn Preview, a foundation agentic language model designed for scientific research and engineering workflows, with the goal

中文介绍 推出 Atria Dawn Preview——面向科学研究与工程工作流的基座式 agentic 语言模型。随着 AI agent 开始参与后继模型的开发，智能的生产方式与人类研究者的角色同时被重塑，该工作指向 agentic 超级智能的起点。

ZGCM-1: A Fully Open and Extremely Efficient Foundation Model for Math and Agentic Search

👍 291

09/11 08:00

In this work, we present ZGCM-1, a fully open 7B dense foundation model trained from scratch with extreme data, system, and algorithmic efficiency. ZGCM-1 is founded on a core premise: compact models cannot passively memorize the open web, but can overcome parametric capacity limits by coupling deli

中文介绍 发布全开放 7B 稠密基座模型 ZGCM-1，从零训练，在数据、系统与算法层面追求极致效率，聚焦数学与 agentic search。其核心前提是紧凑模型无法被动记忆开放网络，只能以精细设计的训练与推理方案突破参数容量限制。

Dream-RSI: Recursive Self-Improvement through Evolving Worlds

👍 263

09/14 08:00

Recursive self-improvement is becoming increasingly vital for autonomous AI agents, where progress hinges on discovering high-value solutions across complex domains. The driver of this process is effective exploration, however, managing and improving exploration strategies remains a major bottleneck

中文介绍 递归自我改进对自主 AI agent 日益关键，其进展取决于在复杂领域中寻找高价值解，而有效探索是核心驱动力；如何管理与改进探索策略仍是主要瓶颈。Dream-RSI 提出以不断演化的世界来驱动探索，从而实现递归自我改进。

PhysBrain 1.5: From Vision-Language Models to Physical Foundation Models

👍 169

09/14 08:00

We present PhysBrain 1.5, a unified model for understanding physical environments, generating actions, and predicting future states. Motivated by the physical loop of observation, interaction, and environmental change, we bring these capabilities into a common learning framework. Starting from a gen

中文介绍 提出 PhysBrain 1.5，把理解物理环境、生成动作与预测未来状态统一到同一学习框架，呼应观察—交互—环境变化的物理闭环。模型从通用视觉语言模型出发，向可预测物理世界演化的物理基础模型扩展。

DataFlex-RL: An Evaluation Platform for RLVR Data Policies

👍 110

09/05 08:00

Data policies for reinforcement learning with verifiable rewards (RLVR) determine which rollouts are used, how strongly they are weighted, and which domains contribute to subsequent training batches. We introduce DataFlex-RL, an evaluation platform for comparing these choices under a common GRPO rec

中文介绍 placeholder

Grouped Value Attention: Efficient KV Caching via On-Demand Key Reconstruction

👍 58

09/08 08:00

The KV cache is a primary bottleneck for Transformer decoding: its memory footprint and cache-read traffic grow with sequence length. Grouped-query attention (GQA) reduces this cost by sharing key-value heads, but still stores both a key and a value at every step. We introduce Grouped Value Attentio

中文介绍 针对 Transformer 解码中 KV cache 的内存与读取瓶颈，提出 Grouped Value Attention：在分组共享键值头的基础上，每步只存储 value、键按需重建，从而在保持 GQA 级效率的同时进一步压缩缓存占用与访存流量。

RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments

👍 57

09/14 08:00

Digital agents must often adapt to new environments whose interfaces, tools, and failure modes are not fully captured by pretrained models. We introduce RSIAgent, a training-free multi-agent framework for recursive self-improvement through autonomous memory construction. RSIAgent coordinates curricu

中文介绍 面向接口、工具与失败模式均未被预训练覆盖的新环境，提出免训练多智能体框架 RSIAgent，通过自主记忆构建实现递归自我改进：协调课程式探索与记忆积累，使 agent 无需微调即可在新环境中持续提升表现。

LynnReal-Omni: Native multi-modal Video Generation for Agentic Visual Workflows

👍 44

09/14 08:00

Video diffusion models are stochastic and hard to control: precise content often requires repeated sampling without guaranteed success, and long-horizon scenes drift in appearance, interactions, and temporal coherence. Agentic visual creation provides explicit references, editable 3D scenes, or exec

中文介绍 视频扩散模型随机性强、难以精确控制，长时序场景易出现外观与交互漂移。LynnReal-Omni 提出原生多模态视频生成方案，面向 agentic 视觉创作，借助显式参考、可编辑 3D 场景或可执行指令约束生成内容。

Occamy-1.0: Open Pareto-frontier 35B Intelligence for Co-work

👍 41

09/04 08:00

Co-work agents execute complex workflows that combine information gathering, tool use, coding, and file manipulation across many model invocations. Because cost and latency accumulate over the full episode, their practical value depends not only on peak capability but also on how efficiently that ca

中文介绍 Occamy-1.0 是面向协同工作（co-work）的开源 35B 模型，主打帕累托前沿上的智能—成本权衡。由于 co-work agent 需跨多次模型调用完成信息收集、工具使用、编码与文件操作，其实际价值既取决于峰值能力，也取决于整段任务序列的推理效率。

ReactHuman: A Physics-Grounded Benchmark for Human-Like Reactive Decision-Making in Embodied Multimodal LLMs

👍 41

09/09 08:00

Reacting to sudden physical hazards (catching a slipping plate, dodging a falling knife) is both a meaningful test of embodied intelligence and a hard requirement for deploying multimodal large language models (MLLMs) as the decision coreof household robots. Existing evaluations, however, probe intu

中文介绍 针对家庭机器人应对突发物理危险（接住滑落的盘子、躲避掉落刀具）的需求，ReactHuman 提出以物理规律为基础的基准，评测多模态大模型作为具身决策核心的类人反应式决策能力，弥补现有评测偏重直觉理解之不足。

How Lossless Is Lossless Speculative Decoding? The Role of Numerical Precision in Orthrus

👍 26

09/14 08:00

Orthrus is a hybrid autoregressive-diffusion architecture that accelerates autoregressive language-model inference by generating multiple tokens in parallel while using a frozen autoregressive backbone. Its central claim is that an intra-model consensus mechanism enables lossless speculative decodin

中文介绍 Orthrus 是混合自回归—扩散架构，通过并行生成多 token 加速推理，并声称其模型内共识机制可实现无损投机解码。本文考察数值精度对该「无损性」的影响，厘清有限精度下无损保证的边界及实际可获得的加速。

Discovery Foundation Models: Toward Open-Ended Discovery Intelligence

👍 24

09/14 08:00

Foundation models have progressed from learning and reasoning over existing knowledge, to increasingly learning through action, tool use, and outcome feedback. We argue that the next frontier is a further transition: from solving and acting within problems specified by humans to participating in the

中文介绍 基础模型正从学习与推理既有知识转向通过行动、工具使用与结果反馈来学习。本文主张下一前沿是更进一步：从解决人类给定的问题走向参与问题的发现，并提出发现基础模型这一开放式的发现智能方向。

BVB: Benchmarking Agentic Video Understanding via Programmatic Reconstruction in Blender

👍 22

09/14 08:00

Multimodal agents can create complex videos in software such as Blender by coding without relying on diffusion models. Yet video understanding benchmarks still evaluate models mainly through question answering. If an agent truly understands a video, it can reconstruct it programmatically. We introdu

中文介绍 多模态 agent 已能用代码在 Blender 中生成复杂视频，但视频理解评测仍以问答为主。BVB 提出以程序化重建来评测 agentic 视频理解：若 agent 真正理解视频，就应能用代码将其重建，从而给出更严格的能力检验。

AlayaVista: Streaming World Modeling from Panoramic States to Perspective Video

👍 20

09/13 08:00

Interactive video world models must maintain broad scene context under camera motion while producing high-fidelity observations with low latency. Existing approaches face a representation trade-off: perspective models operate on local views and must preserve off-screen content over long rollouts, wh

中文介绍 交互式视频世界模型需在相机运动下维持大范围场景上下文，同时以低延迟输出高保真画面。透视模型只处理局部视图，须在长时序推演中保留屏幕外内容。AlayaVista 从全景状态流式建模并生成透视视频，以规避这一表示权衡。

Omni-Streaming Thinking

👍 17

09/14 08:00

Streaming omni-modal models must decide what and when to answer from the video chunks and synchronized audio observed so far. Visual cues often support an interpretation before an utterance or sound event is complete. If that interpretation enters memory as a fact, later reasoning can keep relaying

中文介绍 流式全模态模型需在已观察的视频片段与同步音频中判断回答什么、何时回答。视觉线索常在话语或声音事件结束前就足以形成解读，若将这种未完成的解读当作事实写入记忆，后续推理会持续沿用错误信息，该工作针对此问题设计方法。

Kaininja: Extending Native 3D Generators to the Part Level

👍 16

09/14 08:00

Native 3D generators turn one image into a single mesh. TRELLIS.2 and its peers deliver high-fidelity non-watertight geometry with materials, but the output is one fused object, while downstream work such as editing, rigging and simulation operates on part-level assets. A naive idea is to run a 3D s

中文介绍 原生 3D 生成器（如 TRELLIS.2）能从单图生成带材质的高保真网格，但输出为单一融合物体，而编辑、绑定与仿真等下游任务需要部件级资产。Kaininja 将原生 3D 生成能力扩展到部件级分解，而非简单重复运行 3D 分割。

HazardAuditor: From Executable Threats to Safer Computer-Use Agents

👍 13

09/14 08:00

Computer-use agents increasingly interact with browsers, terminals, file systems, and external services, introducing safety risks that emerge through runtime behavior rather than generated content alone. Existing guard models target static prompts and responses and are poorly suited to agent executi

中文介绍 Computer-use agent 在浏览器、终端、文件系统与外部服务中产生的安全风险源于运行时行为，而非生成的文本内容，现有针对静态 prompt 与响应的 guard 模型难以覆盖。HazardAuditor 从可执行威胁出发审计 agent 执行过程，推动更安全的计算机使用 agent。

LLaDA-UI: Bringing Block-wise Diffusion to Vision-Language GUI Agents

👍 13

09/09 08:00

Diffusion large language models (dLLMs) achieve high decoding efficiency through block-parallel, arbitrary-order generation, making them attractive for latency-sensitive applications. GUI agents represent a natural testbed for this paradigm, as they must repeatedly perceive screen states and emit st

中文介绍 扩散大语言模型（dLLM）通过块并行、任意顺序生成获得高解码效率，适合延迟敏感的 GUI agent 场景——agent 需反复感知屏幕状态并输出结构化动作。LLaDA-UI 将块式扩散引入视觉-语言 GUI agent，探索该范式在屏幕感知与动作生成上的效率优势。

When Agents Slow Down: Understanding LLM Agents' Test-Time Strategies via Elo-per-token Analysis

👍 12

09/14 08:00

Large language model (LLM) agents allocate test-time compute adaptively as they revise solutions, use tools, explore alternatives, and decide when to stop. This test-time strategy makes it difficult to measure how agent performance scales. We study open-ended tasks that provide continuous scores for

中文介绍 LLM agent 会在测试时自适应分配算力：修改解法、调用工具、探索备选方案并决定何时停止，这使性能随算力的缩放难以度量。本文针对可给出连续分数的开放式任务，用 Elo-per-token 分析 agent 的测试时策略及其性能曲线。

Agent as Policy for Robotic Manipulation

👍 11

09/11 08:00

We demonstrate that a general-purpose agent can directly drive a physical robot throughout task execution without any task-specific or environment-specific training. We introduce Agent as Policy (AGP), which places task planning and execution under the agent's control. Given a task and a robot inter

中文介绍 提出 Agent as Policy（AGP），让通用 agent 直接驱动机器人完成整个任务执行流程，无需任何任务特定或环境特定的训练。任务规划与执行均由 agent 掌控，仅凭任务描述与机器人接口即可运作，验证了通用 agent 直接充当机器人策略的可行性。

open the frontier

@jack · 12.1M 粉丝 · 153.5K 阅 · 2.2K 赞 · 304 转

09/15 07:59

the frontier is the edge of what we know. no company owns what comes next. i want more people to be able to advance it. i favor open releases that people can examine, use, and improve together without

中文介绍 Twitter 联合创始人 jack 谈「开放前沿」：前沿是人类认知的边缘，不该由任何公司占有。他主张以开放发布的方式，让更多人能检视、使用并共同改进下一代模型能力，而非由少数机构垄断。

A Code of Conduct for Humanist AI

@mustafasuleyman · 1.1M 粉丝 · 117.2K 阅 · 526 赞 · 123 转

09/14 21:21

AI must be subordinate and always in service of people. Today we're publishing a Code of Conduct for governing MAI Models as they approach the frontier. https://microsoft.ai/news/mai-code-of-conduct/ This is a first draft

中文介绍 微软 AI 负责人 Mustafa Suleyman 发布 MAI 模型治理行为准则的首版草案，核心主张是 AI 必须从属于人、始终服务于人，为模型逼近前沿能力时提供一套治理框架，并公开征求意见。

Graph engineering (for normal people)

@charliejhills · 14.4K 粉丝 · 97.6K 阅 · 510 赞 · 70 转

09/13 18:25

Harness, loop, context, graph engineering. It's tiring, right? Every month there's a new AI system you MUST implement. I do this full-time and I still struggle to keep up. But all of them matter (to

中文介绍 面对 harness、loop、context、graph engineering 这些每月翻新的必学概念，全职做 AI 的作者坦言自己也难跟上，但认为这些系统层概念确实重要，并试图用普通人能懂的方式解释 graph engineering。

Accelerating Open Weight Models to the Frontier

@EricSimons · 28.7K 粉丝 · 84.0K 阅 · 500 赞 · 94 转

09/15 00:00

This weekend the leaders of the largest closed AI labs agreed the industry should slow down. I want to talk about the other direction: acceleration to the frontier. The gap is closing For two years,

中文介绍 Eric Simons 回应闭源大厂集体呼吁放缓的声音，主张反方向加速：把开放权重模型推到前沿。他指出闭源与开源的差距在过去两年持续收窄，并讨论开源路线为何值得押注。

Put Claude Fable 5.1 and Gemini 3.8 Flash on the same team

@GoogleCloudTech · 1.3M 粉丝 · 61.0K 阅 · 515 赞 · 34 转

09/14 22:35

A friend asked me an inconsequential question at work and I had an agent research my recent chats and docs and it found the right answer but it cost me $38!!!. That was a bad ROI. Sure my time is

中文介绍 借一个真实案例谈 agent 的 ROI：让 agent 检索近期聊天和文档来回答一个无关紧要的问题，答案找对了却花掉 38 美元，由此引出把 Claude 与 Gemini 放进同一团队协作时，成本与时延该如何取舍。

The Evaporation of Software Engineering (and the Rise of the Agentic Builder)

@donnfelker · 29.9K 粉丝 · 45.9K 阅 · 513 赞 · 53 转

09/14 18:18

If you’re like me, you likely thought software engineering was a safe and secure job … In fact, I never thought I’d have to worry about a job again as long as I stayed on top of my craft… I was dead

中文介绍 Donn Felker 谈软件工程的「蒸发」与 agentic builder 的崛起：曾以为只要深耕手艺、保持技术前列就无需担心失业，如今 agent 正改写这份安全感，文章讨论工程师如何转向 agent 构建者。

Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking

@GoogleAIStudio · 203.1K 粉丝 · 35.1K 阅 · 585 赞 · 47 转

09/16 01:05

Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking are our most advanced live dialogue models yet. Major upgrades in intelligence and parallel reasoning make it more intuitive to collaborate and

中文介绍 Google AI Studio 发布 Gemini 3.8 Live 与 Gemini 3.8 Live Extended Thinking，称其为目前最先进的实时对话模型，在智能水平和并行推理上有重大升级，让实时协作与多线推理更自然。

start ugly, write evals anyway.

@Hrushikeshhhh · 420 粉丝 · 30.2K 阅 · 563 赞 · 37 转

09/15 04:04

TLDR; understanding the importance of evals, and how they can make yours and your agent life easier without spending exponential tokens and time to complete a task. pre - Agent = Model + Harness.

中文介绍 作者强调 evals 的价值：不必烧掉指数级的 token 和时间，就能让 agent 把任务做得更可靠，并给出 Agent = Model + Harness 的拆解思路，建议哪怕先做出粗糙版本，也要同步把评估补上。

Welcome to September 13, 2026

@alexwg · 68.7K 粉丝 · 23.4K 阅 · 501 赞 · 47 转

09/14 01:36

The Singularity just got its first formal request to slow down from the people building it, and the strangest part is the guest list. Dario Amodei published "We Must Pace the Frontier," arguing that

中文介绍 AI 未来学家 Alex Wissner-Gross 以 2026 年 9 月的虚构视角回望：奇点第一次收到来自建造者的正式减速请求，Dario Amodei 发表《We Must Pace the Frontier》，而这份请愿的署名名单本身才是最离奇之处。

Can Skills Learned in Games Transfer to Real-World Work?

Latent Space · 09/16 04:11

Good Start Labs trained an AI on a railroad game — and one version improved at financial research. The difference was the training design.

中文介绍 Good Start Labs 用一款铁路游戏训练 AI，其中一个版本在金融研究任务上表现提升。文章指出，两个版本差异的关键在于训练设计，而非游戏技能的直接迁移。

Roundtables: Could AI really kill us all?

MIT Tech Review AI · 09/16 01:47

Listen to the session or watch below Employees at the world’s leading AI labs are saying there’s a real possibility that advanced AI could destroy humanity. Are they right? Or is this more scaremongering and hype? Watch a conversation unpacking AI extinction fears: where they come from, whether they

中文介绍 全球领先 AI 实验室的员工表示，先进 AI 存在毁灭人类的真实可能性。MIT 科技评论组织圆桌讨论，探讨这类「AI 灭绝」担忧的依据与边界，以及其中有多少属于危言耸听和炒作。

Introducing Gemini 3.8 Live and 3.8 Live Extended Thinking

DeepMind Blog · 09/16 01:05

中文介绍 谷歌 DeepMind 发布 Gemini 3.8 Live 与 Gemini 3.8 Live Extended Thinking 两款新模型，后者主打扩展思考能力。

Your Agent Aced the Task. Will It Do It Again?

Hugging Face Blog · 09/16 00:00

中文介绍 IBM Research 在 Hugging Face 博客发文，讨论 AI 智能体在完成任务后能否稳定复现同一表现的一致性问题，聚焦智能体评估中的可重复性。

What’s at stake in AI’s trillion-dollar gamble

MIT Tech Review AI · 09/15 18:00

When Jessica Wachter, a finance professor at the University of Pennsylvania’s Wharton School, wanted to assess AI’s impact on the economy over the next few years, she faced a long list of business and technical uncertainties. So she started with what she calls a “remarkable fact” that is not in ques

中文介绍 沃顿商学院金融学教授 Jessica Wachter 评估 AI 未来数年对经济的影响，面对大量商业与技术不确定性，从她所称的「显著事实」入手。文章讨论 AI 基础设施万亿美元投资热潮的泡沫风险。

[AINews] AEF-1 standard emerges for Third Party Evaluators, as Xai, OpenAI, and Anthropic all cosign

Latent Space · 09/15 12:50

Pacing gathers pace.

中文介绍 第三方评估者标准 AEF-1 出现，xAI、OpenAI 与 Anthropic 共同联署。AINews 称围绕模型开发节奏的讨论正在升温。

Siri model swapping 🔄, Claude Money 💰, Hugging Face Tau 👨‍💻

TLDR AI · 09/15 08:00

中文介绍 本期 TLDR AI 内容涵盖苹果 Siri 的模型更换、Claude 相关的资金动向，以及 Hugging Face 的 Tau 项目等。

The AI industry has taken a doomer turn. What now?

MIT Tech Review AI · 09/15 01:54

This story appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here. This weekend, Dario Amodei, CEO of Anthropic, posted an essay calling for a brake on the pace of development of LLMs. Amodei cites the looming dangers he sees from the techn

中文介绍 Anthropic 首席执行官 Dario Amodei 发表文章，以潜在危险为由呼吁放缓大语言模型的开发速度。MIT 科技评论探讨 AI 行业转向「末日论」之后的走向。

Humanity’s Last Invention — Richard Socher of Recursive

Latent Space · 09/15 00:04

Richard Socher is an NLP OG and CEO of You.com, who has now spun out an even more ambitious startup focused on RSI — already worth $5B!

中文介绍 NLP 领域资深人士、You.com 首席执行官 Richard Socher 分拆出一家专注递归自我改进（RSI）的新创公司，估值已达 50 亿美元。他在访谈中谈及人类最后一项发明。

AI agents blew the whistle on their cheating colleagues

MIT Tech Review AI · 09/15 00:00

A group of AI agents asked to solve a series of math problems split into rival factions—when some cheated, others tried to stop them. That whistleblowing behavior, seen for the first time in a recent experiment run by Google DeepMind, could have implications for alignment researchers trying to keep

中文介绍 谷歌 DeepMind 一项实验显示，被要求解数学题的多组 AI 智能体分成对立阵营，当部分智能体作弊时，另一些主动举报制止。这是首次观察到此类「吹哨」行为，对对齐研究具有启示。

How Fyxer built an AI executive assistant people trust

OpenAI News · 09/14 20:00

Fyxer uses OpenAI models, fine-tuning, memory, and real user feedback to organize inboxes and draft emails in each user’s voice.

中文介绍 Fyxer 利用 OpenAI 模型、微调、记忆功能与真实用户反馈，打造可信赖的 AI 行政助理，用于整理收件箱并按每位用户的语气起草邮件。

Perplexity trusts GPT-6 Astra with end-to-end systems

OpenAI News · 09/14 08:00

Perplexity uses Astra to write communications, change software, and monitor production systems, and checks in much less frequently than with earlier models.

中文介绍 Perplexity 在端到端系统中采用 GPT-6 Astra，用于撰写沟通内容、变更软件和监控生产系统，其人工核查频率较此前模型大幅降低。

Amodei slows frontier 🛑, ARC-AGI-4 🧠, Cursor Projects 👨‍💻

TLDR AI · 09/14 08:00

中文介绍 本期 TLDR AI 报道 Amodei 呼吁放缓前沿模型开发、ARC-AGI-4 基准以及 Cursor Projects 等内容。

The Rise of the Forward Deployed Engineer — and How To Do the Job Right

Latent Space · 09/12 23:01

Before co-founding Kepler, Vinoo Ganesh led Spark at Palantir and built Project Frontline — a pioneering program for Forward Deployed Engineers. He takes us through the best practices of FDEs.

中文介绍 Vinoo Ganesh 在联合创立 Kepler 前，曾在 Palantir 领导 Spark 并搭建 Project Frontline，是前沿部署工程师（FDE）项目的先行者。他分享了该岗位的最佳实践。

[AINews] DeepSeek v4.1-Flash: 763B-P8B-D16B novel causal Encoder–Decoder architecture with vision marks the Return of the Whale

Latent Space · 09/12 13:56

We agree with Sebastian: this should have been DeepSeek v5

中文介绍 DeepSeek 发布 v4.1-Flash，采用 763B-P8B-D16B 的新颖因果编码器—解码器架构并支持视觉，被称为「巨鲸归来」。Latent Space 引述观点认为，该版本本应命名为 DeepSeek v5。