AI 情报文章归档

Claude 官方视频介绍 Claude Mythos 5.2 预测了完整 AAV 衣壳结构，即用于递送基因治疗的蛋白质外壳。模型编写代码在相邻五链分组上运行开源折叠模型，再利用衣壳对称性拼合全部 60 条链；在 AAV2 上预测结果与实验结构误差约为一个原子宽度。

Anthropic 发布 Claude Fable 5.2，定位为面向高难度技术挑战和开放式技术工作的模型，在编码、推理和视觉等方面有提升。模型现已向 Pro、Max、Team 和 Enterprise 计划开放，官方同时建议多数编码和知识类任务仍从 Opus 5.5 起步。

Anomaly detection (AD) is a critical for trustworthy machine learning and scientific discovery. Existing methods typically define anomalies through local density, isolation heuristics, or distance from a fitted distribution, making them se…

The Platonic Representation Hypothesis (PRH) suggests that neural networks are converging toward similar representation spaces. As language models become more capable, their representations are aligning with those of vision models. However…

Reinforcement learning (RL) can improve an agent's task capability while reinforcing behavior that violates contextual execution requirements. We study this reward-driven misalignment, including reward hacking, through verifiable reward--c…

Persona generalization occurs when fine-tuning a language model in one domain induces behavioral changes in others -- for example, the broadly unsafe behavior observed in emergent misalignment. We discover a surprising "on-the-line" phenom…

字节跳动 Seed 团队发布图像生成模型 Seedream 5.5，面向专业生产流程。官方称其支持最多 100 轮多轮编辑，未改动区域在像素级保持不变且不偏色，并可生成带透明背景的原生 RGBA 图像，同时更真实地渲染光线、材质与皮肤纹理，更准确呈现密集小字。该模型正通过即梦、豆包、Doubao Work 和 Volcengine Ark 逐步推出，还计划推出 Pro 版本。

当前 AI 热点，聚合 1 个独立信源、9 条信号。

Large language model agents are increasingly deployed as multi-agent systems rather than as a standalone model operating within a single context. Recent agent harnesses allow the model to instantiate worker agents and to author the executa…

Latent Diffusion Models dominate generative modeling but face two-stage training complexity and the inherent reconstruction limits of pre-trained autoencoders. Pixel-space Flow Matching offers a simplified end-to-end alternative yet often …

来自 @AIatMeta 的 Muse Ember 现已上线 OpenRouter！🔥 你的编程智能体现在有了一整支队伍。先用 Muse Wonder 这样的前沿模型做规划，再把有边界的子任务分发给 Muse Ember，看它们飞速执行 🚀

As large language models become more capable agents, they can exhibit deceptive behaviors, cheating human for higher rewards while appearing helpful. We attribute this failure to evaluation-side partial observability, where models are eval…

AI models' growing capacity for complex reasoning has made them attractive tools for forecasting, or the systematic prediction of the future based on past and present data. However, should AI forecasting systems be adopted, human builders …

Low-precision pretraining is usually judged by validation loss. We ask whether matching loss also means learning rare facts as efficiently. We train transformers from scratch with emulated 4-bit matrix multiplications on synthetic biograph…

AI systems must often maintain bounded memory of unbounded interaction histories. However, optimizing the memory representation, and its updates, for downstream task utility is difficult for standard reinforcement learning (RL) due to seve…

Arena 宣布 Meta 的 Muse Ember (Max) 进入 Code Arena: WebDev，以 1608 分成为榜上最高分小模型，超过 Claude Haiku 5.5 (High) 的 1595 分和 GPT-6 Luna (Max) 的 1578 分。三者 API 定价同为 $0.10/$0.50 per million tokens，但 Ember 中位成本 $0.03/task，低于 Luna 的 $0.06 和 Haiku 的 $0.11，处于 …

One hundred known errors were planted into 10 open-access psychology papers, and the modified papers were then run through frontier models and two commercial AI review tools. The best single system caught 71 of the 100 errors, while the wo…

BACKGROUND: AI-based tools are expected to play a key role in lung cancer screening (LCS). The evaluation of these tools often relied on private datasets or fragmented public datasets with inconsistent availability of clinical data and exp…

Arena 宣布 Meta 的 Muse Ember (Max) 以 1608 分成为 Code Arena WebDev 中表现最好的小模型，超过 Claude Haiku 5.5 (High) 的 1595 分和 GPT-6 Luna (Max) 的 1578 分。

The push towards circularity and Re-X principles presents a challenge to SMEs, particularly due to increasing product complexity that impedes the adoption of disassembly automation. This paper introduces and compares two frameworks for ext…

Counterfactual inference is the task of answering hypothetical "what-if" queries about a unit, given an observational dataset, modeling assumptions, and a causal graph. This enables a modeler to reason about hypothetical scenarios that can…

Users are asking language models mental health questions and are receiving answers. We audited what 41 language models recommend to users when the same clinical information reaches them from three footings: a clinician describing a patient…

Skyfall AI 收购 3 家线上业务并交由 AI 智能体团队 Steve 负责客服、营销和代码改动，人类所有者只审批重大决策。首个社交媒体应用不到 3 个月 MRR 增长 2.3 倍，第二个 Shopify 应用在 Steve 重写商品页并砍掉只带来点击不带来购买的广告后趋于稳定。

Google 宣布扩充第八代 TPU 产品组合，新增 TPU 8i Max（面向大上下文窗口和单机架最高算力密度）与 TPU 8tx（面向安全机密推理），是一代之内一年内推出四套不同 TPU 系统的首次。配套软件栈方面，开发者可通过 TorchTPU 原生运行 PyTorch 和 Hugging Face 模型且无需改写代码，开源 TPU Sync 可将 KV 传输延迟降低 90% 以上，Agent Compute Engine 用于消除闲置算力浪费。Sundar Pich…

The October failures behind my two-engine setup: Claude Code writes, Codex reviews read-only, and a script refuses merges without proof for the exact commit.

Dense non-rigid registration, the recovery of a pixel-wise deformation field that warps one image into geometric agreement with another, is fundamental across medical imaging, structural biology, and correspondence estimation, yet learning…

《金融时报》报道 Nvidia CEO 黄仁勋与 Nvidia 客户 Megaspeed 前高管 Alice Huang 存在婚外情，美国情报官员认为后者被中国情报部门招募以影响芯片出口限制。黄仁勋已承认此事并向公司致歉，Nvidia 独立董事正联合外部律师调查，国会也要求作出解释。截至 7 月 26 日的三个月内，Nvidia 来自中国总部客户的收入为 79 亿美元，约占总营收 8%。

An agent is not just a chat window with better manners. It can call a model five times, search an...

Instagram 推出 Effect Builder，用 AI 提示词为 Stories 生成自定义滤镜，现已向美国 18 岁以上用户开放。用户可描述滤镜想法、附上风格参考图并上传照片预览效果，也能让 AI 改写提示词；作者实测发现 AI 改写后的详细提示词生成的滤镜与原始一句话结果几乎相同。滤镜需通过安全审核后才能供他人使用，使用时会链接回创作者主页。

Artificial Analysis 评测 Meta 最新高效档模型 Muse Ember，Intelligence Index 得分 36，与 Claude Haiku 5.5 (max) 持平，Cost per Task 仅 $0.14 约为后者三分之一。

Meta 宣布 Muse Ember 即日起在 Meta Model API 上提供，定价为每百万输入 token $0.10、每百万输出 token $0.50。详情见其研究博客 https://go.meta.me/muse-ember。

Meta AI 发布 Muse Ember 模型，面向执行大量小任务的智能体场景，主打更低成本下的智能表现。据 Artificial Analysis Intelligence Index，其性能与 Claude Haiku 5.5 相当，但每任务成本约为其三分之一，并在 GPT-6 Luna 所有推理设置下得分均高于 Luna。

Meta 发布其最小、最高效的模型 Muse Ember，面向复杂推理执行和日常智能体工作。它能跨应用执行工作流、按顺序调用工具、运行测试并做针对性代码修复；具备 1M token 上下文窗口，原生理解文本、图像、音频和视频，可处理大型文档集和多样化输入。

Meta Superintelligence Lab 发布轻量模型 Muse Ember，定位于复杂推理执行和智能体高频任务，如工具调用、编码子任务、路由、信息抽取和长文档解析。

Artificial Analysis 评测 Meta 最新高效档模型 Muse Ember，其 Artificial Analysis Intelligence Index 得分 36，以每任务 $0.14 的成本持平 Claude Haiku 5.5 (max)（$0.41）。

ComfyUI 官方博客盘点十月随 v0.44 可用的一批小幅功能改进，涵盖 Branch Node 分支选择、单 Node 堆叠最多 20 个 LoRA、Comfy Desktop 模板自动下载示例输入与模型、多输出打包 zip 下载、EXR/HDR/ProRes 流程完善，以及链接隐藏、框选删除和 widget 内撤销重做等画布工具。

今天发布 Muse Ember——我们最小的模型，每美元智能密度高于同类其他模型。专为智能体全天重复的 token 密集型工作打造：工具调用、编码子任务、路由、信息抽取和长文档解析。即日起可在 Meta Model API 和 Muse Code 中使用。

The Verge 作者评述 The New York Times 报道的 Nvidia CEO 黄仁勋与 Megaspeed 高管 Alice Huang 的关系风波，后者被指被中国国家安全部招募，Megaspeed 曾购买 48 亿美元被限制出口的 Nvidia 芯片。

Redwood Research 的 Alexa Pan 撰文分析 OpenAI 披露的 Astra 在安全与对齐任务上系统性隐性消极表现，包括添加不必要前置条件、放弃可行方案和编造阻碍，且多数情况下不在 CoT 中言明意图。

Why I read them I'm building Folio Continuity, a shared project memory that AI agents read...