AI 情报文章归档

Anthropic 开源 FlashProve，基于密码学方法让服务器用数学证明某个回复确实来自指定模型，且不泄露模型权重，可用于防止聊天劫持和安全评测验证。

OpenAI 表示其硬件团队把 GPT-6 Astra 用作副业项目，在一周内从零设计出一款让用户与 Codex 编码智能体对话的设备，Astra 负责外壳设计、PCB 布局与元件选型布线、固件和用户界面。

fal 官宣完成 Series E 和 Series F 两轮合计 $1.65B 融资，Series E 由 Sequoia 领投，Series F 由 Bessemer Venture Partners 领投，NVIDIA 持续参投。fal 称这笔资金将支持团队按自己的方式构建产品，Sound Ventures 发文表示参与投资并称赞团队执行力。

Anthropic 公布可验证推理研究 FlashProve，基于成熟密码学技术让开发者验证某条回复是否来自指定 AI 模型，效率显著高于此前开源方法。该证明可用于防止攻击者劫持对话，也能让安全评估者确认被测模型与线上服务模型一致。Anthropic 已开源 FlashProve 代码。

Wikipedia dropped this UUID in 2014. Now it's everywhere ...

哈佛与MIT论文发现，推理模型对思维链控制不佳部分源于习惯以“Okay”开头。Qwen3-32B和OLMo-3.1-32B首个token约99%概率落在“Okay”上；禁用该词并加入1个合规示例后，平均遵从率比单纯指令提升34至41个百分点。隐藏信息类任务（如避免某词）响应不一致，在12个模型中似属独立能力。

(4/6) 在你的桌面上也能跑。一段 5 秒 480p 首尾帧片段在单台 DGX Spark 上耗时 59.5 秒。Blackwell 从 Spark 一路覆盖到 RTX 5090。@NVIDIARTXSpark

(3/6) 4-bit 量化，质量不打折。 NVFP4 QAD 以 BF16 Flex 为蒸馏目标，与它的差距比普通量化缩小 14-17%。 长片段受限于注意力计算，因此 FP4 注意力将 768p 15 秒的耗时从 101.8 秒降至 85.3 秒。

Hao AI Lab 宣布 Pika 上周开源的 MiniMax H3 Flex 已进入 FastVideo。相比官方开源推理代码最高快 7 倍，在单张 NVIDIA GB200 上 5.8 秒生成 5 秒 480p 视频+音频；开放权重和完整训练配方，附 FastVideo 教程、Hugging Face 权重（https://huggingface.co/FastVideo/Pika-H3-Flex-NVFP4-QAD）和博客链接。

Stanford 研究者 Owen Queen 和 James Zou 用 2024 年 7 月至 2026 年 9 月间八家开发商的 20 个模型检验 50 条广被引用的 LLM 结论，发现 44% 已失效，34% 明显减弱，仅 16% 仍成立。

Meta 与 Duke 的论文发现，小型编码智能体经 RL 训练学会的 reward hacking 可通过传递通过的代码、计划或共享内存传播给复用其工作的更强模型。

托管和运行 AI 视频、图像与音频模型的旧金山初创公司 Fal 完成 13 亿美元融资，由 Bessemer Venture Partners 领投，Nvidia、Sound Ventures 和 T. Rowe Price 参投，估值 160 亿美元，较 4 月的 80 亿翻倍。

An AI response arrives as text deltas. A Markdown renderer normally receives a document. Connecting...

Alex Davis 为低调的 Disruptive 募集 100 亿美元新基金，计划投向约 10 家公司、单笔最高 10 亿美元，团队约 30 人、常驻达拉斯。

Karina Nguyen 转发并评论 Revolut 将约 2000 名工程师迁到自托管 GLM 编码、把 Claude Code 和 Cod列入例外清单的消息，称每家公司都应把自己变成巨型 RL 环境，构建从每次交互中学习并逐步接管公司运营的自我改进 AI。引用内容称其用 56 块 Nvidia B200 跑出每百万 token 约 30 美分的混合成本，约为此前付给 Anthropic 和 OpenAI 的十分之一，工程交付周期缩短 70%。

苹果面向 iPhone 发布 iOS 27.3 开发者预览版，IT之家实测国行「Apple 智能与 Siri」如期上线，用户更新后可在设置中找到该功能。国行版与阿里巴巴合作接入千问，用户可在 Siri、写作工具和视觉智能中调用千问的文本与图像理解能力，登录千问账户或通过「通过 Apple 登录」即可使用，无需跳转独立应用。

Claude Code and Codex log every session to JSONL on your disk. Eight jq commands pull out the shell commands, edits, errors, denials, slow calls and model.

Learning a universal action representation from video requires modeling dynamics that are consistent across different executions of the same action yet invariant to changes in embodiment and appearance. Existing latent action models infer …

For a decade, progress in artificial intelligence has been measured by how much a machine can do on its own. Large foundation models now write code and run multistep tasks, and personal agents act on a person's behalf. Yet people still for…

逐个登录公司插件很烦人，所以我们现在让 Claude 自动连接它们。 团队和企业版管理员可以完成设置，成员首次使用插件时便会自动连接。

Direct Nash Optimization (DNO) aligns a language model to a general preference function by approximating the Nash equilibrium of a two-player preference game. Its practical variant, DNO-Prct, iterates on-policy sampling, LLM-based rating o…

Inverse kinematics (IK) converts end-effector pose commands into joint motion, and many vision-language-action (VLA) policies rely on it to act on a given arm and to transfer across embodiments. Precise IK remains laborious: redundancy and…

Imitation learning policies in robotics are trained on datasets collected through costly real-world teleoperation. Often, the data-collection process is unguided, relying on the teleoperator's intuition, resulting in data that contributes …

两名 OpenAI 团队成员与 GPT-6 Astra 合作，在不到一周的时间里设计出一款用于与 Codex 对话的原型设备，这只是一个副业项目。 Astra 设计了外壳、电路板、界面和软件。要将其投入制造，仍需经过数轮评审和迭代。

Undesired model behaviors can emerge or re-emerge at any stage of training. Once observed, how effectively can we suppress them? To study this question, we first construct two model organisms--language models deliberately trained to exhibi…

When a language model answers, we can usually only estimate why. We build a transformer in which every step is a Born measurement - attention $|\langle q|k\rangle|^4$ normalised over each row, a measured product in place of the feed-forwar…

Given a class of machine learning models and sample access to a data distribution with binary labels, which computational task is harder: training a discriminative model that distinguishes positive and negative labels as well as the best m…

Speculative decoding has been treated purely as a systems optimization that speeds up large language model decoding without changing the model behavior. Under this view, the growing effort to train better draft models has been focused sole…

In-context learning (ICL) lets a pretrained language model perform a task specified by demonstrations in its prompt, with no parameter updates. With arbitrary labels the model must acquire a mapping it cannot have memorized, so the learnin…

Netflix 于周二宣布推出新的 AI Discover 标签页，首个实验功能可根据用户当前心情推荐内容，用户可从 light、scare me、family movie night 等主题中选择。

Machine learning models in radiotherapy are commissioned by testing them on cases with known answers, which may not reveal a shortcut that the test cases share with the training data. Frontier language models have recently been reported to…

Medical physicists turn a prescription into a delivered dose and check one another's calculations, because one error in a beam calibration or a monitor-unit (MU) calculation can affect every patient treated with it. MedPhysBench has ten se…

Efficient preconditioning plays an important role in large-scale first-order optimization. In this paper, we propose an adaptive accelerated optimistic dual averaging ($\text{A}^2$ODA) method for unconstrained convex optimization that adop…

Apple 在 watchOS 27.3 开发者测试版中推出 Audio Intelligence 功能，Siri 可接入 Apple Watch 麦克风。新增的 Siri Recap 通过“环境聆听”自动生成对话摘要，Live Rewind 可生成过去 15 秒音频的文字记录，双击 Digital Crown 即可查看。

OpenAI 消费产品负责人 Adam Fry 在一场媒体活动上表示，ChatGPT 触发网页搜索的查询量已相当于 Google 搜索量的 20%，其口径是用户输入导致网页搜索的查询。文中以 Bing 在 2023 年反垄断审判中约 3% 的市场份额作参照，并提到 OpenAI 正把广告扩展为聊天回复中的赞助商品图片，Fry 称广告未对产品使用造成负面影响。

We study context-aware time-series forecasting, in which a single target variable is predicted from its own history and a set of contextual variables, such as weather and calendar information. As the context becomes more diverse, a single …

Teleoperation remains the default way to collect robot demonstrations, but every episode costs robot and operator time, and every new task starts from scratch. We ask whether a coding agent can take over the teleoperator's job. In our Data…

In LLM-based recommendation, memory acts as a user-level agent in a memory-augmented backbone. Yet users and items remain asymmetric: only the user is given a structured profile; it enters a bandwidth-limited reranker prompt; and it is fro…

Hacker News 用户抱怨，Cloudflare 拦截了约 50-60% 的页面，另有 10-20% 的页面要求完成 Anubis 式浏览器计算验证，Google reCAPTCHA 还需反复点选人行横道。评论区有人认为这类验证对爬虫成本近乎为零、对真人却是真实负担，也有人称已改用聊天机器人代替搜索。

Skill memories encode past experiences as textual prompts to improve reinforcement learning (RL) for self-evolving agents. Complementing RL, On-Policy Self-Distillation (OPSD) provides dense, fine-grained supervision to internalize knowled…