Genie 4 Worlds, Frontier Code 2, AI 药物发现

TLDR AI 2026-10-28 (中文(简体) ).

Genie 4 Worlds, Frontier Code 2, AI 药物发现

与Human-in-the-Loop一起设计了更简单的AI – Agents 要求修订监督框架(赞助商)

快速，暂停，审查是罚款处理大多数基因AI。 但检查agent制造的上百个微量决定中的每一个，并不仅仅是不切实际的——它会因为一开始使用agent而失去任何节省的时间。

这个eBook by Sevener 推出是管理AI agents的人力控制框架，帮助各组织在AI从飞行员转向生产的过程中平衡自主性，准确性和问责制。 你会学会的：

- 为什么传统的HITL模式在AI变得更加自主时破裂？

- 可扩展AI监督的人力控制框架

- RLHF、积极学习和agent监督等技术

- 实施下一代hiltL的实用蓝图

拿电子书来

Google 推出 Genie 4 (5分钟读取)

Genie 4是一款能产生相片现实化，互动环境的世界模型，具有同步音频和实时即时编辑功能。 它可以保持相容的模拟时间可长达两分钟，将Genie 3的持续时间翻一番，同时能改进物理和人物现实主义。

用于设备AI的液态AI 发布 LFM3型号(27分读取)

LFM3是一个由开放的重量文本，视觉，音频，omni，和决定模型组成的家族，构建在一个共享的三重主干线上，从2B参数范围开始。 1.7B文本模型以35%的参数来击败其2.6B前身，以4位精度相适应于约1GB，并支持为在设备上运行的agents提供256K上下文窗口。 LFM3-VL-2B 新增视频输入并匹配了Gemma 4 12B跨62视觉的基准测试，以及家族中每个型号的微调在一个单一的GPU上。

将定理正式化(27分钟读取)

Axiom Math拥有发布，这是Finite Simple Groups(CFSG)分类法(CFSG)的第一个完整的计算机检查证明。 最初的CFSG证明被认为是人类所写的最长的数学证明。 它在由大约100名数学家所撰写的500多篇文章中相接了约15,000页。 Axiom Math的工程师和数学家在24天内指导了数十个AI agents的群，在一个在Lean验证助手中产生了大约4,310万行代码的项目中。

Anthropic Versus Anthropic (17分钟读取)

Anthropic正在努力平衡大胆的安全倡议和令人失望的行动决定。 首席执行官Dario Amodei一直捍卫着他们昂贵的、面向安全的战略，然而，人们对一致性、治理以及AI开发的潜在风险仍然感到关切。 它们的内部文化既反映了雄心勃勃的努力，也反映了传统商业方法的令人不安的迹象，引发了关于它们对AI安全和全球责任的承诺的辩论。

世界的AI芯片在哪里？(1分钟读取)

AI芯片进口探险家估计自2022年以来AI按国家计算抵达，加强了全球AI芯片供应链的绘图。 它补充了关于AI芯片组件、销售和数据中心的现有数据。 这些资源共同详细概括了大赦国际在全世界的计算分配和控制。

👨💻

工程和研究

控制一个agent很难。 当是agent舰队时会发生什么？(发起人。

AI agents获得更多的自主权。 他们获取企业数据，使用工具，采取行动，把传统的治理留在尘中。 11月4日，AIUC和Collibra专家将展示如何应用AIUC-1标准，监督直播agents，并在他们下线时进行干预。 学习如何保持控制作为你的agent舰队的尺度。

主动视频 Agents (22分钟读取)

StreamVibe-VL是一种视觉语言模式，它从连续观测中推断出隐含的用户用意，并随着这些用意的变化而适应。

Anthropic's Claude Refute the Expressional Time hypothesis, with a Lean proof (GitHub Repo) 互联网档案馆的存檔，存档日期2014-12-22.

Anthropic 称是一个内部研究的Claude模型，在无人输入的情况下，发现了一种确定性算法，它计算出3-SAT公式在2^O(n log n/√log n)时间内的满意任务，这将使3-SAT子责任人和反驳ETH. 这个寄存器持有由AI模型所撰写的纸和"精益4"形式化，约8分钟后建成，虽然Anthropic 指出没有机器检查"精益声明"是否与纸相匹配，结果称没有关于实际SAT解析。 在OpenAI自己的ETH反驳数小时后降落。

Meta的可穿戴AI互动模拟器(GitHub Repo)

InteractiveWorld是培训和评价可穿戴的AI助手的模拟框架。 它模拟一个助手通过用户眼睛观察的交互，与他们交流，并接收更新的第一人称视图。

睡眠框架：LLMs在持续学习中保留知识的新方法(读取11分)

研究人员开发出一种新的方法，使大型语言模型能够整合出新的事实并增强推理，而不会发生灾难性的遗忘或静态的知识截断。 睡眠框架为大型语言模型提供了连续的周期性生命周期，使其能够不断学习并巩固脆弱的短期环境，成为稳定的长期参数。 在活性互动和离线内存整合之间进行替换，可以使LLMs不断获取新事实，将长文本知识保留到10M令牌，并改进复杂的推理。

FrontierCode 2.0 (13分钟读取)

FrontCode 2.0提供了更广泛的覆盖范围，更高的可合并性标准，以及更大的任务范围，同时在很大程度上保留了FrontCode 1.0.的方法。 基准测试现在容纳出人意料但可接受的行为。 任务来自更多的储存库，涉及更多的类别，分级标准现在规定了更广泛的正确性和设计要求。 与FrontierCode 1.1相比，参考解决方案现在平均修改了67%的文件和89%的代码行，其代码基数为25%。

Anthropic的药物开发过程(4分钟读取)

Anthropic与6个制药和生物技术组织合作，探索Claude如何从早期研究到临床试验准备等加速药物开发。 它还针对罕见儿童疾病的发布内部药物发现方案，Claude处理大量的科学和操作工作。

Anthropic 收购单人生物技术(4分钟读取)

Anthropic拥有由科迪·特兰巴格(Cody Tranbarger)经营的一人一生物技术收购。 Mithril Medicine是一家创办于使用AI开发疾病治疗方法的起步企业。 加入Anthropic意味着Mithril能够以更多的资源继续运行其程序，这可以让Tranbarger启动多几个程序。 他们的目标是扩大初创企业的新工作方式，然后将其置于全世界所有生物制药的手中。

JEV帮助你的agent决定下一步该怎么做。 Bluehost给了它一个24/7的跑步地点

你的agent不应该依赖你的笔记本电脑开着以99.99%的超时保证率运行它，所以在关闭后继续工作

构建更智能的 Agents (赞助者)

探索Neo4j NODES的上下文图和代理内存，这是一次开发者会议，与安德鲁·恩格共同主讲100场现场技术会+基调。

免费在线 24小时

登记册

Nvidia将投资3亿元，用于Meta的"扩大AI伙伴关系"的股权提供(2分读出).

Meta打算加强Nvidia基础设施的发展，公司正在考虑在监测agent行为等其他领域合作。

比较向您的代码提交，解决散装冲突，保持编码( 5分钟读取)

10月为Visual Studio 推出更新了一个直接承诺比较功能，一个新的Copilot Chat体验来改进agent工作流程，以及一个热出选项来保存未保存的工作。

Glasser现在让AI agents购买每个API呼叫的保费数据(读取3分)

Glasser允许AI agents每次API呼叫购买保费数据，提供80+供应商2000多个端点的访问。

Gemini 机器人健美理性2现可访问Google Cloud (7分读取).

Gemini机器人健美理性2现可于Google Cloud上取用，将高级斯皮提奥-时空推理和机器人长视距规划结合起来。

获取最有趣的AI故事与突破

与 1 100 000 名读者一起每天发一封电子邮件

本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

TLDR AI 2026-10-28

Genie 4 worlds 🌍, Frontier Code 2 👨💻, AI drug discovery 🧬

Together With Human-in-the-Loop Was Designed for Simpler AI – Agents Demand a Revised Oversight Framework (Sponsor)

Prompt, pause, review was fine for dealing with most generative AI. But inspecting each of the hundreds of microdecisions an agent makes is not just impractical – it loses any time saved by using the agent in the first place.

This eBook by Seekr introduces a human-in-control framework for governing AI agents, helping organizations balance autonomy, accuracy, and accountability as AI moves from pilot to production. You'll learn:

- Why traditional HITL models break when AI becomes more autonomous

- The human-in-control framework for scalable AI oversight

- Techniques like RLHF, active learning, and agent supervision

- A practical blueprint for implementing next-generation HITL

Get the eBook.

Google Introduces Genie 4 (5 minute read)

Genie 4 is a world model that generates photorealistic, interactive environments with synchronized audio and real-time prompt-based editing. It can maintain consistent simulations for up to two minutes, doubling Genie 3's duration, while improving physics and character realism.

Liquid AI Launches LFM3 Models for On Device AI (27 minute read)

LFM3 is a family of open weight text, vision, audio, omni, and decision models built on a shared tri-hybrid backbone, starting in the 2B parameter range. The 1.7B text model beats its 2.6B predecessor with 35% fewer parameters, fits in about 1 GB at 4 bit precision, and supports a 256K context window for agents running on device. LFM3-VL-2B adds video input and matches Gemma 4 12B across 62 vision benchmarks, and every model in the family fine-tunes on a single GPU.

Formalizing the Enormous Theorem (27 minute read)

Axiom Math has released the first complete computer-checked proof of the Classification of Finite Simple Groups (CFSG). The original CFSG proof is considered to be the longest mathematical proof ever written by humans. It spans roughly 15,000 pages in more than 500 articles written by about 100 mathematicians. Engineers and mathematicians at Axiom Math directed swarms of dozens of AI agents over 24 days in a project that produced approximately 43.1 million lines of code in the Lean proof assistant.

Anthropic Versus Anthropic (17 minute read)

Anthropic is grappling with balancing bold safety initiatives and disappointing operational decisions. CEO Dario Amodei has defended their costly, safety-oriented strategies, yet concerns persist over alignment, governance, and potential risks from AI development. Their internal culture reflects both ambitious efforts and troubling signs of a conventional business approach, sparking debate about their commitment to AI safety and global responsibilities.

Where are the world's AI chips? (1 minute read)

The AI Chip Imports explorer estimates AI compute arrivals by country since 2022, enhancing the mapping of the global AI chip supply chain. It complements existing data on AI chip components, sales, and data centers. These resources collectively provide a detailed overview of AI compute allocation and control worldwide.

👨💻

Engineering & Research

Controlling one agent is hard. What happens when it's an agent fleet? (Sponsor)

AI agents are taking on more autonomy. They access enterprise data, use tools, take actions, and leave traditional governance in the dust. On Nov 4, AIUC and Collibra experts will show how to apply AIUC-1 standards, supervise live agents, and intervene when they go off-script. Learn how to stay in control as your agent fleet scales.

Proactive Video Agents (22 minute read)

StreamVibe-VL is a vision-language model that infers implicit user intentions from continuous observations and adapts as those intentions change.

Anthropic's Claude Refutes the Exponential Time Hypothesis, With a Lean Proof (GitHub Repo)

Anthropic says an internal research Claude model found, with no human input, a deterministic algorithm that counts the satisfying assignments of 3-SAT formulas in 2^O(n log log n/√log n) time, which would make 3-SAT subexponential and refute ETH. This repository holds the paper and a Lean 4 formalization written by AI models that builds in about 8 minutes, though Anthropic notes no machine checks that the Lean statement matches the paper and the result says nothing about practical SAT solving. It landed hours after OpenAI's own ETH refutation.

Meta's Interactive Simulator for Wearable AI (GitHub Repo)

InteractiveWorld is a simulation framework for training and evaluating wearable AI assistants. It models interactions where an assistant observes through a user's eyes, communicates with them, and receives updated first-person views.

The sleep framework: A new way for LLMs to retain knowledge in continual learning (11 minute read)

Researchers have developed a new way to enable large language models to integrate new facts and enhance reasoning without catastrophic forgetting or static knowledge cutoffs. The sleep framework equips large language models with a continuous periodic lifecycle, enabling them to continually learn and consolidate fragile short-term context into stable long-term parameters. Alternating between active interaction and offline memory consolidation allows LLMs to continuously acquire new facts, retain long-context knowledge up to 10M tokens, and improve complex reasoning.

FrontierCode 2.0 (13 minute read)

FrontierCode 2.0 provides broader coverage, higher mergeability standards, and a larger task scope while largely retaining the methodology of FrontierCode 1.0. The benchmark now accommodates surprising but acceptable behavior. Tasks are drawn from more repositories and span more categories, and grading criteria now impose more extensive correctness and design requirements. Reference solutions now modify 67% more files and 89% more lines of code on average compared to FrontierCode 1.1, in codebases that are 25% larger.

Anthropic's Drug Development Process (4 minute read)

Anthropic partnered with six pharmaceutical and biotechnology organizations to explore how Claude could accelerate drug development, from early research to clinical trial preparation. It also launched internal drug discovery programs targeting rare childhood diseases, with Claude handling substantial scientific and operational work.

Anthropic acquired a one-person biotech (4 minute read)

Anthropic has acquired a one-person biotech run by Cody Tranbarger. Mithril Medicine is a startup founded to develop cures for diseases using AI. Joining Anthropic means Mithril will be able to continue running its programs with more resources, and it will allow Tranbarger to start several more programs. They aim to scale the startup's new way of working, then put it in the hands of all biopharma, all over the world.

JEV helps your agent decide what to do next. Bluehost gives it somewhere to run 24/7. (Sponsor)

Your agent shouldn't depend on your laptop staying open. Run it on an always-on VPS with a 99.99% uptime guarantee, so it keeps working after you log offGet started

Build Smarter Agents (Sponsor)

Explore context graphs and agentic memory at Neo4j NODES, a developer conference featuring 100 live technical sessions + keynote with Andrew Ng.

Free. Online. 24hrs

Register

Nvidia to Invest $3 Billion in Meta's Equity Offering as They Expand AI Partnership (2 minute read)

Meta intends to increase the development of Nvidia infrastructure, and the companies are considering working together in other areas like monitoring agent behavior.

Compare Commits to Your Code, Resolve Conflicts in Bulk, Keep Coding (5 minute read)

The October update for Visual Studio introduces a direct commit comparison feature, a new Copilot Chat experience for improved agent workflows, and a Hot Exit option to preserve unsaved work.

Glasser now lets AI agents buy premium data per API call (3 minute read)

Glasser allows AI agents to buy premium data per API call, offering access to over 2,000 endpoints from 80+ providers.

Gemini Robotics Embodied Reasoning 2 is now available on Google Cloud (7 minute read)

Gemini Robotics Embodied Reasoning 2 is now available on Google Cloud, integrating advanced spatio-temporal reasoning and long-horizon planning for robots.

Get the most interesting AI stories and breakthroughs delivered in a free daily email.

Join 1,100,000 readers for one daily email

Auto-fetched and translated by AI Gem Box. All rights belong to the original author.