春秋带剑 | 公众号 chunqiudaijian
1. 新闻
1. Anthropic推出个人智能体Claude Duet
10月22日凌晨上的。
这些AI智能体主打的卖点是，以前的AI得你跟它说它才能懂，现在个人智能体会主动帮你干活，帮你整理你的工作、生活一切内容。现在Meta有Muse，Manus有Cue，OpenAI有Dots，马斯克有grok bot，国内有阿里千问Go，蚂蚁灵光2.0开启内测，这已经超越了三国杀的程度，马上就能快进到五代十国了！
参考资料：AI智能体三国杀：Claude Duet凌晨突袭上线
2. Anthropic两名重要员工于10月24日离职
参考资料：突发！Claude Code痛失两员大将
3. Anthropic正式交表IPO
笑点：5%到7%的收入，来自违规客户
其它细节：
- 上市落地后，5300名员工将分得297亿美元的股票；两位创始人已提前套现46.6亿美元。
- 达里奥将持有的特殊股份，控制50.1%的投票权。
- 今年第二季度，Anthropic已经实现营业盈利。
- 公司代码库超90%的提交是Claude写的，人均年创收拉到660万美元。
- 超8300亿美元的算力开支。
参考资料：Anthropic正式交表IPO，7%收入来自违规账号。。
4. 10月26日，马斯克分享了由其旗下Grok Imagine制作的AI生成版《奥德赛》第一集
该系列计划推出多集。马斯克7月曾承诺年底前用Grok Imagine制作一部"在历史上准确无误"的完整《奥德赛》电影。外媒指出，从视觉效果看，该视频集中暴露了当前生成式AI视频的典型缺陷：牙齿完美得不自然、头发质感轻盈得几乎像水、皮肤呈现奇怪的模糊感（源于模型对光照和透视理解不足），背景存在纹理弹出等问题。
5. OpenAI和Anthropic分别发表论文证明ETH（指数时间假说）是错的
但是时间还是很长，所以在实践上暂时不用担心加密算法被迅速破解，但在学术理论上，很多论文的理论基础都被推翻了。
理论计算机学者正在组织验证。
参考资料：一文看懂OpenAI刚刚捅出的计算机界大篓子
6. 谷歌发布世界模型Genie 4
通过语言描述和当前画面，Gemini可以引导Genie 4直接生成物理世界的运动效果动画（包括音频）
暂时还没有开放使用。
参考资料：Google 重磅发布 Genie 4：世界模型终于活了
7. 谷歌发布新语言模型Gemini 4 carbon，重回AAII（Artificial Analysis Intelligence Index）第一
北美豆包又站起来了！
现在新模型排名都咬得很紧啊，感觉每隔多久就会听说一款新的模型霸榜。感觉大家都没有真正形成什么护城河，就是砸算力砸数据，氪金出奇迹。
暂时还没有开放使用。
8. 智元从10月22日起连续七个工作日每天发布具身智能全栈技术
10月22日：具身作业数据集RealWorld-7M
10月23日：作业平台Orchestrator OS
10月26日：安规级重载具身机器人精灵G2 Max，面向拆码垛、物料转运、产线供料等场景
10月27日：具身Agent框架GenieMaster
10月28日：仿真平台MirrorWorld Pro 3.0
10月29日：基座大模型Voyager-3B
10月30日：群体经验学习范式（PEL）白皮书
11月2日至3日将进行机器人不间断直播作业。
参考资料：
9. 重庆通用人工智能研究院发布明镜平台和带有社交能力的通用具身智能体通通
还有说价值观什么什么的，我感觉是通稿
参考资料：从屏幕走进真实场景：一场大会照见通用人工智能新未来
10. 苹果在MacBook Pro上引入触控屏
笑点是2010年乔布斯本人在发布会上说触控屏在演示时看起来很酷，但实际用不了多久就会疲劳，时间一长，手臂像是要断掉一样。
参考资料：16年前的苹果发布会名场面翻红！乔布斯嘲讽触控屏笔记本 台下笑成一片_笔记本电脑_什么值得买
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
WeChat Official Account cunqiudaijian
1. Public information
1. Anthropic presents a personal AI agentClaude Duet
On the morning of October 22nd.
The main point of the AZAI agent is that the former AI needs you to tell it to understand, and now the individual AI agent will help you work, sort out your work, everything in your life. Now Meta has Muse, Manus has Cue, OpenAi has Dots, Mask has Grok Bot, Ali has asked Go, an antslight 2.0 on the inside, which is beyond the triple killing and will soon reach the five generations of ten!
Reference: AZAI agent Trinational Kill: Claude Duet hit the line early in the morning
2. Two key employees of Anthropic separated on 24 October
Reference: Sudden! Claude Code lost two generals.
3. Anthropic official submission IPO
Smile: 5 to 7 per cent of income from irregular clients
Other details:
- Upon landing, 5,300 employees will be given shares of $29.7 billion; the two founders have prepaid $4.66 billion.
- Darío will hold a special share and control 50.1 per cent of voting rights.
- In the second quarter of this year, Anthropic has made a profit.
- Over 90 per cent of the submission in the company code repository was written by Claude, with an annual per capita income of $6.6 million.
- Over $83 billion in computing expenses.
Reference: Anthropic officially submitted IPO, with 7 per cent of income derived from irregular accounts.
4. On 26 October, Mask shared the first edition of the AI-generated version of " Odyssey ", produced by Grok Imagine under his flag.
The series is being launched in multiple volumes. Mask committed himself in July to producing a complete "History-Accurate" "Odyssey" film by the end of the year. In terms of visual effects, the video highlights typical deficiencies in the current generation of AI videos: perfect teeth, light hair-massing almost like water, strange ambiguity in the skin (resulting from poor model understanding of light and perceiving) and problems with texture ejection.
5. Open AI and Anthropic published separate papers proving that ETH was wrong.
However, it was still a long time, so there was little to worry in practice about the rapid breakdown of encryption algorithms, but in academic theory the theoretical basis of many of the papers had been overturned.
Theoretical computer scholars are organizing validation.
References: A book that understands what OpenAi just pulled out of the computer world.
6. Google publish world model Genie 4
Gemini can direct Genie 4 directly to the physical world by describing and current images (including audio)
For the time being, it is not open.
References: Google, weighing publish Genie 4: The world model finally survived.
Google publish new language model Gemini 4 carbon, back to Artificial Analysis Intelligence Index first
North America Doubao stands up again!
Now, the rankings of the new models are very tight, and it feels like every other time a new model is heard. It's like you're not really forming a moat, it's a calculus. Kryptonian is a miracle.
For the time being, it is not open.
8. publish Smart Holistic Technology for seven consecutive working days from 22 October
22 October: PersonalWorld-7M
23 October: OPERATIONAL OS
October 26th: Ambient heavy load of robotic elf G2 Max, for decorative scenes, trans-shipment of materials, wire supply, etc.
27 October: Genee Master with the Agent Framework
28 October: MirrorWorld Pro 3.0
29 October: Base LLMVoyager-3B
30 October: Group Learning Model (PEL) White paper
Uninterrupted live robotic broadcasting will take place from 2 to 3 November.
References:
9. Chongqing General Artificial Intelligence Institute publish Mirror Platform and Socially Capable Universal AI agent Communication
And values or something. I feel like a draft.
References: Moving into the real scene from the screen: a congress to see the new global artificial intelligence future
10. Apple introduces touch screen on MacBook Pro
The joke was that Jobos himself said at the publish conference in 2010 that touch screens looked cool at the demonstration, but that it would actually take a little longer to get tired, for a long time, as if his arm was to be broken.
Reference: Apple publish was famous 16 years ago! Jobos mocked the touch screen notebook and laughed under the stage into a laptop.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.