谷歌发了 Gemini 4 Carbon，在 Artificial Analysis 智能指数 v5 上 59 分排第一，压过了 Claude Opus 5.5 和 GPT-6 Astra。离上一版 Argon 才两周。不过现在只开放给少数网络安全合作方，普通人还用不上。 blog.google/innovation-and-a… 两周一个大版本，这节奏有点离谱

Higgsfield 出了 Cinema Studio 4.0，一次能生成 30 秒的多镜头场景，最多喂 50 张参考图。上一版 3.5 还只有 15 秒、9 张参考。 prnewswire.com/news-releases… AI 做短片越来越像在当导演了

德国初创 NVision 在 Nature 上发了单分子量子比特：一个专门设计的有机卡宾分子，用光初始化和读出、用微波操控，量子态能撑到毫秒级。筛了 150 多个候选分子，合成了 40 个才挑中它。 nature.com/articles/s41586-0… 以后量子芯片像 OLED 屏那样镀一层分子膜，这思路挺野的

Axiom put the classification of finite simple groups into Lean. The original proof runs ~15,000 pages. AI agent swarms wrote ~43.1M lines of Lean in 24 days and caught a flaw in one published argument. axiommath.ai/research/the-en… The Monster, kernel-checked

Quantinuum 和 GE Vernova 开始合作，要在 Helios 离子阱机器上跑 Fermi-Hubbard 模型，研究高温超导是怎么来的。之前他们已经在 Helios 上测到过 s 波和 d 波配对关联。GE 惦记的是以后更轻的发电机、电机和电缆。 quantinuum.com/blog/quantinu… 高温超导机制卡了快四十年，拿量子模拟去啃挺对路的

SANXINGDUI is the first AI feature film with a fully AI-generated cast to get a theatrical permit in China. 100 min of sci-fi from Bona Film Group, made with Vidu, no real actors or digital doubles. globenewswire.com/news-relea… Curious how it holds up on a big screen

百度开源了 Unlimited-OCR-2，3B 参数、MIT 协议，上百页的文档一次喂进去直接解析。在他们新出的百页级测试 MarathonDocBench 上 77.1 分，Gemini-3.8-Flash 是 28.6，Qwen3.8-Max 是 22.9。 huggingface.co/baidu/Unlimit… 测试集是自家出的，不过 3B 能跑成这样挺香

Anthropic secretly gave Claude Code a coin-flip lean on 4 judgment calls, then handed it to 50 expert analysts. They went with the agent's lean 86% of the time, and 77% later explained their picks in its words. anthropic.com/research/agent… Human in the loop, sure

Anthropic 也发了篇作者栏写着 Claude 的论文：3-SAT 有 2^o(n) 时间的确定性算法，直接推翻指数时间假设 ETH，Lean 4 已验证。核心算法是内部模型 10 月 16 日自己找到的，没人插手。 hexagonmath.org/2610.00577v1 跟 OpenAI 撞车了，纯渐近结果，实际用不上，但这个月理论 CS 是真热闹

蚂蚁开源了 Ling-3.1-flash-1019，MIT 协议，560B 总参数、每 token 激活 25B，1M 上下文。只是在原版上多做了一轮后训练，Artificial Analysis 智能指数就从 41 涨到 44，跟 DeepSeek V4.1 Flash 1017 只差 1 分。 huggingface.co/inclusionAI/L… 国产开源这一档现在挤得不行

Liquid AI's new LFM3 is a family of ~2B open-weight models, and the text one fits in about 1 GB at 4-bit. On their 65 everyday Mac agent tasks it beat Qwen3.5-2B 64% to 48%, and got stuck <1% of the time vs 15%. liquid.ai/blog/lfm3 Laptop-sized agents are getting real

Google DeepMind's Genie 4 is out in research preview. Real-time playable worlds that now make their own audio, characters included, and you can type mid-sim to change the weather. Scenes hold ~2 min, double Genie 3. deepmind.google/blog/genie-4… Still drifts on long runs, they admit

商汤开源了个 27B 的视频模型 SenseNova 6.8 Code2Video，思路挺特别：不直接生成像素，而是写代码再渲染成片，想改颜色、节奏就用大白话跟它说。基于 Qwen3.8-27B，Apache 2.0，官方说能连跑 10 小时、2000 多轮。 huggingface.co/sensenova/Sen… 做科普动画的应该会喜欢

Pasqal's 256-atom array watched a 2D Ising magnet relax way slower than expected after a quench, right where three classical methods stop agreeing. They say they can't explain it yet. arxiv.org/abs/2608.07178 Not an advantage claim, but this is the fun part of analog simulators

OpenAI 一口气放出三个大结果： 1. 局部光滑猜想全维度成立，顺带拿下 Kakeya 猜想 2. 指数时间假设 ETH 被证伪，Lean 已验证 3. 严格证明 4D 的 QCD 和纯 Yang-Mills 存在 openai.com/index/building-sc… 第三个学物理的看了真有点震撼，不过千禧年奖还差质量间隙，等 Gross 他们的结论

中国科大在单个钙离子里做了自主量子纠错：用离子的多个能级编码逻辑比特，错误的熵转到振动模式里，再靠旁边的辅助离子冷却带走，不用测量也不用反馈。加噪声测试下逻辑比特寿命 11.6ms，物理比特才 0.9ms。 nature.com/articles/s41567-0… 离子阱的玩法是真多

Anthropic is reportedly looking to raise up to $100B, with the roadshow starting Nov 9. But Q3 revenue growth slowed to 45% q/q ($16.8B vs $11.6B), and some investors have reportedly cut valuation talk from ~$2T to ~$1.5T. investing.com/news/stock-mar… Even the "discount" is $1.5T

Kimi K3.1 Pro 发了，月之暗面这次干了件挺少见的事：故意把网络安全相关数据从训练和 RL 环境里删掉，说不想开源一个把网络攻击能力推到最前沿的模型。自家漏洞挖掘测试 81.0%，不删的版本是 90.5%。权重 11 月 4 日前放出。 kimi.ai/blog/kimi-k3-1-pro 开源厂商主动给自己降一档，有意思

CapCut is coming to ChatGPT as a plugin. Rough out a video idea in chat, CapCut builds a first cut, then you fix it by talking: hold this shot longer, keep the music under my voice. Coming in the next few months. capcut.com/newsroom/capcut-w… Timeline editing by chat, finally

做 AI 视频的可以看看这句：Anthropic 在招股书里明说，他们主动不做图像和视频生成模型，把算力留给研究和安全。上半年光算力和基础设施就花了 89 亿美元。 sec.gov/Archives/edgar/data/… 视频这块短期别指望 Claude 了，还得看 Veo、可灵、Seedance