12 Nov 2026. Google today turned AI judgment into a product category, launching the Gemini Decisions API alongside its first decision model, Gemini Decision 1, which scores 66 on the Artificial Analysis Decision Index v2.1 — 7 points behind leaders Drex Max v1.1 and GLiDE v1.2 at 73. The API generates no prose: it answers yes/no questions, multiple-choice with up to 255 options, and custom-tier scoring, each with a calibrated 0-1 confidence (calibration error 0.025, versus GPT-6 Luna's 0.072). A single request accepts up to 256K tokens of mixed text, JSON, images, audio, video and PDF, and can judge up to 1,000 questions per call, with input priced at $0.10 per million tokens. Separately, DeepMind's AlphaProof Code achieved the first code-level mathematical proof of the Signal protocol's security, and AlphaGenome 2 doubled DNA variant detection for gene expression effects.
Section Synthesis & Key Takeaways: Today's briefing curates 425 high-signal developments across 120 global and regional sources.
Models (5 picks): 「DeepMind releases AlphaGenome 2 DNA model with doubled variant detection」（The move from tissue-level averages to single-cell-type resolution changes what genetics labs can do with variant interpretation: rare-disease teams can now pinpoint which specific cell type a variant disrupts rather than relying on broad tissue signals. Computational biologists and genomics researchers can start testing it immediately via the free AlphaGenome API for non-commercial work. Compared with prior variant-effect predictors, the conditional architecture and cross-modal imputation reduce the need to run every assay before planning follow-up experiments）; 「Nvidia 发布 GR00T 2 世界动作模型，驱动机械臂与人形机器人」（机器人领域一直被操控数据卡脖子，Nvidia 这次用 40 万小时人类视频加两万多小时遥操作数据的组合，把训练成本降下来了。做具身智能或人形机器人的团队值得认真读一下预训练配方）.
Products (5 picks): 「Google 上线 Gemini Decisions API：一次最多判 1000 题，输入每百万 Token 0.1 美元」（做分类、路由、审核这类只需要判断的场景，现在可以不用解析大模型输出文本了。Google 这版一次能挂 1000 个问题还支持音视频输入，比 OpenAI 同名产品更激进，做客服分流、内容审核的开发者可以直接去 AI Studio 拿免费 Key 试）; 「Gemini 4 Carbon 上线 Databricks，OfficeQA Pro v2 得分 66.7%」（Databricks 用户当天就能用上 Google 这款主打编程和长任务智能体的新模型，OfficeQA Pro v2 成绩比上一代 Argon 高了近 9 个百分点，还能用 Unity Gateway 统一管用量和开销，做企业文档处理的团队可以关注）.
Industry (5 picks): 「Anthropic IPO 路演披露多芯片战略：Claude 已跑通四大芯片平台」（Anthropic 上市路演把家底摊开了讲：Claude 同时跑在 AWS、Google、NVIDIA 和 AMD 四种芯片上，接入 AMD 只花 3 天还是自家智能体 Mythos 干的活。想了解大厂怎么压低算力成本、以及“算力乘数”这个概念的，这篇值得点进去看原始转录）; 「Google DeepMind 用 AlphaProof Code 首次从代码层面证明 Signal 协议安全性」（AI 第一次把形式化验证做到了代码层面，被验证的还是 Signal 这种保护数十亿人消息的加密系统。做密码学、安全审计或者关心端到端加密的人应该看看，AI 定理证明正在从数学竞赛走向真实软件验证）.
Papers (3 picks): 「Apple Research Links Gating to In-Context Learning in State Space Models」（The non-obvious takeaway is that SSM weaknesses on retrieval may be fixable through gating design rather than requiring a wholesale return to Transformers. Model architects and ML engineers evaluating efficient alternatives to attention should study the paper's analysis to guide architecture choices. Compared with prior empirical comparisons between SSMs and Transformers, this work offers a mechanistic explanation rather than just benchmark observations, which is useful for anyone building long-context inference systems）; 「用能量视角统一机器学习：一篇 arXiv 教程把 GAN、diffusion、BERT 都看作 EBM」（做 ML 研究或想搞懂各模型之间关系的人值得花时间。它把 GAN、diffusion、BERT、对比学习全塞进同一个能量框架，差别只在配分函数和采样怎么处理。看懂这篇，很多论文里的公式其实是在说同一件事，而且能量组合还能直接推理出新算法思路）.
How-to (3 picks): 「SGLang Megakernels Lift Muse Spark Open Decoding to 1,109 Tokens/s」（Kernel-launch overhead, not compute, is often the real ceiling in multi-GPU inference, and this write-up shows concrete numbers and code proving it. Engineers building or tuning serving stacks on Blackwell-class hardware get CUDA/PTX examples and measured timelines they can adapt. Teams benchmarking against vLLM or TensorRT-LLM deployments will find the launch-count reduction a useful reference point）; 「AWS shows push-based pipeline to automate Agent Registry updates」（The non-obvious point is that registry freshness can be solved with event routing rather than polling jobs or stored credentials, using CloudTrail as the change source of truth. Platform and MLOps engineers building on Amazon Bedrock AgentCore should care, since the same EventBridge-to-Lambda wiring generalises to other agent metadata stores. Teams in multi-account AWS organisations get a reference pattern that works without cross-account secret sharing）.
