https://ai.h5x.app/AI 技术解读与实践zh-CNhttps://ai.h5x.app/articles/nemotron-voice/https://ai.h5x.app/articles/nemotron-voice/NVIDIA 的万圣节互动装置把识别、对话、语音合成与面部动画连成实时系统。真正值得借鉴的是它对状态和资源预算的处理。Sat, 31 Oct 2026 16:00:00 GMThttps://ai.h5x.app/articles/granite-modular/https://ai.h5x.app/articles/granite-modular/Granite Switch 4.1.1 把语音与决策能力做成可更新的适配器。模块化交付给模型系统带来了版本管理的新问题。Sat, 31 Oct 2026 16:00:00 GMThttps://ai.h5x.app/articles/protostar-data/https://ai.h5x.app/articles/protostar-data/Project ProtoStar 聚焦数据、模型与代码的可发现和可复用。对于 RAG 与科研 Agent，元数据和来源链同样是产品能力。Sat, 31 Oct 2026 16:00:00 GMThttps://ai.h5x.app/articles/mistral-large4/https://ai.h5x.app/articles/mistral-large4/MoE 选择专家，稀疏注意力选择上下文位置。这两种机制解决不同问题，模型选型需要回到自己的负载。Sat, 31 Oct 2026 16:00:00 GMThttps://ai.h5x.app/articles/apple-specialized/https://ai.h5x.app/articles/apple-specialized/AnySpeechRecognizer 与 AnyDecisionModel 提供可选择模型的 Swift 接口，说明端侧 AI 正在从单一聊天模型走向任务组合。Sat, 31 Oct 2026 16:00:00 GMThttps://ai.h5x.app/articles/lfm3-edge/https://ai.h5x.app/articles/lfm3-edge/Liquid AI 将文本、视觉、音频与决策模型放在共享骨干上。专业化能力需要与设备资源、数据和评测一起设计。Sat, 31 Oct 2026 16:00:00 GMThttps://ai.h5x.app/articles/novelty-judges/https://ai.h5x.app/articles/novelty-judges/Ai2 的 novelty benchmark 说明，提示、比较方式和检索配置都会改变模型对创新性的判断。自动评测本身也需要评测。Sat, 31 Oct 2026 16:00:00 GMThttps://ai.h5x.app/articles/inference-cost/https://ai.h5x.app/articles/inference-cost/Cloudflare 的代码评审系统案例表明，缓存命中率高并不意味着总成本低。成本优化需要观察工作量、上下文与质量。Sat, 31 Oct 2026 16:00:00 GMThttps://ai.h5x.app/articles/typed-extraction/https://ai.h5x.app/articles/typed-extraction/Cloudflare 公布的提取模型把数值、日期和置信区间作为直接输出。应用仍需要定义校验与不确定性处理。Sat, 31 Oct 2026 16:00:00 GMThttps://ai.h5x.app/articles/context-engineering/https://ai.h5x.app/articles/context-engineering/这是一篇基础阅读。Anthropic 对上下文筛选、压缩与长期任务状态的讨论，仍可作为构建可靠 Agent 的工程检查表。Sat, 31 Oct 2026 16:00:00 GMT