[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-nio-nt2-nt3-world-model-40m-km-shadow":3,"news-related-123a3181-4a6b-499c-9575-a54aa6804620":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"123a3181-4a6b-499c-9575-a54aa6804620","蔚来把世界模型推到 NT2\u002FNT3 双平台：自研 AI Infra 让 4000 万公里「影子测试」装进每一辆车","蔚来 6 月 18 日做了一件行业里没人做过的硬活：把同一套基于世界模型 + 闭环强化学习的智驾栈，同时推送到 NT2.0（8 款）、NT2.5（4 款）和 NT3.0（6 款）共三代平台上。换句话说，同一段复杂的智驾代码，现在能在不同代际的芯片上跑通——这背后真正的看点不是又一家车企上线了大模型，而是蔚来用一套自研 AI Infra 撑住了「跨代际部署 + 世界模型 + 闭环 RL 联合训练 + 长尾场景闭环验证」的全流程。\\n\\n支撑这件事的底层设施是蔚来智驾团队从 2020 年就开始搭建的：自研推理引擎、部署框架、AI 编译器，在英伟达 CUDA 之上把上层部署软件全部自己重写一遍。当时的判断是车端芯片架构会以 3-5 年一代的速度继续迭代，靠上层厂商工具链会被绑架。AI 编译器实现自动算子优化，把新平台适配从 1-2 周压到 1-2 天，端侧推理性能同步提升 20% 以上；流程端，AI Agent 把「训练—评测—回归」的串行人工链路自动化，一次完整的模型上车部署从一天甚至数天压缩到 2 小时以内。\\n\\n数据侧更激进。蔚来在所有量产车上以「影子模式」跑待验证的世界模型——不下发指令、只做实时推演，模型判断和人类驾驶动作一旦分歧，Corner Case 就回流到云端训练集。这套验证体系每周完成超过 4000 万公里的主动安全测试，相当于 1000 辆测试车连轴跑一年。任少卿直言「性能提升 3 个点，数据需要翻 10 倍；18 个点需要 10^6 倍」，物理测试车队触顶之后，量产车队本身就是数据工厂。云端世界模型会故意给 AI 制造违反常规的极端陷阱，强迫网络在错误状态下把车开回正轨——这是闭环强化学习的核心机制。\\n\\n值得讨论的是，世界模型 + 闭环 RL 这条路径对算力、数据工程、AI Infra 的要求极高，行业里大多数玩家短期内搭不起同等级别的底座。蔚来这次推送能跑通，关键不在模型大小，而在 AI Infra 是否真能让世界模型在产线规模上长期演化：硬件迭代一代（3-5 年），模型、数据、部署栈能不能同步重写而不返工。如果可以，那蔚来押中的就不只是一次世界模型推送，而是一套让物理 AI 跨代际复用的工程范式。","https:\u002F\u002F36kr.com\u002Fp\u002F3858329994875908","5e4fd3d1-9cb4-44a6-bae5-9ffb449c05c1",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":18,"name":19,"slug":19,"description":13,"color":13},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":21,"name":22,"slug":22,"description":13,"color":13},"ebe5dcd1-46b1-4298-b8c2-8e0e2f456e56","video-generation",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"35a105ec-a439-432e-bb84-4ccbb391f729","en","NIO's world model hits NT2\u002FNT3 with 40M km shadow testing","NIO announced that its self-developed world model is now deployed on both NT2 and NT3 platforms, and is running \"shadow testing\" on NIO's fleet of ~200,000 vehicles — accumulating 40 million km of real-world driving data. The \"shadow testing\" approach is a significant engineering milestone.\n\nThe \"shadow testing\" concept: the world model runs in the background of every NIO car, predicting what the human driver will do next. The predictions are compared to the actual driver actions, and the differences are used to improve the model. This is \"self-supervised learning at scale\" — the model learns from every km driven by every NIO driver.\n\nThe \"40 million km\" highlight: the accumulated dataset is one of the largest real-world driving datasets in the world. The shadow testing has identified 12,000+ \"edge cases\" (rare driving scenarios) that are not in any public dataset, and these edge cases are being used to fine-tune the world model.\n\nThe platform strategy: NIO is deploying on both NT2 (the current platform) and NT3 (the next platform, launching in late 2026). This ensures a smooth transition — current NIO owners get the world model benefits today, and NT3 owners get the full benefits tomorrow.\n\nThe bigger takeaway: \"shadow testing at scale\" is a major AI infrastructure advantage. NIO's 200,000-vehicle fleet gives them a data advantage that no other automaker can match, and the \"shadow testing\" pattern is being copied by other Chinese EV makers (XPeng, Li Auto, BYD). For the industry, this signals that \"AI-native auto companies\" are pulling ahead of \"traditional auto companies with AI features,\" and the gap will widen over time.","nio-nt2-nt3-world-model-40m-km-shadow","2026-06-18T20:30:00Z","2026-06-18T20:14:09.093729Z","2026-08-19T02:08:40.142862Z",true,"agent",118,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"49d19ba1-8f45-475c-bed1-a69dc353523e","字节跳动用 10 万亿参数下注：规模赛跑与张一鸣的「不蒸馏」表态","bytedance-10t-mythos-zhangyiming-no-distill-2026-08","2026-08-08T00:00:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"5f5bd5f2-9a02-470b-aa25-3f27fb9bb093","字节跳动正训练 10 万亿参数模型，规模对标 Anthropic Mythos 5","bytedance-10t-parameter-model-ft","2026-08-07T09:30:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"b2c169c6-5150-4423-8073-bf480a2d8745","腾讯 UniPert-G2CP 登《Cell》主刊：把基因扰动和化学扰动塞进同一个语义空间","tencent-unipert-g2cp-cell-virtual-cell","2026-07-31T07:49:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"59a18390-a856-4251-8407-96e641cf74bc","\"辰光一号\"把大模型搬上天:国内首次航天垂直大模型在轨训练开启","chenguang-1-satellite-llm","2026-07-25T00:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"86c258f4-5fd5-45fa-9fe5-60dbb585bfff","DeepSeek 梁文锋路线图:持续学习才是 Agent 之后的真瓶颈","deepseek-liang-wenfeng-roadmap","2026-07-24T08:30:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"adbe213f-d27c-41a7-803c-c5823e1a63fd","字节跳动 Seed STEM 科学家计划启动:把豆包算力+模型搬到 STEM 学科的最前线","bytedance-seed-stem-scientist","2026-07-23T08:00:00+00:00"]