怪谈百物语
AI时代需要三座“工厂”，怎么建、谁买单？_我的网站
A |
The third China-US Track 1.5 Dialogue was held in Beijing on Tuesday, bringing together government officials, scholars and business representatives from both countries for discussions on constructive strategic stability and practical cooperation.
Co-hosted by the International Department of the Central Committee of the Communist Party of China (IDCPC) and the Asia Society, the dialogue was attended by about 40 representatives. Participants included officials from relevant Chinese government departments and local governments, experts from think tanks and universities, as well as figures from US strategic and business circles, including John L. Thornton, co-chair of the Asia Society Board of Trustees.
Under the theme "Constructive Strategic Stability - Practical China-US Cooperation in the New Era," the participants held candid, in-depth and constructive exchanges, according to the IDCPC's release on its WeChat account on Tuesday.
Addressing the dialogue, Liu Haixing, head of the International Department of the CPC Central Committee, said the two sides have recently followed the direction set by the two heads of state, taken practical steps to strengthen communication and dialogue, advanced a positive agenda and achieved notable results.
Liu called on the two countries to continuously enlarge common interests and make cooperation the anchor of positive stability. He urged the two countries to uphold mutual benefit and reject zero-sum games so that competition remains within bounds, to seek common ground while managing differences, and observe the bottom line of no conflict and no confrontation to achieve lasting peace and stability.
Participants discussed ways to enhance political mutual trust, expand economic and trade cooperation, and explore cooperation in artificial intelligence, science and technology, CCTV News reported.
Established in 2023, the China-US Track 1.5 Dialogue operates between formal government-to-government diplomacy, known as Track 1, and purely nongovernmental exchanges, known as Track 2. The second dialogue was held on June 13, 2024, during which participants held candid, in-depth and constructive exchanges around the theme of "sharing perspectives and finding solutions on key and urgent issues."
As a channel involving both government and nongovernmental participants, the Track 1.5 mechanism can help China and the US communicate more effectively, identify common ground and better understand the consensus reached by the two heads of state, Zhou Mi, a senior research fellow at the Chinese Academy of International Trade and Economic Cooperation, told the Global Times on Tuesday.
It can also help translate the consensus into greater mutual understanding and concrete action, Zhou said.
Discussions on AI at this year's dialogue could help the two sides identify shared interests despite existing frictions and technological challenges, creating opportunities for practical cooperation and laying the groundwork for follow-up efforts, he noted.
Wang Yiwei, a professor at the Renmin University of China's School of International Studies, told the Global Times on Tuesday that the Track 1.5 dialogue serves as a bridge between Track 1 and Track 2, allowing participants to discuss sensitive issues and explore policy space while conveying frontline assessments and feasible proposals to decision-making departments.
Liu outlined practical pathways for China-US cooperation in the new era. First, the two sides should uphold mutual respect, peaceful coexistence and win-win cooperation, steer China-US relations to overcome the "Thucydides Trap," and create a new paradigm of major-country relations.
Second, the two countries should jointly shoulder their responsibilities as major powers, make constructive interaction an important driver of international cooperation, address global challenges together and inject greater stability into the world, Liu noted.
Third, they should expand exchanges and cooperation in trade, public health, agriculture, tourism, people-to-people ties, law enforcement and cooperation between localities, thereby strengthening the resilience of bilateral relations and working toward a better future, Liu said.
Overcoming the "Thucydides Trap" fundamentally rejects the notion that conflict between major powers is inevitable, Wang Yiwei said. It reflects an effort by China and the US to pursue a path different from the historical confrontation between an established power and a rising one.
Such a model does not seek to avoid competition altogether, but to ensure that competition has boundaries, differences remain manageable and cooperation retains sufficient space, Wang Yiwei said.
To put this approach into practice, the two countries should respect each other, avoid crossing red lines concerning core interests and reduce the risk of strategic miscalculation. They should expand areas where their interests converge and establish regular cooperation mechanisms on global issues such as climate change, public health, food security and AI safety, turning shared responsibilities into joint action, he noted.
As the world's two largest economies, China and the US should recognize that many global challenges cannot be effectively addressed without coordination between them, particularly at a time of intensifying competition. Constructive engagement is therefore not a temporary expedient but a long-term responsibility that major countries should undertake for the world, Wang Yiwei said.
The dialogue came a day after China's top political advisor Wang Huning met with US delegates attending the event in Beijing on Monday.
During their meetings, the US side said that the US-China relationship is at a special juncture, adding that the sound rapport between the two heads of state constitutes the most valuable and important part of bilateral ties. The US side agrees that both sides should act on the consensus of building a constructive US-China relationship of strategic stability, advance corporate cooperation, expand people-to-people and cultural exchanges, guard against risks such as those posed by artificial intelligence, and deliver benefits to both countries and the rest of the world, according to Xinhua.
。 我们正在迈入Agentic AI（智能体）时代。在AI时代，Token（词元）正成为最大的消耗品，而Agent正在成为消耗Token的主体。 面对这样的变革，7月18日，在摩尔线程“词元时代 万物智能”主题论坛上，摩尔线程创始人、董事长兼CEO张建中给AI产业的“卖铲人”下了新定义：AI产业需要AI工厂，而这个工厂分为三类，分别是模型训练工厂、词元生产工厂、智能体生产工厂。 张建中阐述三大“AI工厂”概念 模型训练工厂：打造万卡级长稳训推的硬底座 训练一个前沿大模型，成本动辄数亿甚至上十亿元。模型能不能达到前沿水平、能不能快速从语言扩展到多模态乃至物理AI，与训练工厂的基础设施能力息息相关。 一个合格的模型训练工厂需要具备以下几个条件：全功能——下一个要训练的模型可能是语言、视觉、世界模型或具身大脑，模型训练工厂必须什么都能训；高扩展——随着集群训练规模扩大，训练工厂需要保持极高的线性扩展效率；高精度——训练损失曲线和评测指标必须对齐国际主流；高稳定——万卡级长连续训练，中断一次代价巨大；软件栈完备——从算子到框架再到强化学习，缺一环就跑不通全链路。 摩尔线程以夸娥（KUAE）智算集群为核心，逐一回应了这些要求。在万卡级规模下，集群保持95%的线性扩展效率，有效训练时长占比超过90%，训练精度与国际主流计算卡持平。软件栈方面，摩尔线程从底层MUSA SDK出发，兼容Torch、Triton、TileLang等算子生态，支持Megatron-LM、DeepSpeed、FSDP等主流训练框架，并提供MT-HTA仿真系统（SimuMax/SimuInfer）在训练前预估时长与效果。 在夸娥万卡级集群和训练套件之外，摩尔线程还针对MoE架构，集成了DeepEP通信库、算子融合等先进技术，原生支持Per-Block FP8精度，使算力得以显著提升。 WAIC 2026摩尔线程展台 “我经常听到国内的模型公司担心，用国产的训练平台会不会耽误时间，会不会影响追赶世界先进水平。”张建中在演讲中传达出用户的疑问。 但摩尔线程已经通过实践对这一质疑给出了直接反击：MoE-236B在万卡级集群上从零起步完成了完整训练，基于25T+高质量语料，精度与损失曲线同国际主流高度一致；北京大学5D世界模型EvoPhys-World在夸娥集群上完成全栈原生训练，连续37天登顶斯坦福WorldScore“世界生成”维度；北京智源研究院RoboBrain 2.5在MTT S5000千卡集群上完成全流程训练，国产算力首度跑通具身大脑训练。 在WAIC 2026的展台上，记者还看到摩尔线程在WAIC现场首发的MTT C256超节点：首创一层Scale-up网络，实现标准单宽机柜128卡全互联、并柜扩展可达256卡极速互联，采用计算与交换一体化的高密设计，可大幅提高智算中心GPU部署密度，成为支撑长稳训推的硬件底座。 词元生产工厂：把每一份算力变成高质量Token 如果说训练工厂解决的是“训得出”，词元工厂解决的就是“供得上”——而且要供的是高质量Token。
B | 国家信息中心大数据发展部副主任魏颖在论坛上给出了一个判断：词元时代，产业正从卖电、卖算力升维到卖Token——按实际消耗付费，告别粗放的打包兜售，迈向按需付费的精细运营。 但用不同算力设施、不同模型生产的Token质量可能千差万别，Token的质量也直接决定了用户的消费意愿。中国工程院院士、清华大学计算机科学与技术系教授郑纬民在演讲中阐述，从技术指标来看，Token质量的差别主要体现在四个方面——首字延迟（TTFT）、吞吐量（TPS，每秒生成Token数）、并发请求数（同时服务多少请求）以及质量保障（SLO）。首字延迟越低、吞吐量越大、并发请求数越高、质量保障越好的Token对于用户而言越有价值。 郑纬民阐述何为高质量Token “如何规模化地生产高质量Token？”成为词元生产工厂的核心命题。 郑纬民对推理计算实现的过程进行了详细剖析：他表示，大模型推理涵盖两个关键过程——Prefill（预填充，简称P）把用户请求一次性处理成中间格式，是计算密集型；Decode（生成，简称D）逐字吐出结果，要求内存带宽高、延迟小，是带宽密集型。 这两大过程对硬件的诉求截然不同。过去很多计算芯片既要做P过程又做D过程，存在资源错配。而P过程与D过程的物理分离，意味着P过程与D过程可以分别部署在不同的算力芯片上，实现“各取所长”。 基于PD分离的思想，摩尔线程推出了异构推理方案：由摩尔线程的算力芯片S5000完成Prefill过程，同时Decode依然安排在国际主流卡上，以此实现异构组合、分池部署。
C | 但令人惊喜的是，这样的组合带来的不是“1+1=2”的结果，而是“1+1”远大于2的效果。张建中解释道：“在这样的情况下，约3台S5000+2台国际主流卡，性能相当于9台国际主流卡。” 这一方案也得到了行业伙伴的支持。趋境科技创始人艾智远在论坛上直言，S5000是他们经过全面评测后选择的当前阶段最优Prefill方案——原生FP8精度无需额外适配、稳定支撑超长上下文、生态迁移顺畅、供应链可靠。在Prefill为王的编程场景里，4台S5000的Prefill性能能够超过1台B300。趋境科技目前日均供应万亿级高质量Token，下半年这一数字将达数万亿。 模型厂商的判断也在印证这一转变。MiniMax副总裁薛子钊在论坛上表示，从整体训练和推理来看，今年是国产芯片非常重要的拐点，越来越多真实的线上场景开始承担起推理和训练任务，MiniMax很快也会有非常大规模的国产芯片推理集群真实上线承担任务。他特别提到，包括摩尔线程在内的多家国产芯片企业都在规划超节点，互联规模超过英伟达当前的72卡方案，“这一点国产芯片是做得比较领先的”——这让大参数、长上下文的推理场景能获得非常高的吞吐；而同构或异构的PD分离，则能把各类卡的性能发挥到最大化，整体降低系统推理成本。 智能体生产工厂：从数字世界到物理世界 第三座工厂是最少被提及、却可能最具想象空间的。
D | 张建中把它定义为"连接数字世界与物理世界"的基础设施——AI最终落地到每个消费者身边，需要两种智能体：数字智能体和物理智能体（具身智能）。
E | 数字智能体的核心要求是：能跨应用控制、能记忆、能调用工具完成复杂工作流、支持主流开源与闭源框架。摩尔线程自研了全域数字智能体“小麦”，掌握60余项技能，支持38款APP的跨应用控制，通过前端多模型、丰富知识库的灵活配置，做到更快、更准、更懂用户。 WAIC 2026上摩尔线程展示智能体生产工厂 物理智能体的要求则更为严苛。机器人需要在虚拟环境中完成训练与验证，才能部署到真机，实现Sim-to-Real（从虚拟到现实）。 摩尔线程为此打造了具身智能仿真平台MT Lambda，从底层算力、核心引擎到上层框架构建了完整解决方案。平台集成了三大物理仿真引擎——开源的MuJoCo和Newton，以及摩尔线程自研的AlphaCore物理引擎。在图形渲染方面，平台集成了3DGS、自研AI生成式渲染（AGR），以及光线追踪（Ray-Tracing）——MTT S5000是国内唯一具备Ray-Tracing能力的智算芯片，这意味着仿真环境中可以实现高度拟真的数字孪生渲染，让机器人在虚拟世界中充分训练后再部署到真机。WAIC展台上，一只机器狗的仿真训练动作与真机跳起动作几乎一模一样。平台还支持百卡到千卡级的大规模集群扩展，满足海量仿真训练需求。
F | 京东集团技术委员会主席、京东云总裁曹鹏则从产业应用端印证了三大工厂的价值。他表示，AI基础设施层面正发生一个很大的转变，包括国产GPU、国产推理引擎、国产模型在内的全国产化体系正越来越多地被行业认可和使用，“摩尔线程就是一个非常典型的案例。”在模型工厂中，京东结合MTT S5000进行模型训练、适配与调优，刚刚发布的一系列JoyAI大模型，相当一部分训练算力正是依托摩尔线程提供的国产化算力；在Token侧，京东自研的xLLM推理引擎与MTT S5000做了深度适配，支撑集团每天大量的Token应用。 曹鹏透露，京东正在全面布局无人配送、仓储机器人、家政服务机器人等物理世界场景，其"女娲"项目计划采集超过1000万小时的音视频数据用于训练具身大脑模型，而基于摩尔线程的计算与渲染优势，双方在数据生成和具身仿真领域开展了深度合作。现在，京东内部已部署400多个智能应用、4万多个智能体，每天Token消耗已近10万亿，从研发、采销到快递配送、客服，遍布各业务环节。 运维侧的智能体实践同样值得关注。无问芯穹技术副总裁吴保东分享了Infini-AIOps运维智能体在摩尔线程集群上的部署成果：调用摩尔线程的专家知识库与SDK，把故障处理评分从人工的50分提升到86分，平均处理时长从1小时缩短到20分钟以内，效率提升5倍；复杂问题的解决率从不到45%跃升至90%。无问芯穹与信通院联合发布的AISHPerf运维智能体评测基准，弥补了GPU集群运维智能体缺乏统一标准的不足。用吴保东的话说，智能体不仅要“造得出来”，还要“管得住、跑得稳”。 作者丨姬晓婷 编辑丨张心怡 美编丨马利亚 监制丨连晓东。
G |
Current article:http://paa0y.yuejinhuapennuochoudengying.cyou/klsnv/h0vv.html
Published on:14:45:06