破晓之爱

SC expresses concern over hate speeches made against minority communities_我的网站

一 |     2026年7月18日，在上海世博展览馆召开的“世界人工智能大会”现场，安谋科技举办了一场聚焦端侧AI的发布会，NPU产品线负责人、首席架构师舒浩博士带来了《Agentic AI时代下，“周易”NPU的全场景进化》的演讲，并披露了安谋科技自研的全新 “周易X3 Pro NPU”。         “我们正在一个行业变局的浪潮中，从过去的万物互联正要走向未来的万物智能。

二 |     New Delhi, Oct 21 (UNI) Expressing its serious concern over the way, some of the hate speeches were made against the minority communities during recent religious congregations, the Supreme Court questioned as to where we have reached, what have we reduced religion to?

"Where have we reached? What have we reduced religion to? What have we reduced God to? It is tragic," a two-judge bench of the Supreme Court, headed by Justice K M Joseph and also comprising Justice Hrishikesh Roy, observed.

"The statements are certainly very shocking for a country that is to be religion neutral," the Supreme Court said, and expressed shock at some of the statements and hate speech made against minority communities during certain recent religious congregations.

The Apex Court was hearing a plea filed by the petitioner, Shaheen Abdullah, seeking urgent direction, interventions and or orders to stop the hate speeches made against the Muslim community.

"A country which is a democracy and religion neutral, you are saying it there should be action in IPC. But it's against one community. It is sad," the Supreme Court observed.

Senior advocate Kapil Sibal, appearing for the petitioner, Abdullah, presented a recent "Hindu Sabha" as an example where BJP MP from West Delhi, Parvesh Verma, called for the "total boycott" of "these people" apparently referring to Muslims.

The Supreme Court said, in its order that "the respondent (authorities) will issue directions to their subordinates in this regard without looking at the religion of the accused, so that the secular nature of India is preserved.*

The Supreme Court also in its order directed that the respective governments and police authorities should take suo motu action in cases of hate speeches and they should not wait for lodging of a formal complaint.

"Such action is needed to be taken to preserve secular character of the country," the Supreme Court observed and expressed its shock over the way the hate speeches are made.

Abdullah, the petitioner, said the Muslim community is being "targeted and terrorised" by the participation of the members of the ruling political party in delivering the hate speeches.

The Supreme Court in its order directed the Delhi, Uttar Pradesh and Uttarakhand police to take Suo Motu cognizance action against any instance of hate speech in the State without waiting for a complaint to be filed.

The Supreme Court also made it clear that if there is a lack of action by the police and authorities, it will result in contempt of court against the concerned authority or person.

The Supreme Court also sought action taken report from Delhi police on alleged hate speeches made by Bharatiya Janata Party (BJP) Member of Parliament, Parvesh Verma and others in Delhi on October 9 against Muslim community.

Sibal argued that we have filed many complaints. This Court or administration never takes action! Always status reports. These people are participating in events on daily basis.

"The Administration does nothing. We keep coming to Court," Sibal pleaded for action.

While reading the alleged speeches of Varma, the Court read, "If need be, we will slit their throats", (Bench reads). It said, it is very sad and tragic.

Sibal said, Yes, him and team. He is an MP of the party (BJP).

"We need an SIT (Special Investigation Team) to make sure that it doesn't get repeated," Sibal pleaded to the Supreme Court.

UNI SM SHK2105。

三 | AI计算底座应该如何进化？我的答案是需要让AI智能体走进每一台边端侧设备，同时让每一台边端侧设备变成我们的智能伙伴。而承载这一愿景的，正是安谋科技自研的周易NPU。”舒浩说道。         四个关键洞察：AI计算底座必须全面进化          过去一年多来，AI行业从技术、产品到应用场景都发生了巨变。舒浩认为，针对AI智能体的推理场景，AI计算底座必须进行全面的进化。         舒浩的这一判断，主要源自四个维度的观察：          其一，模型能力大幅提升，推理需求从云端走向边端。舒浩指出，当前AI模型在逻辑推理、软件工程能力和Agent能力三个方向上取得了飞跃式进展。逻辑推理基准测试中，AI已经超越人类专家，能够独立处理复杂任务；软件工程层面，AI已接近生产力级别，可像中高级专家一样审阅文档、编写代码甚至修复Bug；而在Agent能力上，过去半年尤其迅猛，现在的AI已经可以熟练操作电脑、使用浏览器，生成和编辑各类文件，从简单的工具演变为可以与人协同的工作伙伴。         模型能力已经完全达到实用化要求，为AI应用爆发奠定了扎实的基础。而更关键的是，中小模型（如30B参数级别）的能力正在快速逼近旗舰模型，这意味着未来的推理计算需求不会只集中在云端，边端侧AI推理将迎来爆发前夜的窗口期。         其二，推理范式从单轮对话演变为AI智能体多轮协作。过去使用云端AI，是用户提问上传、云端计算返回答案的一次性交互。这种模式对网络依赖重，缺乏上下文，无法处理复杂场景。

四 | 而未来的AI智能体时代，将是一个集感知、记忆、推理、行动持续循环的闭环模式。大量请求需要实时本地处理，需要保持长上下文以应对复杂问题，同时兼顾数据隐私和主动理解场景的能力。         舒浩将其总结为“持续感知+上下文记忆+多模态理解+主动行动”的集合。他强调，这种范式转变“一定会催生NPU计算架构的改变”，NPU需要从一次性的推理加速器进化为AI智能体的计算底座。         其三，不同应用场景推理工作负载高度分化，统一硬件无法适配所有场景。舒浩展示了一组AI推理Token消耗数据：代码编程以52%的占比成为绝对主力，紧随其后的是角色扮演、工程技术、语言翻译、科学计算等。但比起Token消耗，更值得关注的是不同应用场景对系统资源需求的差异。比如，代码编程场景对Prompt和KV Cache需求大，对推理延迟有一定要求，但成本不敏感；角色扮演对推理延迟要求极高，追求实时交互体验；语言翻译则对系统成本最为敏感。         “整个云端推理负载已经分化，不同场景对性能指标、系统资源的要求都不相同。”舒浩断言，“理论上我们没法使用统一的硬件解决方案来适配各种不同的使用场景。”          其四，边端侧AI推理场景需求差异极大。舒浩将边端侧AI场景划分为四大类，各自的核心诉求截然不同：手机、PC、可穿戴等个人终端设备，核心诉求是电池续航、个人隐私和本地快速响应；摄像头、门锁、自动驾驶、机器人等视觉与空间智能场景，核心诉求是多模态传感、低延迟处理和功能安全；工业网关、边缘工控机等工业与能源边缘场景，核心诉求是任务确定性、设备可靠性和长生命周期；家庭AI Box、边缘服务器等边缘节点与AI本地盒子，核心诉求则是多路并发、24小时在线和云边协同。         边端侧AI推理场景对于系统需求的巨大差异，也再度反映了无法用统一硬件方案高效适配所有场景。         周易X3-Pro计算架构：统一、分层、场景化          面对这一核心矛盾，舒浩给出了安谋科技的“周易”NPU的演进路径：一个全场景部署、极致系统效率加高部署效率的计算底座。         回顾“周易”NPU的发展路线图，从2018年的Z1 NPU开始起步，到现在完整经历了CNN时代、CNN+Transformer时代、大模型时代，如今正迈向Agentic AI时代。         去年发布的“周易”X3 NPU被定位为“专为大模型而生的端侧AI效率新标杆”。而随着AI智能体的兴起，安谋科技团队正在研发新一代“周易”X3-Pro架构，定位为“面向Agentic AI全场景高效的计算架构”。         舒浩逐层剖析了X3-Pro的核心思路：通过统一架构、灵活的实例化配置来覆盖各种不同场景。未来产品线将分为三个梯度：X3-Pro Lite系列覆盖低功耗端侧推理场景，X3 Premium系列覆盖中等算力推理场景，X3-Pro Subsys系列覆盖大算力复杂推理场景。

五 |          支撑这一产品矩阵的，是“统一的分层场景化NPU计算架构”。舒浩解释道，“统一、分层、场景化”三个关键词贯穿设计始终。         具体来说，在应用层，面对边端侧各种各样的推理场景，通过统一的软件层进行适配，保证高部署效率；在中间的核心软硬件适配层，上层定义统一的ISA和统一硬件界面，下层则引入差异化的硬件抽象，以对接不同的硬件层；硬件层在统一计算架构下，可设计可穿戴专用NPU（强调专用架构和高计算效率）、端侧通用NPU（强调通用架构和高灵活性）、存算一体NPU乃至非冯·诺依曼创新架构（强调高灵活性和极致系统效率）。         舒浩将这套设计哲学凝练为十六个字：“一次适配，多芯复用，面向场景，灵活配置”。

六 | 软件层面，统一的软件架构让客户可以在多个行业和产品形态中复用历史的模型适配、调优和部署经验；硬件层面，面向场景的统一架构配合多种灵活适配方式，高效满足不同场景对性能、功耗、灵活性、可编程性的差异化要求。         在关键特性升级方面，据介绍，X3-Pro相较去年发布的X3在多个维度实现了跃升：通用可编程性上支持更多数据精度和更高效的统一编程模型；专用计算效率上引入更多混合精度加速计算；灵活可配置性上实现更强大的单核算力、更灵活的Cluster配置以及可配置的内存层次架构；面向未来的可扩展性上，创新性地支持了非冯·诺依曼架构。         “让AI智能体走进每台边端侧设备。”舒浩在演讲最后，提出了希望与生态伙伴共同定义Agentic AI时代边端侧智能产品范式的愿景。         小结：          从CPU产品线的“星辰”平台到NPU产品线的“周易”X3-Pro，安谋科技正在边端侧AI的广阔战场上，布下一盘覆盖全场景、兼顾效率与灵活性的棋局。当AI智能体从云端走向真实世界的每一个角落，底层的计算底座进化，才刚刚开始。         编辑：芯智讯-浪客剑。

Current article:http://n9gey20.binhannuxiangtulediatacang.cfd/8ezk/20260826/7561.html

Published on:15:12:06