Skip to content

Translation in progress Some entries still include text in the original language.

DeepSeek · News

Oct 13 · Tuesday

DeepSeek · News

Latest publications and developments
Oct 7Wed
  1. CSA Labs Research, NIST CAISICSA Labs Research · 2 sources · 2 reports · 51

    NIST renames CAISI to CAISSI, focusing on superintelligence evaluation and standards

    The official website name of NIST's former CAISI center has been changed to the Center for Superintelligence Innovation and Standards Advancement (CAISSI), with functions including serving as the primary window for industry engagement with the U.S. government, promoting testing and collaborative research on commercial SI systems, working with NIST agencies to develop guidelines and best practices for measuring and improving SI system safety, assisting industry in developing voluntary standards, and establishing voluntary agreements with private-sector SI developers and evaluators, leading unclassified evaluations of SI capabilities that could endanger national security, with a focus on verifiable areas such as cybersecurity, biosecurity, and chemical weapons. CSA Labs analyzed Executive Order EO 14434, signed on September 29, 2026, which requires executive branch agencies to replace Artificial Intelligence and AI with Super Intelligence and SI in non-statutory communications, and imposes no safety or compliance requirements on the private sector; Section 3(a) defines the new terms as artificial intelligence under 15 U.S.C. § 9401(3), and therefore does not change existing regulations, contracts, or grants, and Section 2(b) explicitly does not require modification of previously issued regulations or presidential documents.

Oct 3Sat
  1. 量子位量子位 · 28

    DeepSeek elastic computing team expands recruitment, DSec sandbox supports Agent training

    DeepSeek's elastic computing team is heavily recruiting senior engineers, and its DSec sandbox infrastructure has supported all training, evaluation, and data preprocessing workflows for DeepSeek-V4. A single DSec extended shard has about 160 servers, 30,000 CPU cores, and 250TB of memory, serving about 3 million sandboxes daily, with over 380,000 online at peak. DSec optimizes resources through on-demand image loading, memory sharing, and CPU scheduling, and has observed Agents attempting to read residual answers and forge RPC requests in production environments, which have been addressed with measures such as AppArm.

Sep 28Mon
  1. Tech Policy Press · 22

    中非关系如何塑造 AI 地缘政治

    中国正通过关键矿产、数字基础设施投资和中国开放权重模型的低成本落地三条路径,将经济影响力转化为其在非洲的 AI 布局优势。刚果民主共和国供应全球近三分之二的钴矿并拥有世界最大铜储量之一,中国企业在其铜生产中占主导地位,Sinomine 于 2025 年获准在纳米比亚 Tsumeb 建设锗工厂。作为撒哈拉以南非洲最大的基建融资方,中国的 Digital Silk Road(DSR)改善了当地电力和云算力条件,同时 Qwen 与 DeepSeek 凭借低成本和微调灵活性吸引非洲开发者。

Sep 24Thu
  1. POLITICO Europe Technology · 35

    全球AI安全治理努力可能不得不在美国缺席下推进

    联合国大会期间,多国领导人及OpenAI CEO Sam Altman、Anthropic CEO Dario Amodei呼吁建立AI全球治理框架,Anthropic主张制定模型测试的"共同全球标准"。由芬兰和挪威牵头、22国通过的AI安全宣言未获美国和中国签署,特朗普在联大演讲中明确拒绝任何AI"全球主义控制方案"。前美国官员指出,美国不参与将限制相关努力的有效性,而中美领导人本周会晤仅将顺带讨论AI。

Sep 1Tue
  1. Tech Policy PressTech Policy Press · 43

    Russia's first large model law takes effect on September 1

    Russia's first comprehensive law targeting large foundation models took effect on September 1; the law was signed by Putin on July 26, and its core is to establish a support and control framework for eligible large model developers, rather than to manage AI risks. The law establishes two statuses, sovereign and domestic, requiring developers to be Russian legal entities, with generation and storage of user queries to be completed in data centers within Russia, and subject to assessments consistent with Russian law and traditional values, though the assessment procedures and award rules still await government resolutions. Those granted status can receive state support, participate in policy-making, and enter protected markets, and the government can also designate application scenarios where only these two types of models may be used.

Jul 31Fri
  1. Tech Policy Press · 29

    Mozilla CTO: Is open-source AI really a dangerous path?

    Mozilla CTO Raffi Krikorian believes that open-source AI is not a dangerous path: capability revocation is not a property of the model itself, but a term of the API contract, applying only to the lab's own customers. Once weights are released, they cannot be taken back. He cites Mozilla's *State of Open Source AI* report, stating that open-source models handle about one-third of global AI workloads yet earn only about 4% of revenue, and 79% of developers already build with open-source models.

Jul 20Mon
  1. Import AI · 43

    UK AISI 评测:开源权重模型与闭源前沿的网络能力差距正在缩小

    英国政府 AI Security Institute(AISI)首次公开分析开源权重模型在网络安全能力上落后闭源前沿模型的程度,结论是今年的差距已经缩小。在 70 项细分网络能力评测上,GLM-5.2 最接近早 4.3 个月发布的 Claude Opus 4.6,DeepSeek-V4-Pro 介于 Claude Opus 4.5 与 GPT-5 之间;而 2025 年大部分时间测得的差距为 6 到 10 个月。在名为 The Last Ones 的长周期网络靶场上差距更大,GLM-5.2 仅达到 Opus 4.5 的水平,DeepSeek V4-Pro 低于 Sonnet 4.5。

Jul 13Mon
  1. The EU AI Act Newsletter · 35

    EU AI Act Newsletter #106:呼吁强制执行通用人工智能模型规则

    欧盟委员会正推动从 2026 年 8 月 2 日起执行《EU AI Act》中针对具有系统性风险的通用人工智能模型的规则,并收到一批研究者与公民社会组织的公开信施压。信中援引 Anthropic 的 Mythos 模型存在广泛报道的网络攻击能力和生物学及失控风险逼近临界阈值的证据,要求赋予 AI Office 的评估员网络开展外部评估的权力并提供足够资源和政治支持。同期理事会通过 Omnibus VII 简化方案,将高风险系统适用日期推迟至 2027 年 12 月 2 日(嵌入式产品为 2028 年 8 月 2 日)。

Jul 6Mon
  1. 中国网信网 · 54

    中央网信办通报清朗AI技术滥用治理专项行动第一阶段处置结果

    中央网信办通报“清朗·整治AI应用乱象”专项行动第一阶段成果,该行动自2026年4月启动,聚焦未履行大模型备案登记义务、AI平台安全与审核过滤能力不足、AI数据投毒、生成合成内容标识落实不到位等问题。第一阶段累计处置违规网站、应用程序、智能体等AI产品1.4万余款,清理违法违规信息600余万条,处置账号2.6万余个,下架违规AI商品1300余个、违规开源数据集9个。北京、上海、浙江、江苏、广东等地网信部门采取联动巡查、细化平台要求、多维度整改、专项举报窗口、多部门协调监管等措施。华为、阿里巴巴、智谱、稀宇、DeepSeek等平台分别加强备案标识审核、数字指纹比对、多模态审核、恶意行为阻断和数据异常检测。下一步将开展第二阶段工作,重点整治利用AI制作虚假信息、散播不良信息、假冒仿冒、侵害未成年人权益和网络水军等问题。

May 12Tue
  1. IAPS · 41

    IAPS 政策备忘录:美国应对 AI 蒸馏攻击的行政与国会行动仍需补强

    IAPS 发布政策备忘录,认为白宫与国会在应对 AI 蒸馏攻击上已有动作,但力度与威胁规模不匹配。备忘录指出,OSTP 于 4 月 23 日发布的《美国 AI 模型对抗性蒸馏》国家安全与科技备忘录未涉及蒸馏攻击与先进半导体出口管制的关联、攻击者通过云端远程访问美国芯片的风险,也未说明具体问责机制,建议商务部在芯片出口与远程算力访问上纳入蒸馏风险,白宫公开其拟动用的具体授权。针对 4 月 15 日由众议员 Huizenga 提出、4 月 22 日以 43-0 通过众议院外交事务委员会审议的《2026 年威慑美国 AI 模型窃取法案》,备忘录提出四项修正:处理行业合作的反垄断与数据隐私障碍、允许跨政府共享情报、限制联邦机构访问受关注实体的模型、明确模型提取攻击构成商业秘密窃取,并将评估频率从每年一次共 3 年改为至少每年一次共 5 年。

Mar 23Mon
  1. IAPS · 43

    IAPS 评估 H200 对华出口的影响:DeepSeek 算力或接近美国数据中心

    IAPS 发布 Issue Brief,评估美国商务部 BIS 1 月半导体出口规则下 Nvidia H200 对华销售的影响。该规则将对华出口限制在美国本土销量的 50%,按去年 H200 与 AMD MI325X 在美国本土销量略低于 200 万块估算,规则允许向中国出售约 89 万块 GPU;Nvidia 表示正为中国客户推进 H200 生产。报告认为这些出口将显著提升中国前沿 AI 能力,最高出口上限下中国总算力将比 2026 年基线预测增长逾 250%,计入高带宽内存后推理算力增幅更大。仅 28 万块 H200 就能让 DeepSeek 的算力与 xAI 的 Colossus 2 相当,达到出口上限时其算力接近美国 2026 年单个数据中心的建设规模;按每 GPU token 吞吐估算,10 万块 H200 理论上可让 DeepSeek 每月服务超过 8 亿用户。

Mar 20Fri
  1. IAPS · 42

    IAPS 提议由市场化出口审计机构协助美国芯片出口管制执法

    IAPS 在一份工作论文中提出,美国出口管制执法可由独立营利性出口审计机构协助,这些机构由美国工业与安全局(BIS)认证和监管,进口方须为特定类别出口聘请经认证的审计公司,审计方通过现场检查等方式发现潜在的转运行为。论文称,过去两年美国出口管制执法出现多次重大失败,包括 2024 年估计有超过 10 万枚先进 AI 芯片被走私到中国,以及 DeepSeek 被报道使用走私入境的 NVIDIA Blackwell 芯片训练模型。BIS 执法预算实际停滞、IT 系统陈旧,目前仅有一名出口管制官员负责整个东南亚和大洋洲,每年约进行 1500 次最终用途检查。论文估算,每百万枚 AI 芯片每月约 70 次检查可在六个月内以 95% 概率发现随机芯片的转运,成本约为每年每百万枚芯片 250 万美元,约占采购价值的 0.005%。

Dec 18, 2025Thu
  1. CAIS · 43

    AI Safety Newsletter #67:特朗普签署阻止州级 AI 立法的行政令

    美国特朗普总统签署行政令,试图阻止各州自行监管 AI,通过联邦资金与执法手段评估、挑战并限制州级法律,并以一套尚未明确的联邦框架取而代之。该行政令无法单方面推翻州法,而是指示联邦机构依据现行法律解释发起挑战或扣留相关联邦资金:司法部长将组建工作组挑战其认为繁重的州级 AI 法律,商务部长将认定违规州不具备联邦宽带资金资格,FCC 将研究是否要求 AI 开发者向监管者提供标准化模型信息,FTC 需就现行反不公平或欺骗性商业行为法如何适用于 AI 模型发布指引。白宫 AI 与科学顾问将起草一部国家 AI 法,以覆盖与联邦政策冲突的州级规则,同时保留儿童安全、AI 基础设施、州政府自用 AI 等领域的州法。

You have reached the end