理解不断变化的 AI。https://ainews.imrenagi.com/zh-Hanshttps://ainews.imrenagi.com/zh/articles/2026-09-03-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-09-03-daily-digest/一份带来源链接的简报，涵盖新的Google和Meta模型、网络安全修复、选举深度伪造、云电脑、数据控制、推荐来源、空间模型以及学校AI政策。Wed, 02 Sep 2026 22:13:32 GMTai-newsdaily-digestdaily-digestai-agentsdeveloper-toolsai-securityai-privacyai-policyhttps://ainews.imrenagi.com/zh/articles/2026-09-03-mml-008-logarithms-turn-products-into-sums/https://ainews.imrenagi.com/zh/articles/2026-09-03-mml-008-logarithms-turn-products-into-sums/学习把对数看作逆幂，推导乘积转求和的法则，并使用 NumPy 对数分数区分被 float64 舍入为零的概率乘积。Wed, 02 Sep 2026 21:10:00 GMTmath-for-aimathematical-foundationslogarithmslog-likelihoodnumerical-stabilityhttps://ainews.imrenagi.com/zh/articles/2026-09-02-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-09-02-daily-digest/关于Astra网络安全防护、Anthropic训练暂停、新前沿模型、AI基础设施、欧盟监管和生成式界面的来源链接简报。Tue, 01 Sep 2026 22:18:00 GMTai-newsdaily-digestdaily-digestai-safetyai-securityai-modelsai-infrastructureai-policyhttps://ainews.imrenagi.com/zh/articles/2026-09-02-mml-007-powers-roots-and-model-scaling/https://ainews.imrenagi.com/zh/articles/2026-09-02-mml-007-powers-roots-and-model-scaling/学习指数律与根式，然后使用对数刻度探索器和 NumPy 观察为什么模型维度或上下文翻倍会产生四倍数量。Tue, 01 Sep 2026 21:14:00 GMTmath-for-aimathematical-foundationsexponentsscaling-lawscomplexityhttps://ainews.imrenagi.com/zh/articles/2026-09-01-google-search-console-generative-ai-report-measurement-guide/https://ainews.imrenagi.com/zh/articles/2026-09-01-google-search-console-generative-ai-report-measurement-guide/Google 现在在全球报告生成式 AI 展现。使用这份逐字段工作表衡量可见性，但不要把它误认为流量。Tue, 01 Sep 2026 01:12:00 GMTai-newsai-applicationsgoogle-search-consoleai-searchsearch-analyticsmeasurementpublishershttps://ainews.imrenagi.com/zh/articles/2026-09-01-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-09-01-daily-digest/一份带来源链接的简报，关注 ChatGPT 的广告经济、Claude 账户安全、音乐训练诉讼、预测模型，以及塑造实际 AI 采用的控制机制。Mon, 31 Aug 2026 22:14:00 GMTai-newsdaily-digestdaily-digestai-productsai-securityai-policyai-researchai-infrastructurehttps://ainews.imrenagi.com/zh/articles/2026-09-01-mml-006-sums-products-and-loss-aggregation/https://ainews.imrenagi.com/zh/articles/2026-09-01-mml-006-sums-products-and-loss-aggregation/学习求和与乘积记号，将上下界展开为循环，并使用 NumPy 的求和与平均把逐样本误差变成均方误差目标。Mon, 31 Aug 2026 21:16:21 GMTmath-for-aimathematical-foundationssummationloss-functionsnumpyhttps://ainews.imrenagi.com/zh/articles/2026-08-31-ai-research-reproducibility-cost-code-data-environment/https://ainews.imrenagi.com/zh/articles/2026-08-31-ai-research-reproducibility-cost-code-data-environment/一项涵盖 918 篇论文的研究绘制了复现 AI 研究的成本。用其发现为代码、数据、算力和审阅者时间编制预算。Mon, 31 Aug 2026 13:10:24 GMTai-newsresearchreproducibilityai-researchresearch-artifactsmachine-learningevaluationopen-sciencehttps://ainews.imrenagi.com/zh/articles/2026-08-31-math-for-ai-roc-precision-recall-imbalanced-data/https://ainews.imrenagi.com/zh/articles/2026-08-31-math-for-ai-roc-precision-recall-imbalanced-data/通过一个阳性率为 1% 的检测器，逐步把阈值转换为 ROC 和精确率-召回率点，并为 30 个告警的审核预算选择运行点。Mon, 31 Aug 2026 01:20:30 GMTmath-for-aiprobability-and-distributionsroc-aucprecision-recallimbalanced-dataclassificationthresholdsmodel-evaluationhttps://ainews.imrenagi.com/zh/articles/2026-08-31-mml-005-indices-subscripts-and-tensor-addresses/https://ainews.imrenagi.com/zh/articles/2026-08-31-mml-005-indices-subscripts-and-tensor-addresses/学习读取张量下标，将批次-词元-特征地址转换为 NumPy，并发现 AI 代码中看似有效的轴错误。Sun, 30 Aug 2026 21:06:24 GMTmath-for-aimathematical-foundationsindicestensorsnotationhttps://ainews.imrenagi.com/zh/articles/2026-08-30-openai-cursor-model-access-cutoff-migration-plan/https://ainews.imrenagi.com/zh/articles/2026-08-30-openai-cursor-model-access-cutoff-migration-plan/OpenAI 提议于 11 月 12 日结束在 Cursor 中的模型访问。使用这份固定任务矩阵和 30/14/7 天演练来测试替代方案并保留回滚能力。Sun, 30 Aug 2026 09:11:34 GMTai-newsai-applicationsopenaicursorcoding-agentsmodel-migrationdeveloper-toolsservice-continuityhttps://ainews.imrenagi.com/zh/articles/2026-08-30-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-08-30-daily-digest/一份带来源链接的简报，关注代理式 IDE 安全、区域模型分发、获许可知识，以及在实践中信任 AI 系统所需的证据。Sat, 29 Aug 2026 22:05:19 GMTai-newsdaily-digestdaily-digestai-agentsai-securityai-productsopen-source-aiai-researchai-policyhttps://ainews.imrenagi.com/zh/articles/2026-08-30-mml-004-sets-membership-and-datasets/https://ainews.imrenagi.com/zh/articles/2026-08-30-mml-004-sets-membership-and-datasets/通过发现并修复 AI 训练、验证和测试划分中的重叠，学习集合、隶属关系、并集、交集和分割。Sat, 29 Aug 2026 21:06:55 GMTmath-for-aimathematical-foundationssetsdatasetsnotationhttps://ainews.imrenagi.com/zh/articles/2026-08-29-temperature-scaling-calibration-ece/https://ainews.imrenagi.com/zh/articles/2026-08-29-temperature-scaling-calibration-ece/通过 softmax，利用六个预测拟合温度，计算 ECE，并了解为什么更好的似然仍可能带来更差的分箱校准分数。Sat, 29 Aug 2026 15:17:04 GMTmath-for-aiprobability-and-distributionstemperature-scalingcalibrationsoftmaxexpected-calibration-errorprobabilityhttps://ainews.imrenagi.com/zh/articles/2026-08-29-mml-003-domain-codomain-and-valid-inputs/https://ainews.imrenagi.com/zh/articles/2026-08-29-mml-003-domain-codomain-and-valid-inputs/学习定义域和陪域如何规定有效的模型输入与承诺的输出类型，然后诊断形状、范围和未定义运算导致的失败。Sat, 29 Aug 2026 14:24:00 GMTmath-for-aimathematical-foundationsfunctionsdomainsmodel-inputshttps://ainews.imrenagi.com/zh/articles/2026-08-29-mml-002-functions-as-machines/https://ainews.imrenagi.com/zh/articles/2026-08-29-mml-002-functions-as-machines/从变量走向函数：计算输入—规则—输出的映射，组合简单计算，并理解为什么同一个模型可以处理许多示例。Sat, 29 Aug 2026 14:23:00 GMTmath-for-aimathematical-foundationsfunctionspredictionpythonhttps://ainews.imrenagi.com/zh/articles/2026-08-29-mml-001-variables-values-and-ai-models/https://ainews.imrenagi.com/zh/articles/2026-08-29-mml-001-variables-values-and-ai-models/通过区分变量名与其当前值，学习如何阅读模型方程，并在一个微型评分模型中追踪输入、参数和输出。Sat, 29 Aug 2026 14:22:00 GMTmath-for-aimathematical-foundationsvariablesmodel-parametersbeginnerhttps://ainews.imrenagi.com/zh/articles/2026-08-29-pipette-on-device-ai-benchmark-interpretation/https://ainews.imrenagi.com/zh/articles/2026-08-29-pipette-on-device-ai-benchmark-interpretation/Pipette 发布了许多配置下的延迟、吞吐量、内存和质量结果，但不同手机、运行时和量化方式并不能组成一个干净的排名。Sat, 29 Aug 2026 09:02:47 GMTai-newsai-applicationspipetteon-device-aibenchmarksquantizationllama-cppedge-aiinferencehttps://ainews.imrenagi.com/zh/articles/2026-08-29-kubernetes-inference-perf-model-server-benchmark/https://ainews.imrenagi.com/zh/articles/2026-08-29-kubernetes-inference-perf-model-server-benchmark/Kubernetes Inference Perf 让模型服务器比较更加一致，但其结果仍取决于提示词、负载模式、token 数量和集群。Sat, 29 Aug 2026 05:05:00 GMTai-newsai-applicationskubernetesinference-perfllm-servingbenchmarkingai-infrastructureperformance-testinghttps://ainews.imrenagi.com/zh/articles/2026-08-29-world-cup-arena-leakage-free-llm-benchmark/https://ainews.imrenagi.com/zh/articles/2026-08-29-world-cup-arena-leakage-free-llm-benchmark/WorldCup Arena 在比赛结果产生前测试 LLM 预测，但评分方式会改变排名，公开档案也不包含每一次原始模型调用。Sat, 29 Aug 2026 03:04:45 GMTai-newsresearchworldcup-arenallm-benchmarksforecastingmodel-evaluationsdata-leakagereproducibilityhttps://ainews.imrenagi.com/zh/articles/2026-08-29-anthropic-model-hardware-standard-safety-gate/https://ainews.imrenagi.com/zh/articles/2026-08-29-anthropic-model-hardware-standard-safety-gate/Anthropic 的模型硬件标准为代理提供了通用设备接口，但物理联锁、身份、时序、观测和恢复仍然是外部控制。Sat, 29 Aug 2026 01:09:00 GMTai-newspolicy-and-safetyanthropicmodel-hardware-standardai-agentsphysical-aisafety-testinghttps://ainews.imrenagi.com/zh/articles/2026-08-29-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-08-29-daily-digest/一份带来源链接的最新简报，关注自动化对齐、代理安全、开放模型、云访问、消费者部署和 AI 政策。Sat, 29 Aug 2026 00:06:59 GMTai-newsdaily-digestdaily-digestai-agentsai-securityopen-source-aiai-infrastructureai-policyai-developer-toolshttps://ainews.imrenagi.com/zh/articles/2026-08-28-chatgpt-bocconi-critical-thinking-study-rubric/https://ainews.imrenagi.com/zh/articles/2026-08-28-chatgpt-bocconi-critical-thinking-study-rubric/一项涉及 1,053 名学生的试验发现，ChatGPT 提高了经过评分的建议质量，但其原创性指标是计算得出的，研究也没有衡量之后的学习。Fri, 28 Aug 2026 13:05:00 GMTai-newsai-applicationschatgptai-educationcritical-thinkingcausal-reasoningevaluationrubricsoriginalityhttps://ainews.imrenagi.com/zh/articles/2026-08-28-gpt-5-6-kiro-cost-per-completed-task/https://ainews.imrenagi.com/zh/articles/2026-08-28-gpt-5-6-kiro-cost-per-completed-task/OpenAI 和 AWS 报告称，GPT-5.6 Terra 在 Kiro 中完成 Terminal-Bench 任务的成本大约降低了 82%，但仓库工作还会增加审查和修复成本。Fri, 28 Aug 2026 11:09:00 GMTai-newsai-applicationsgpt-5-6kirocoding-agentsdeveloper-toolsmodel-evaluationengineering-costhttps://ainews.imrenagi.com/zh/articles/2026-08-28-openai-brazil-commercial-operation/https://ainews.imrenagi.com/zh/articles/2026-08-28-openai-brazil-commercial-operation/OpenAI 现在已在圣保罗开展商业运营。应将本地团队和开发者项目与采用量主张、产品访问以及数据驻留区分开来。Fri, 28 Aug 2026 09:05:00 GMTai-newsai-applicationsopenaibrazildeveloper-communityapidata-residencyai-adoptionhttps://ainews.imrenagi.com/zh/articles/2026-08-28-anthropic-risk-report-auditability/https://ainews.imrenagi.com/zh/articles/2026-08-28-anthropic-risk-report-auditability/阅读 Anthropic 的《2026 年 8 月风险报告》时，应区分报告覆盖日期、公司评级、披露的失败、删节内容和外部审查。Fri, 28 Aug 2026 07:00:00 GMTai-newspolicy-and-safetyanthropicresponsible-scaling-policyai-safetyrisk-reportsai-governancemodel-evaluationhttps://ainews.imrenagi.com/zh/articles/2026-08-28-openai-admin-plugin-permission-boundaries/https://ainews.imrenagi.com/zh/articles/2026-08-28-openai-admin-plugin-permission-boundaries/OpenAI 的管理插件将工作区操作带入聊天，同时角色、已启用能力、审批、权威状态和回滚仍然是彼此独立的边界。Fri, 28 Aug 2026 05:07:22 GMTai-newsai-applicationsopenaichatgpt-workcodexpluginsworkspace-administrationaccess-controlhttps://ainews.imrenagi.com/zh/articles/2026-08-28-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-08-28-daily-digest/今天的证据涵盖据报道的模型中心收购、AI 算力交易、物理代理试验、网络安全事件、内存投资和隐私控制。Thu, 27 Aug 2026 22:08:38 GMTai-newsdaily-digestdaily-digestai-infrastructureai-agentsphysical-aiai-securityopen-source-aiai-economicshttps://ainews.imrenagi.com/zh/articles/2026-08-27-google-agent-platform-evaluations-ga-production-drift/https://ainews.imrenagi.com/zh/articles/2026-08-27-google-agent-platform-evaluations-ga-production-drift/Google 现在可以连接代理测试和生产监测。在比较分数之前，了解哪些版本、样本、评审者、轨迹、成本和控制措施必须保持可见。Thu, 27 Aug 2026 09:05:00 GMTai-newsai-applicationsgoogle-cloudai-agentsagent-evaluationsllm-as-a-judgeobservabilityproduction-monitoringhttps://ainews.imrenagi.com/zh/articles/2026-08-27-visual-studio-byom-agent-reliability-preview/https://ainews.imrenagi.com/zh/articles/2026-08-27-visual-studio-byom-agent-reliability-preview/Visual Studio 18.10 Insiders 可以将本地和托管模型连接到 Agent 模式，但一次实测表明，端点可用并不意味着编码工作可靠。Thu, 27 Aug 2026 05:09:30 GMTai-newsai-applicationsvisual-studiocoding-agentslocal-aiollamamodel-evaluationdeveloper-toolshttps://ainews.imrenagi.com/zh/articles/2026-08-27-asi-bench-autonomous-scientific-research-evaluation/https://ainews.imrenagi.com/zh/articles/2026-08-27-asi-bench-autonomous-scientific-research-evaluation/ASI-Bench 在人类指导逐步减少时测试科学代理。它最大的分数下降揭示了构建流程的问题，而不是关于超级智能的证据。Thu, 27 Aug 2026 03:05:24 GMTai-newsresearchasi-benchai-agentsai-for-sciencescientific-agentsbenchmarksevaluationshttps://ainews.imrenagi.com/zh/articles/2026-08-27-thomson-reuters-proprietary-legal-ai-model-explained/https://ainews.imrenagi.com/zh/articles/2026-08-27-thomson-reuters-proprietary-legal-ai-model-explained/Thomson Reuters 发布了用于 CoCounsel 的专有模型，也开放了一个较小版本的权重。以下是买方可以验证的内容，以及仍属于供应商主张的内容。Thu, 27 Aug 2026 01:08:00 GMTai-newsmodel-releasesthomson-reuterslegal-aienterprise-aimodel-evaluationopen-weight-modelsdomain-specific-modelshttps://ainews.imrenagi.com/zh/articles/2026-08-27-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-08-27-daily-digest/十项当前进展显示，AI 正从模型发布转向语音界面、工业机器人、代理沙箱、企业动作，以及围绕它们的电力和劳动力系统。Wed, 26 Aug 2026 23:03:14 GMTai-newsdaily-digestdaily-digestai-agentsai-infrastructuredeveloper-toolsai-securityphysical-aienterprise-aiai-economicshttps://ainews.imrenagi.com/zh/articles/2026-08-26-openai-jalapeno-inference-chip-benchmark-explained/https://ainews.imrenagi.com/zh/articles/2026-08-26-openai-jalapeno-inference-chip-benchmark-explained/OpenAI 的 Jalapeño 推理芯片在其发布的测试中领先，但延迟、功耗标准化、可用性和工作负载匹配度限制了采购结论。Wed, 26 Aug 2026 13:05:00 GMTai-newsai-applicationsopenaijalapenoinferenceai-chipsbenchmarksdata-centershttps://ainews.imrenagi.com/zh/articles/2026-08-26-twitch-amazon-ai-training-opt-out/https://ainews.imrenagi.com/zh/articles/2026-08-26-twitch-amazon-ai-training-opt-out/Twitch 现在允许创作者拒绝部分未来的 Amazon 生成式 AI 训练，但这个开关并不承诺删除历史数据，也不公开下游执行情况。Wed, 26 Aug 2026 11:03:11 GMTai-newspolicy-and-safetytwitchamazonai-training-dataprivacy-controlscreator-rightsdata-provenancehttps://ainews.imrenagi.com/zh/articles/2026-08-26-google-deepmind-sima-2-eve-online-offline-agent-evaluation/https://ainews.imrenagi.com/zh/articles/2026-08-26-google-deepmind-sima-2-eve-online-offline-agent-evaluation/DeepMind 正在离线环境中启动 EVE 代理研究。以下是持久世界为何需要针对记忆、规划、恢复和安全的不同测试。Wed, 26 Aug 2026 09:12:00 GMTai-newsresearchgoogle-deepmindsima-2eve-onlineai-agentsagent-evaluationslong-horizon-planninghttps://ainews.imrenagi.com/zh/articles/2026-08-26-apple-foundation-models-on-device-private-cloud-evaluation/https://ainews.imrenagi.com/zh/articles/2026-08-26-apple-foundation-models-on-device-private-cloud-evaluation/Apple 的通用 Swift 会话 API 可以访问端侧模型、Private Cloud Compute 或其他提供商，但这些路径在隐私、容量和可用性上各不相同。Wed, 26 Aug 2026 07:04:01 GMTai-newsai-applicationsapple-foundation-modelson-device-aiprivate-cloud-computemodel-routingevaluationswifthttps://ainews.imrenagi.com/zh/articles/2026-08-26-nvidia-avo-arc-agi-3-harness-result/https://ainews.imrenagi.com/zh/articles/2026-08-26-nvidia-avo-arc-agi-3-harness-result/NVIDIA AVO 通过了 ARC-AGI-3 的公开集，但该基准没有隔离工具链的贡献，也没有在私有任务上测试泛化。Wed, 26 Aug 2026 03:08:00 GMTai-newsresearchnvidia-avoarc-agi-3ai-agentsagent-harnessesbenchmarksevaluationshttps://ainews.imrenagi.com/zh/articles/2026-08-26-github-copilot-model-deprecations-fallback-matrix/https://ainews.imrenagi.com/zh/articles/2026-08-26-github-copilot-model-deprecations-fallback-matrix/六个 Copilot 模型将在 9 月 1 日退役，其中一个建议替代模型又将在九天后退役。产品名称比模型目录更稳定。Wed, 26 Aug 2026 01:04:00 GMTai-newsai-applicationsgithub-copilotmodel-deprecationdeveloper-toolsevaluationmigrationreliabilityhttps://ainews.imrenagi.com/zh/articles/2026-08-26-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-08-26-daily-digest/十二项最新进展显示，AI 正从模型发布走向定制芯片、持久代理上下文、代理优先基础设施、安全边界和可衡量的劳动影响。Tue, 25 Aug 2026 22:05:00 GMTai-newsdaily-digestdaily-digestai-agentsai-infrastructuredeveloper-toolsai-securityai-researchon-device-aienterprise-aihttps://ainews.imrenagi.com/zh/articles/2026-08-25-heir-homomorphic-encryption-private-inference-feasibility/https://ainews.imrenagi.com/zh/articles/2026-08-25-heir-homomorphic-encryption-private-inference-feasibility/Google 的 HEIR 编译器可以将多个模型转换为加密计算，但密文膨胀、内存、延迟和模型变化仍然是决定性因素。Tue, 25 Aug 2026 13:00:00 GMTai-newsai-applicationsheirhomomorphic-encryptionprivate-aiencrypted-inferencemlirprivacyevaluationhttps://ainews.imrenagi.com/zh/articles/2026-08-25-sentence-transformers-v6-multivector-retrieval-benchmark/https://ainews.imrenagi.com/zh/articles/2026-08-25-sentence-transformers-v6-multivector-retrieval-benchmark/Sentence Transformers v6 增加了多向量检索，而其自身结果显示，平均质量仅有适度提升，同时存储成本却大幅增加。Tue, 25 Aug 2026 01:00:00 GMTai-newsai-applicationssentence-transformerscolbertretrievalragembeddingsevaluationvector-searchhttps://ainews.imrenagi.com/zh/articles/2026-08-25-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-08-25-daily-digest/十二项带有来源链接的进展显示，AI 公司正在专业化模型、转移资本和人才，并让代理更多接触法律、企业、开发者和轨道系统。Mon, 24 Aug 2026 22:05:00 GMTai-newsdaily-digestdaily-digestai-agentsai-modelsdeveloper-toolsenterprise-aiai-infrastructureai-researchai-safetyhttps://ainews.imrenagi.com/zh/articles/2026-08-24-hugging-face-three-million-models-model-selection-workflow/https://ainews.imrenagi.com/zh/articles/2026-08-24-hugging-face-three-million-models-model-selection-workflow/Hugging Face 的模型数量已超过三百万，但点赞和下载量反映的是关注度，而不是许可证适配性、来源、硬件成本或任务质量。Mon, 24 Aug 2026 15:00:00 GMTai-newsmodel-releaseshugging-faceopen-weightsmodel-selectionmodel-evaluationmodel-cardsmodel-licenseslocal-inferencehttps://ainews.imrenagi.com/zh/articles/2026-08-24-meaningful-human-review-automated-decision-audit-log/https://ainews.imrenagi.com/zh/articles/2026-08-24-meaningful-human-review-automated-decision-audit-log/只有当审查员能够看到相反证据、暂停自动化路径、推翻自动化决定并纠正实际结果时，工作流中的人员才算进行了有意义的审查。Mon, 24 Aug 2026 13:00:00 GMTai-newspolicy-and-safetyhuman-reviewautomated-decisionsai-governanceaudit-logsalgorithmic-accountabilityappealsgdprhttps://ainews.imrenagi.com/zh/articles/2026-08-24-cloudflare-wrangler-mcp-optional-oauth-scopes/https://ainews.imrenagi.com/zh/articles/2026-08-24-cloudflare-wrangler-mcp-optional-oauth-scopes/Cloudflare 现在允许用户为 Wrangler 和 MCP 客户端拒绝可选 OAuth 作用域，但必需权限、令牌处理和工具行为仍是独立风险。Mon, 24 Aug 2026 11:00:00 GMTai-newspolicy-and-safetycloudflareoauthwranglermodel-context-protocolmcpleast-privilegeagent-securityhttps://ainews.imrenagi.com/zh/articles/2026-08-24-mcp-roadmap-deployment-compatibility-matrix/https://ainews.imrenagi.com/zh/articles/2026-08-24-mcp-roadmap-deployment-compatibility-matrix/MCP 的规范、扩展、SDK 和传输层正在以不同速度演进。路线图中的优先事项并不会让两个产品实现互操作。Mon, 24 Aug 2026 09:00:00 GMTai-newsai-applicationsmodel-context-protocolmcpsoftware-architecturesdkinteroperabilitydeploymenthttps://ainews.imrenagi.com/zh/articles/2026-08-24-ai-gateway-user-insights-anomaly-response/https://ainews.imrenagi.com/zh/articles/2026-08-24-ai-gateway-user-insights-anomaly-response/Cloudflare AI Gateway User Insights 可以检测异常的会话成本，但警报无法确定意图、归因每个请求，也无法遏制相关活动。Mon, 24 Aug 2026 07:08:00 GMTai-newspolicy-and-safetycloudflareai-gatewayanomaly-detectionincident-responseai-agentscost-governancehttps://ainews.imrenagi.com/zh/articles/2026-08-24-gemini-3-7-flash-thinking-level-cost-per-accepted-task/https://ainews.imrenagi.com/zh/articles/2026-08-24-gemini-3-7-flash-thinking-level-cost-per-accepted-task/Gemini 3.7 Flash 将于 2027 年 1 月涨价，而最便宜的推理等级仍取决于验收率、重试、延迟和回退成本。Mon, 24 Aug 2026 05:04:53 GMTai-newsmodel-releasesgeminimodel-evaluationapi-costthinking-levelsai-engineeringhttps://ainews.imrenagi.com/zh/articles/2026-08-24-llama-cpp-v0-2-0-stable-vs-nightly-local-inference/https://ainews.imrenagi.com/zh/articles/2026-08-24-llama-cpp-v0-2-0-stable-vs-nightly-local-inference/llama.cpp 现在在 nightly 构建之外也提供语义化版本发布。它的第一个稳定标签与一个构建标签共享代码，说明渠道名称描述的是策略，而不是质量。Mon, 24 Aug 2026 03:06:15 GMTai-newsai-applicationsllama-cpplocal-aiinferenceopen-sourceversioningbenchmarkshttps://ainews.imrenagi.com/zh/articles/2026-08-24-waymo-1000-tops-edge-ai-inference-budget/https://ainews.imrenagi.com/zh/articles/2026-08-24-waymo-1000-tops-edge-ai-inference-budget/Waymo 披露其定制边缘 AI 计算能力超过 1,000 TOPS，但没有公布精度、功耗、利用率、内存带宽和实测延迟。Mon, 24 Aug 2026 01:10:00 GMTai-newsai-applicationsedge-aiwaymoai-acceleratorsinferencebenchmarkingsystems-engineeringhttps://ainews.imrenagi.com/zh/articles/2026-08-24-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-08-24-daily-digest/十二项带来源链接的发展显示，AI 正进入受保护的工具服务器、可重复的代理基准测试、高效的本地推理、数据运营、物理世界和重视隐私的设备。Sun, 23 Aug 2026 22:12:00 GMTai-newsdaily-digestdaily-digestai-agentsdeveloper-toolsai-infrastructureai-evaluationlocal-aiai-researchprivacyhttps://ainews.imrenagi.com/zh/articles/2026-08-23-agent-harness-safety-lifecycle-test-matrix/https://ainews.imrenagi.com/zh/articles/2026-08-23-agent-harness-safety-lifecycle-test-matrix/HarnessRisk 报告不同模型与代理 harness 组合之间存在广泛的安全差异，并发现配置阶段比其测试的其他生命周期阶段更脆弱。Sun, 23 Aug 2026 13:03:00 GMTai-newspolicy-and-safetyai-agentsagent-harnessessecurity-testingprompt-injectionevaluationsincident-responsehttps://ainews.imrenagi.com/zh/articles/2026-08-23-claude-skills-api-security-review-checklist/https://ainews.imrenagi.com/zh/articles/2026-08-23-claude-skills-api-security-review-checklist/Claude Skills 可以捆绑指令、脚本和依赖。正式可用让不可变版本与运行时边界变得更加重要。Sun, 23 Aug 2026 09:06:00 GMTai-newspolicy-and-safetyclaudeagent-skillssoftware-supply-chainprompt-injectionsecurityhttps://ainews.imrenagi.com/zh/articles/2026-08-23-google-contrail-ai-airspace-trial/https://ainews.imrenagi.com/zh/articles/2026-08-23-google-contrail-ai-airspace-trial/蓝天行动将测试 AI 指导的高度变化能否减少繁忙北大西洋空域中的增温凝结尾迹，而不只是减少选定航班的尾迹。Sun, 23 Aug 2026 05:11:30 GMTai-newsai-applicationscontrailsaviationclimate-aievaluationair-traffic-controlhttps://ainews.imrenagi.com/zh/articles/2026-08-23-cross-model-kv-cache-transfer-compatibility-test/https://ainews.imrenagi.com/zh/articles/2026-08-23-cross-model-kv-cache-transfer-compatibility-test/NVIDIA 研究人员报告，一种 KV 缓存映射带来 25 倍的组件级加速，而其他几个模型对却损失了目标模型的大部分任务准确率。Sun, 23 Aug 2026 01:07:00 GMTmath-for-aimodels-and-datakv-cachellm-inferencemodel-routinglatencyevaluationhttps://ainews.imrenagi.com/zh/articles/2026-08-23-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-08-23-daily-digest/涵盖代理协议、开发者工具、模型定价、劳动力、监管和计算基础设施的十项有来源链接的 AI 进展。Sat, 22 Aug 2026 22:12:49 GMTai-newsdaily-digestdaily-digestai-agentsdeveloper-toolsai-policyai-infrastructurehttps://ainews.imrenagi.com/zh/articles/2026-08-22-stripe-openrouter-neutral-model-routing-audit/https://ainews.imrenagi.com/zh/articles/2026-08-22-stripe-openrouter-neutral-model-routing-audit/Stripe 已同意收购 OpenRouter。提供商选择、路由元数据、定价、隐私、回退行为和可移植性，说明了中立性意味着什么，也说明了它不能意味着什么。Sat, 22 Aug 2026 13:04:57 GMTai-newsai-applicationsopenrouterstripemodel-routingai-infrastructureai-governancehttps://ainews.imrenagi.com/zh/articles/2026-08-22-slack-code-agent-permissions-review-checklist/https://ainews.imrenagi.com/zh/articles/2026-08-22-slack-code-agent-permissions-review-checklist/Slack Code 让团队看见编码代理的工作，但 Slack 不负责控制仓库合并、CI 完整性、构建产物或生产部署。Sat, 22 Aug 2026 09:04:16 GMTai-newspolicy-and-safetyslackcoding-agentsgithubsoftware-supply-chaindevsecopshttps://ainews.imrenagi.com/zh/articles/2026-08-22-deepseek-v4-flash-vision-exp-pilot/https://ainews.imrenagi.com/zh/articles/2026-08-22-deepseek-v4-flash-vision-exp-pilot/DeepSeek 已将图像输入加入其快速 API 产品线，但预览版尚未证明它在截图、图表和文档上都能稳定表现。Sat, 22 Aug 2026 05:07:05 GMTai-newsmodel-releasesdeepseekvision-modelsapimodel-evaluationreliabilityhttps://ainews.imrenagi.com/zh/articles/2026-08-22-linus-ai-debugging-instrumentation-workflow/https://ainews.imrenagi.com/zh/articles/2026-08-22-linus-ai-debugging-instrumentation-workflow/AI 助手帮助 Linus Torvalds 迭代完成 24 个调试补丁并启动内核 18 次，但在证据找到错误之前，它反复劝他放弃调查。Sat, 22 Aug 2026 01:10:49 GMTai-newsai-applicationsai-debugginglinuxsoftware-engineeringcoding-agentsevaluationhttps://ainews.imrenagi.com/zh/articles/2026-08-22-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-08-22-daily-digest/涵盖代理架构、开发者工具、网络防御、多模态 API、开放推理和国家算力的十二项 AI 动态，均附有来源链接。Fri, 21 Aug 2026 22:12:01 GMTai-newsdaily-digestdaily-digestai-agentsdeveloper-toolscybersecurityai-infrastructurehttps://ainews.imrenagi.com/zh/articles/2026-08-21-grok-4-6-cost-per-task-evaluation/https://ainews.imrenagi.com/zh/articles/2026-08-21-grok-4-6-cost-per-task-evaluation/Grok 4.6 结合了较低的标价和报告中很强的基准成绩，但重试、失败任务、延迟和工具费用可能逆转表面上的优势。Fri, 21 Aug 2026 13:18:00 GMTai-newsmodel-releasesgrokmodel-evaluationbenchmarksapi-costcoding-agentshttps://ainews.imrenagi.com/zh/articles/2026-08-21-mojo-open-source-compiler-what-changed/https://ainews.imrenagi.com/zh/articles/2026-08-21-mojo-open-source-compiler-what-changed/Mojo 编译器现在可以从源代码构建和修改。AI News 复现了构建过程，但许可和贡献治理仍是分开的限制。Thu, 20 Aug 2026 23:18:34 GMTai-newsai-applicationsmojoopen-sourcecompilermlirdeveloper-toolsmaxhttps://ainews.imrenagi.com/zh/articles/2026-08-21-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-08-21-daily-digest/十二项动态连接了 AI 辅助的网络攻击、金融代理、合成网络内容、本地模型、监管、监控和气候研究。Thu, 20 Aug 2026 22:15:08 GMTai-newsdaily-digestdaily-digestai-securityai-agentslocal-aideveloper-toolsai-policyai-researchhttps://ainews.imrenagi.com/zh/articles/2026-08-20-openai-astra-safety-pause/https://ainews.imrenagi.com/zh/articles/2026-08-20-openai-astra-safety-pause/OpenAI 因 Astra 的网络安全担忧暂停前沿训练。了解暂停了什么、新监控如何工作，以及为什么 20% 的估算范围很窄。Thu, 20 Aug 2026 02:16:07 GMTai-newspolicy-and-safetyopenaiastracybersecurityai-safetymodel-monitoringreinforcement-learninghttps://ainews.imrenagi.com/zh/articles/2026-08-20-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-08-20-daily-digest/十二项进展连接起更快的推理硬件、模型路由、企业隐私、开发者平台、本地 AI 工具，以及基础设施政策。Wed, 19 Aug 2026 22:13:41 GMTai-newsdaily-digestdaily-digestai-infrastructuredeveloper-toolsopen-modelsai-safetyai-businessai-researchhttps://ainews.imrenagi.com/zh/articles/2026-08-20-alphaevolve-matrix-multiplication-exponent/https://ainews.imrenagi.com/zh/articles/2026-08-20-alphaevolve-matrix-multiplication-exponent/AlphaEvolve 帮助降低了已知最佳的矩阵乘法指数上界。了解改变了什么、如何认证，以及为什么它不会让今天的 GPU 更快。Wed, 19 Aug 2026 18:09:44 GMTmath-for-ailinear-algebraalphaevolvematrix-multiplicationcomputational-complexityoptimizationjaxhttps://ainews.imrenagi.com/zh/articles/2026-08-19-openai-ohio-data-center-8gw-800mw/https://ainews.imrenagi.com/zh/articles/2026-08-19-openai-ohio-data-center-8gw-800mw/OpenAI 在俄亥俄州的数据中心协议提出 8 GW 的目标，但首个明确阶段是 800 MW，且仍取决于电力、许可和建设。Wed, 19 Aug 2026 10:15:45 GMTai-newsai-applicationsopenaidata-centersai-infrastructurenvidiaenergyohiohttps://ainews.imrenagi.com/zh/articles/2026-08-19-claude-watermark-eu-ai-act-engineering/https://ainews.imrenagi.com/zh/articles/2026-08-19-claude-watermark-eu-ai-act-engineering/Claude 可以标记部分生成的文本和文件，但这些信号证明的是处理历史，而不是作者身份、真实性或完整的第 50 条合规性。Wed, 19 Aug 2026 02:11:57 GMTai-newspolicy-and-safetyclaudeanthropicai-watermarkingeu-ai-actcontent-provenanceai-compliancehttps://ainews.imrenagi.com/zh/articles/2026-08-19-google-spirit-airlines-ai-training-data-deal/https://ainews.imrenagi.com/zh/articles/2026-08-19-google-spirit-airlines-ai-training-data-deal/Google 对 Spirit Airlines 数据的竞价涵盖企业消息、代码和运营记录。去标识化解决个人身份问题，但无法单独解决机密性或下游使用问题。Wed, 19 Aug 2026 00:37:32 GMTai-newspolicy-and-safetygooglespirit-airlinestraining-datadeidentificationdata-governanceprivacyenterprise-aihttps://ainews.imrenagi.com/zh/articles/2026-08-19-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-08-19-daily-digest/十二项进展将前沿模型安全措施与青少年保护、开放开发者工具、专用芯片、运营故障和实际 AI 部署连接起来。Tue, 18 Aug 2026 22:13:56 GMTai-newsdaily-digestdaily-digestai-safetydeveloper-toolsai-infrastructureai-agentsai-businessclimate-techplatformshttps://ainews.imrenagi.com/zh/articles/2026-08-19-snowflake-github-actions-script-injection/https://ainews.imrenagi.com/zh/articles/2026-08-19-snowflake-github-actions-script-injection/不受信任的 issue 文本在 Snowflake 的 GitHub Actions 工作流中变成了 shell 代码。尽管是一个自主智能体发现并调整了这条路径，但注入方式本身是传统的。Tue, 18 Aug 2026 18:22:14 GMTai-newspolicy-and-safetygithub-actionsci-cdsecurityscript-injectionai-codingdevsecopshttps://ainews.imrenagi.com/zh/articles/2026-08-18-qwen3-8-27b-local-deployment-16gb/https://ainews.imrenagi.com/zh/articles/2026-08-18-qwen3-8-27b-local-deployment-16gb/紧凑的 Qwen3.8-27B 量化版本可以在 16GB GPU 上运行，但长上下文、视觉、并发和运行时内存使这个标题并不完整。Tue, 18 Aug 2026 16:19:24 GMTai-newsmodel-releasesqwenlocal-aiopen-modelsquantizationgpuinferencehttps://ainews.imrenagi.com/zh/articles/2026-08-18-daily-digest/https://ainews.imrenagi.com/zh/articles/2026-08-18-daily-digest/覆盖基础设施、数据、本地模型、安全、研究、经济学和版权的十项进展，展示了 AI 技术栈变化之快。Mon, 17 Aug 2026 23:49:55 GMTai-newsdaily-digestdaily-digestai-infrastructuretraining-dataopen-modelsai-securityworld-modelsai-economicscopyrighthttps://ainews.imrenagi.com/zh/articles/2026-08-17-vectors-to-embeddings/https://ainews.imrenagi.com/zh/articles/2026-08-17-vectors-to-embeddings/直观理解向量、点积、余弦相似度，以及嵌入为何能把含义转换成几何关系。Sun, 16 Aug 2026 17:00:00 GMTmath-for-ailinear-algebravectorsembeddingsbeginnerhttps://ainews.imrenagi.com/zh/articles/2026-08-16-evaluations-shape-ai-products/https://ainews.imrenagi.com/zh/articles/2026-08-16-evaluations-shape-ai-products/实践性地了解评估选择如何影响模型声明、产品决策，以及读者应当检查 AI 公告中的哪些内容。Sat, 15 Aug 2026 17:00:00 GMTai-newsai-applicationsevaluationsai-productsbenchmarks