{"id":391,"date":"2026-05-10T07:23:08","date_gmt":"2026-05-09T23:23:08","guid":{"rendered":"http:\/\/www.faiyi.com\/?p=391"},"modified":"2026-05-10T07:23:08","modified_gmt":"2026-05-09T23:23:08","slug":"ai%e5%8a%a8%e6%80%81%e6%af%8f%e6%97%a5%e7%ae%80%e6%8a%a5-2026-05-10","status":"publish","type":"post","link":"http:\/\/www.faiyi.com\/?p=391","title":{"rendered":"AI\u52a8\u6001\u6bcf\u65e5\u7b80\u62a5 2026-05-10"},"content":{"rendered":"

\u65e5\u671f\uff1a2026-05-10<\/p>\n

\u672c\u671f\u805a\u7126\uff1a\u91cd\u70b9\u5173\u6ce8\u6a21\u578b\u53d1\u5e03\u4e0e release notes\u3001\u5b98\u65b9 engineering blog\u3001AI coding \/ agent \/ SRE\u3001\u8bc4\u6d4b\u699c\u5355\u53d8\u5316\u3001\u5f00\u53d1\u8005\u5b9e\u8df5\u535a\u5ba2\u3001\u6846\u67b6\u751f\u6001\u3001\u5f00\u6e90\u6a21\u578b\u4e0e\u771f\u5b9e\u7528\u6237\u89c6\u89d2\uff1b\u5f53 HN\u3001Reddit\u3001Hugging Face \u7b49\u793e\u533a\u6e90\u53ef\u8bbf\u95ee\u65f6\u4f18\u5148\u7eb3\u5165\u3002<\/p>\n

\n
    \n
  1. \n

    Artificial Analysis \u6700\u65b0\u6a21\u578b\u6392\u540d\u89c2\u5bdf<\/strong>\uff08Artificial Analysis\uff09<\/p>\n

    \u4e2d\u6587\u6458\u8981\uff1a<\/strong>Artificial Analysis \u662f\u4e1a\u754c\u77e5\u540d\u7684\u7b2c\u4e09\u65b9 AI \u6a21\u578b\u8bc4\u6d4b\u5e73\u53f0\uff0c\u63d0\u4f9b\u6db5\u76d6\u667a\u80fd\u6c34\u5e73\u3001\u8f93\u51fa\u901f\u5ea6\u3001\u5ef6\u8fdf\u3001\u4ef7\u683c\u53ca\u4e0a\u4e0b\u6587\u7a97\u53e3\u7b49\u591a\u7ef4\u5ea6\u7684\u6a21\u578b\u5bf9\u6bd4\u670d\u52a1\u3002\u5176 Intelligence Index v4.0 \u7efc\u5408\u4e86 GDPval-AA\u3001Terminal-Bench Hard\u3001SciCode\u3001Humanity's Last Exam \u7b49\u591a\u9879\u6743\u5a01\u8bc4\u6d4b\u3002\u5f53\u524d\u699c\u5355\u663e\u793a\uff0cGPT-5.5 (xhigh) \u4ee5 60 \u5206\u4f4d\u5c45\u699c\u9996\uff0cClaude Opus 4.7 (Max Effort) \u4e0e Gemini 3.1 Pro Preview \u5e76\u5217 57 \u5206\u3002\u5f00\u6e90\u6a21\u578b\u65b9\u9762\uff0cKimi K2.6 \u4ee5 54 \u5206\u9886\u5148\u3002\u5e73\u53f0\u8fd8\u63d0\u4f9b\u4ef7\u683c-\u8d28\u91cf\u66f2\u7ebf\u5206\u6790\u3001\u7f13\u5b58\u5b9a\u4ef7\u5bf9\u6bd4\u53ca\u7aef\u5230\u7aef\u54cd\u5e94\u65f6\u95f4\u7b49\u5b9e\u7528\u6570\u636e\uff0c\u5e2e\u52a9\u5f00\u53d1\u8005\u548c\u4f01\u4e1a\u6839\u636e\u5b9e\u9645\u9700\u6c42\u9009\u62e9\u6700\u9002\u5408\u7684\u6a21\u578b\u3002<\/p>\n

    English Summary:<\/strong> Artificial Analysis is a leading third-party AI model evaluation platform offering multi-dimensional model comparisons across intelligence, output speed, latency, pricing, and context windows. Its Intelligence Index v4.0 aggregates benchmarks including GDPval-AA, Terminal-Bench Hard, SciCode, and Humanity's Last Exam. Current rankings show GPT-5.5 (xhigh) leading with a score of 60, while Claude Opus 4.7 (Max Effort) and Gemini 3.1 Pro Preview tie at 57. Among open weights models, Kimi K2.6 leads with 54. The platform also provides price-quality curve analysis, cache pricing comparisons, and end-to-end response time metrics to help developers and enterprises select optimal models for their needs.<\/p>\n

    \u539f\u6587\u94fe\u63a5<\/a><\/p>\n<\/li>\n

  2. \n

    Introducing Claude Opus 4.7<\/strong>\uff08Anthropic News\uff09<\/p>\n

    \u4e2d\u6587\u6458\u8981\uff1a<\/strong>Anthropic \u6b63\u5f0f\u53d1\u5e03 Claude Opus 4.7\uff0c\u8fd9\u662f Opus 4.6 \u7684\u91cd\u5927\u5347\u7ea7\u7248\u672c\uff0c\u5728\u9ad8\u7ea7\u8f6f\u4ef6\u5de5\u7a0b\u4efb\u52a1\u4e0a\u8868\u73b0\u5c24\u4e3a\u7a81\u51fa\u3002\u8be5\u6a21\u578b\u5728\u590d\u6742\u3001\u957f\u65f6\u95f4\u8fd0\u884c\u7684\u4efb\u52a1\u4e2d\u5c55\u73b0\u51fa\u66f4\u5f3a\u7684\u4e25\u8c28\u6027\u548c\u4e00\u81f4\u6027\uff0c\u80fd\u591f\u7cbe\u786e\u9075\u5faa\u6307\u4ee4\u5e76\u5728\u62a5\u544a\u524d\u81ea\u884c\u9a8c\u8bc1\u8f93\u51fa\u3002Opus 4.7 \u7684\u89c6\u89c9\u80fd\u529b\u663e\u8457\u63d0\u5347\uff0c\u652f\u6301\u9ad8\u8fbe 2576 \u50cf\u7d20\u7684\u957f\u8fb9\u5206\u8fa8\u7387\uff08\u7ea6 375 \u4e07\u50cf\u7d20\uff09\uff0c\u662f\u524d\u4ee3\u6a21\u578b\u7684\u4e09\u500d\u4ee5\u4e0a\u3002\u5728\u4e13\u4e1a\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u66f4\u4f73\u7684\u5ba1\u7f8e\u548c\u521b\u9020\u529b\uff0c\u80fd\u751f\u6210\u66f4\u9ad8\u8d28\u91cf\u7684\u754c\u9762\u3001\u5e7b\u706f\u7247\u548c\u6587\u6863\u3002Anthropic \u8fd8\u5f15\u5165\u4e86\u65b0\u7684 "xhigh" \u52aa\u529b\u7ea7\u522b\uff0c\u5e76\u5728 Claude Code \u4e2d\u63a8\u51fa "\/ultrareview" \u547d\u4ee4\u548c\u6269\u5c55\u7684 auto mode \u529f\u80fd\u3002\u8be5\u6a21\u578b\u5df2\u5168\u9762\u4e0a\u7ebf Claude \u4ea7\u54c1\u3001API \u53ca\u5404\u5927\u4e91\u5e73\u53f0\u3002<\/p>\n

    English Summary:<\/strong> Anthropic officially released Claude Opus 4.7, a major upgrade from Opus 4.6 with notable improvements in advanced software engineering, particularly on the most difficult tasks. The model demonstrates greater rigor and consistency on complex, long-running tasks, follows instructions precisely, and verifies its own outputs before reporting back. Opus 4.7 features substantially enhanced vision capabilities, supporting images up to 2,576 pixels on the long edge (~3.75 megapixels), more than triple previous Claude models. It shows better taste and creativity on professional tasks, producing higher-quality interfaces, slides, and docs. Anthropic also introduced a new "xhigh" effort level, along with "\/ultrareview" command and expanded auto mode in Claude Code. The model is available across all Claude products, API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.<\/p>\n

    \u539f\u6587\u94fe\u63a5<\/a><\/p>\n<\/li>\n

  3. \n

    Featured An update on recent Claude Code quality reports<\/strong>\uff08Anthropic Engineering\uff09<\/p>\n

    \u4e2d\u6587\u6458\u8981\uff1a<\/strong>Anthropic \u5de5\u7a0b\u56e2\u961f\u53d1\u5e03\u5173\u4e8e\u8fd1\u671f Claude Code \u8d28\u91cf\u95ee\u9898\u7684\u8be6\u7ec6\u590d\u76d8\u62a5\u544a\uff0c\u786e\u8ba4\u4e09\u4e2a\u72ec\u7acb\u95ee\u9898\u5bfc\u81f4\u7528\u6237\u4f53\u9a8c\u4e0b\u964d\uff0c\u5e76\u5df2\u5168\u90e8\u4fee\u590d\u3002\u7b2c\u4e00\u4e2a\u95ee\u9898\u662f 3 \u6708 4 \u65e5\u5c06\u9ed8\u8ba4\u63a8\u7406\u52aa\u529b\u7ea7\u522b\u4ece "high" \u6539\u4e3a "medium"\uff0c\u5bfc\u81f4\u6a21\u578b\u8868\u73b0\u4e0b\u964d\uff0c\u5df2\u4e8e 4 \u6708 7 \u65e5\u56de\u6eda\uff0c\u73b0\u9ed8\u8ba4\u4f7f\u7528 "xhigh"\uff08Opus 4.7\uff09\u6216 "high"\uff08\u5176\u4ed6\u6a21\u578b\uff09\u3002\u7b2c\u4e8c\u4e2a\u95ee\u9898\u662f 3 \u6708 26 \u65e5\u5b9e\u65bd\u7684\u7f13\u5b58\u4f18\u5316\u5b58\u5728 bug\uff0c\u5bfc\u81f4\u4f1a\u8bdd\u95f2\u7f6e\u8d85\u8fc7\u4e00\u5c0f\u65f6\u540e\u4f1a\u6301\u7eed\u6e05\u9664\u5386\u53f2\u63a8\u7406\u8bb0\u5f55\uff0c\u4f7f Claude \u663e\u5f97\u5065\u5fd8\u548c\u91cd\u590d\uff0c\u5df2\u4e8e 4 \u6708 10 \u65e5\u4fee\u590d\u3002\u7b2c\u4e09\u4e2a\u95ee\u9898\u662f 4 \u6708 16 \u65e5\u6dfb\u52a0\u7684\u51cf\u5c11\u5197\u957f\u8f93\u51fa\u7684\u7cfb\u7edf\u63d0\u793a\u8bcd\u610f\u5916\u5f71\u54cd\u4e86\u7f16\u7801\u8d28\u91cf\uff0c\u5df2\u4e8e 4 \u6708 20 \u65e5\u56de\u6eda\u3002Anthropic \u5411\u6240\u6709\u8ba2\u9605\u8005\u91cd\u7f6e\u4f7f\u7528\u9650\u989d\uff0c\u5e76\u627f\u8bfa\u6539\u8fdb\u5185\u90e8\u6d4b\u8bd5\u6d41\u7a0b\u548c\u4ee3\u7801\u5ba1\u67e5\u673a\u5236\u3002<\/p>\n

    English Summary:<\/strong> Anthropic's engineering team published a detailed postmortem on recent Claude Code quality issues, confirming three separate problems that degraded user experience, all now resolved. First, on March 4, the default reasoning effort was changed from "high" to "medium," reducing model performance; this was reverted on April 7, with current defaults now set to "xhigh" for Opus 4.7 and "high" for other models. Second, a caching optimization shipped on March 26 contained a bug that continuously cleared historical reasoning after sessions were idle for over an hour, causing Claude to appear forgetful and repetitive; fixed on April 10. Third, a system prompt change on April 16 to reduce verbosity inadvertently hurt coding quality; reverted on April 20. Anthropic reset usage limits for all subscribers and committed to improving internal testing and code review processes.<\/p>\n

    \u539f\u6587\u94fe\u63a5<\/a><\/p>\n<\/li>\n

  4. \n

    Scaling Managed Agents: Decoupling the brain from the hands<\/strong>\uff08Anthropic Engineering\uff09<\/p>\n

    \u4e2d\u6587\u6458\u8981\uff1a<\/strong>Anthropic \u5de5\u7a0b\u535a\u5ba2\u6df1\u5165\u4ecb\u7ecd\u4e86 Managed Agents \u7684\u67b6\u6784\u8bbe\u8ba1\u7406\u5ff5\u2014\u2014\u5c06"\u5927\u8111"\uff08Claude \u53ca\u5176 harness\uff09\u4e0e"\u624b"\uff08\u6c99\u7bb1\u548c\u5de5\u5177\uff09\u89e3\u8026\u3002\u8fd9\u79cd\u8bbe\u8ba1\u501f\u9274\u4e86\u64cd\u4f5c\u7cfb\u7edf\u865a\u62df\u5316\u786c\u4ef6\u7684\u7ecf\u5178\u601d\u8def\uff0c\u901a\u8fc7\u5b9a\u4e49\u901a\u7528\u63a5\u53e3\uff08session\u3001harness\u3001sandbox\uff09\u4f7f\u5404\u7ec4\u4ef6\u53ef\u72ec\u7acb\u6f14\u8fdb\u548c\u66ff\u6362\u3002\u89e3\u8026\u524d\uff0c\u6240\u6709\u7ec4\u4ef6\u8fd0\u884c\u4e8e\u5355\u4e00\u5bb9\u5668\uff0c\u5bfc\u81f4\u6545\u969c\u6392\u67e5\u56f0\u96be\u3001\u5b89\u5168\u8fb9\u754c\u6a21\u7cca\u53ca\u542f\u52a8\u5ef6\u8fdf\u9ad8\u3002\u89e3\u8026\u540e\uff0charness \u4ee5\u65e0\u72b6\u6001\u65b9\u5f0f\u8fd0\u884c\uff0c\u901a\u8fc7\u5de5\u5177\u8c03\u7528\u4e0e\u6c99\u7bb1\u4ea4\u4e92\uff0c\u5bb9\u5668\u6545\u969c\u53ef\u88ab\u6355\u83b7\u5e76\u91cd\u8bd5\uff1b\u4f1a\u8bdd\u65e5\u5fd7\u6301\u4e45\u5316\u5b58\u50a8\uff0c\u652f\u6301\u4ece\u4efb\u610f\u70b9\u6062\u590d\uff1b\u51ed\u8bc1\u4e0e\u6267\u884c\u73af\u5883\u5206\u79bb\uff0c\u6d88\u9664\u4e86\u63d0\u793a\u8bcd\u6ce8\u5165\u5bfc\u81f4\u51ed\u8bc1\u6cc4\u9732\u7684\u98ce\u9669\u3002\u8be5\u67b6\u6784\u4f7f p50 \u9996 token \u5ef6\u8fdf\u964d\u4f4e\u7ea6 60%\uff0cp95 \u964d\u4f4e\u8d85 90%\uff0c\u5e76\u652f\u6301\u591a brain \u4e0e\u591a hand \u7684\u7075\u6d3b\u7ec4\u5408\u3002<\/p>\n

    English Summary:<\/strong> Anthropic's engineering blog details the architectural design of Managed Agents, decoupling the "brain" (Claude and its harness) from the "hands" (sandboxes and tools). Inspired by operating systems' virtualization of hardware, this approach defines generic interfaces (session, harness, sandbox) allowing components to evolve and be replaced independently. Previously, all components ran in a single container, making debugging difficult, security boundaries unclear, and startup latency high. After decoupling, harnesses run statelessly and interact with sandboxes via tool calls; container failures are caught and retrievable; session logs are durably stored for recovery from any point; credentials are separated from execution environments, eliminating prompt injection risks. This architecture reduced p50 time-to-first-token latency by ~60% and p95 by over 90%, while supporting flexible combinations of multiple brains and hands.<\/p>\n

    \u539f\u6587\u94fe\u63a5<\/a><\/p>\n<\/li>\n

  5. \n

    So you\u2019ve heard these AI terms and nodded along; let\u2019s fix that<\/strong>\uff08TechCrunch AI\uff09<\/p>\n

    \u4e2d\u6587\u6458\u8981\uff1a<\/strong>TechCrunch \u53d1\u5e03\u4e86\u4e00\u4efd\u5168\u9762\u7684 AI \u672f\u8bed\u8bcd\u6c47\u8868\uff0c\u65e8\u5728\u5e2e\u52a9\u8bfb\u8005\u7406\u89e3\u4eba\u5de5\u667a\u80fd\u9886\u57df\u4e0d\u65ad\u6d8c\u73b0\u7684\u4e13\u4e1a\u672f\u8bed\u3002\u6587\u7ae0\u6db5\u76d6\u4e86\u4ece\u57fa\u7840\u6982\u5ff5\u5230\u524d\u6cbf\u6280\u672f\u7684 20 \u4f59\u4e2a\u5173\u952e\u672f\u8bed\uff0c\u5305\u62ec AGI\uff08\u901a\u7528\u4eba\u5de5\u667a\u80fd\uff09\u3001AI Agent\uff08\u667a\u80fd\u4f53\uff09\u3001Chain of Thought\uff08\u601d\u7ef4\u94fe\uff09\u3001Coding Agents\uff08\u7f16\u7801\u667a\u80fd\u4f53\uff09\u3001Deep Learning\uff08\u6df1\u5ea6\u5b66\u4e60\uff09\u3001Diffusion\uff08\u6269\u6563\u6a21\u578b\uff09\u3001Distillation\uff08\u77e5\u8bc6\u84b8\u998f\uff09\u3001Fine-tuning\uff08\u5fae\u8c03\uff09\u3001Hallucination\uff08\u5e7b\u89c9\uff09\u3001Inference\uff08\u63a8\u7406\uff09\u3001LLM\uff08\u5927\u8bed\u8a00\u6a21\u578b\uff09\u3001Neural Network\uff08\u795e\u7ecf\u7f51\u7edc\uff09\u3001Open Source\uff08\u5f00\u6e90\uff09\u3001Reinforcement Learning\uff08\u5f3a\u5316\u5b66\u4e60\uff09\u3001Token\uff08\u8bcd\u5143\uff09\u3001Training\uff08\u8bad\u7ec3\uff09\u3001Weights\uff08\u6743\u91cd\uff09\u7b49\u3002\u6bcf\u4e2a\u672f\u8bed\u90fd\u914d\u6709\u901a\u4fd7\u6613\u61c2\u7684\u89e3\u91ca\u548c\u5b9e\u9645\u5e94\u7528\u573a\u666f\u8bf4\u660e\uff0c\u662f\u4e00\u4efd\u9002\u5408\u6280\u672f\u4eba\u5458\u548c\u975e\u6280\u672f\u4eba\u5458\u53c2\u8003\u7684\u5b9e\u7528\u6307\u5357\u3002<\/p>\n

    English Summary:<\/strong> TechCrunch published a comprehensive glossary of AI terms to help readers understand the ever-growing specialized vocabulary in artificial intelligence. The article covers over 20 key terms ranging from foundational concepts to cutting-edge technologies, including AGI (Artificial General Intelligence), AI Agent, Chain of Thought, Coding Agents, Deep Learning, Diffusion, Distillation, Fine-tuning, Hallucination, Inference, LLM (Large Language Model), Neural Network, Open Source, Reinforcement Learning, Token, Training, and Weights. Each term includes accessible explanations and real-world application contexts, making it a practical reference for both technical and non-technical readers navigating the AI landscape.<\/p>\n

    \u539f\u6587\u94fe\u63a5<\/a><\/p>\n<\/li>\n

  6. \n

    Cloudflare Ships Dynamic Workflows, Bringing Durable Execution to Per-Tenant and Per-Agent Code<\/strong>\uff08InfoQ AI\/ML\uff09<\/p>\n

    \u4e2d\u6587\u6458\u8981\uff1a<\/strong>Cloudflare \u53d1\u5e03 Dynamic Workflows\uff0c\u4e00\u4e2a MIT \u8bb8\u53ef\u7684\u5f00\u6e90\u5e93\uff0c\u6269\u5c55\u5176\u6301\u4e45\u5316\u6267\u884c\u5f15\u64ce\u4ee5\u652f\u6301\u6309\u79df\u6237\u3001\u4ee3\u7406\u6216\u8bf7\u6c42\u52a8\u6001\u52a0\u8f7d\u5de5\u4f5c\u6d41\u4ee3\u7801\u3002\u6b64\u524d Cloudflare Workflows \u8981\u6c42\u5de5\u4f5c\u6d41\u4ee3\u7801\u968f\u90e8\u7f72\u7ed1\u5b9a\uff0c\u800c Dynamic Workflows \u901a\u8fc7 Worker Loader \u5728\u8fd0\u884c\u65f6\u8def\u7531\u5230\u5bf9\u5e94\u79df\u6237\u7684\u4ee3\u7801\uff0c\u5b9e\u73b0\u6b65\u9aa4\u4f11\u7720\u3001\u4e8b\u4ef6\u7b49\u5f85\u7b49\u7279\u6027\u4e0d\u53d8\u3002\u8be5\u65b9\u6848\u4e0e Artifacts\u3001Dynamic Workers\u3001Sandboxes \u7ec4\u5408\uff0c\u53ef\u5c06 CI\/CD \u7b49\u573a\u666f\u7684\u51b7\u542f\u52a8\u65f6\u95f4\u4ece\u5206\u949f\u7ea7\u964d\u81f3\u6beb\u79d2\u7ea7\uff0c\u4f7f\u5e73\u53f0\u80fd\u4ee5\u63a5\u8fd1\u96f6\u95f2\u7f6e\u6210\u672c\u670d\u52a1\u6570\u5343\u4e07\u79df\u6237\u3002<\/p>\n

    English Summary:<\/strong> Cloudflare released Dynamic Workflows, an MIT-licensed library that extends its durable execution engine to support per-tenant, per-agent, or per-request dynamic code loading at runtime. Unlike the previous requirement of binding workflow code at deployment, a Worker Loader now routes execution to the correct tenant's code when the engine wakes up.<\/p>\n

    \u539f\u6587\u94fe\u63a5<\/a><\/p>\n<\/li>\n

  7. \n

    [AINews] Anthropic growing 10x\/year while everyone else is laying off >10% of their workforce<\/strong>\uff08Latent Space\uff09<\/p>\n

    \u4e2d\u6587\u6458\u8981\uff1a<\/strong>\u636e\u4e8c\u7ea7\u5e02\u573a\u53ca\u4f20\u7edf\u5a92\u4f53\u62a5\u9053\uff0cAnthropic \u5728\u300c\u5947\u8ff9\u822c\u7684\u7b2c\u4e00\u5b63\u5ea6\u300d\u5b9e\u73b0\u5e74\u5316\u589e\u957f 80 \u500d\u540e\uff0c\u4f30\u503c\u5df2\u8fbe 1\u20131.2 \u4e07\u4ebf\u7f8e\u5143\uff0c\u8d85\u8d8a OpenAI \u6210\u4e3a\u5168\u7403\u7b2c 11\u201315 \u5927\u6700\u6709\u4ef7\u503c\u516c\u53f8\u3002\u4e0e\u6b64\u540c\u65f6\uff0cBlock\u3001Coinbase\u3001Cloudflare \u7b49\u516c\u53f8\u5374\u4ee5\u300cAI \u5c31\u7eea\u300d\u4e3a\u7531\u88c1\u5458 10%\u201340% \u4ee5\u4e0a\uff0c\u5f62\u6210\u9c9c\u660e\u5bf9\u6bd4\u3002\u6587\u7ae0\u6307\u51fa\u771f\u6b63\u7684 AI \u7ea2\u5229\u76ee\u524d\u4e3b\u8981\u96c6\u4e2d\u5728\u786c\u4ef6\u4e0e\u80fd\u6e90\u9886\u57df\uff0c\u8f6f\u4ef6\u884c\u4e1a\u5c1a\u672a\u540c\u7b49\u53d7\u76ca\uff0c\u7ecf\u6d4e\u96c6\u4e2d\u5ea6\u6b63\u903c\u8fd1\u6ce1\u6cab\u533a\u95f4\u3002<\/p>\n

    English Summary:<\/strong> According to secondary market and traditional media reports, Anthropic is now valued at $1\u20131.2 trillion after an 80x annualized growth "miracle Q1," officially overtaking OpenAI as the 11th\u201315th most valuable company globally. This contrasts sharply with layoffs at Block (40%), Coinbase (14%), and Cloudflare (20%), all citing AI readiness. The article notes that current AI growth has mostly benefited hardware and energy rather than software, pushing economic concentration toward bubble territory.<\/p>\n

    \u539f\u6587\u94fe\u63a5<\/a><\/p>\n<\/li>\n

  8. \n

    Halliburton enhances seismic workflow creation with Amazon Bedrock and Generative AI<\/strong>\uff08AWS ML Blog\uff09<\/p>\n

    \u4e2d\u6587\u6458\u8981\uff1a<\/strong>Halliburton \u4e0e AWS \u751f\u6210\u5f0f AI \u521b\u65b0\u4e2d\u5fc3\u5408\u4f5c\uff0c\u57fa\u4e8e Amazon Bedrock\u3001Nova\u3001Knowledge Bases \u548c DynamoDB \u6784\u5efa\u5730\u9707\u6570\u636e\u5904\u7406\u5de5\u4f5c\u6d41\u52a9\u624b\u3002\u8be5\u7cfb\u7edf\u901a\u8fc7\u610f\u56fe\u8def\u7531\u5c06\u81ea\u7136\u8bed\u8a00\u67e5\u8be2\u5206\u7c7b\u4e3a\u5de5\u4f5c\u6d41\u751f\u6210\u6216\u6280\u672f\u95ee\u7b54\uff0c\u5229\u7528 Claude 3.5 \u751f\u6210\u53ef\u6267\u884c\u7684 YAML \u5de5\u4f5c\u6d41\uff0c\u5b9e\u73b0\u6210\u529f\u7387 84\u201397%\uff0c\u5c06\u539f\u672c\u9700\u8981\u6570\u5206\u949f\u7684\u624b\u5de5\u914d\u7f6e\u7f29\u77ed\u81f3\u79d2\u7ea7\uff0c\u6548\u7387\u63d0\u5347\u8d85\u8fc7 95%\u3002<\/p>\n

    English Summary:<\/strong> Halliburton partnered with the AWS Generative AI Innovation Center to build an AI assistant for seismic data processing using Amazon Bedrock, Nova, Knowledge Bases, and DynamoDB. An intent router classifies natural language queries into workflow generation or Q&A, with Claude 3.5 generating executable YAML workflows. The solution achieves 84\u201397% success rates and reduces workflow creation time from minutes to seconds, representing over 95% efficiency improvement.<\/p>\n

    \u539f\u6587\u94fe\u63a5<\/a><\/p>\n<\/li>\n

  9. \n

    Running Codex safely at OpenAI<\/strong>\uff08OpenAI News\uff09<\/p>\n

    \u4e2d\u6587\u6458\u8981\uff1a<\/strong>OpenAI \u53d1\u5e03\u5b98\u65b9\u535a\u5ba2\uff0c\u9610\u8ff0\u5176\u5185\u90e8\u5982\u4f55\u5b89\u5168\u90e8\u7f72 Codex \u7f16\u7801\u4ee3\u7406\u3002\u6838\u5fc3\u63aa\u65bd\u5305\u62ec\uff1a\u901a\u8fc7\u6c99\u7bb1\u4e0e\u5ba1\u6279\u7b56\u7565\u63a7\u5236\u6267\u884c\u8fb9\u754c\uff0c\u5229\u7528 Auto-review \u6a21\u5f0f\u81ea\u52a8\u6279\u51c6\u4f4e\u98ce\u9669\u64cd\u4f5c\uff1b\u5b9e\u65bd\u6258\u7ba1\u7f51\u7edc\u7b56\u7565\u9650\u5236\u51fa\u7ad9\u8bbf\u95ee\uff1b\u4f7f\u7528 OS \u5bc6\u94a5\u94fe\u5b58\u50a8 CLI \u4e0e MCP OAuth \u51ed\u8bc1\u5e76\u5f3a\u5236 ChatGPT \u4f01\u4e1a\u5de5\u4f5c\u7a7a\u95f4\u767b\u5f55\uff1b\u901a\u8fc7\u89c4\u5219\u533a\u5206\u5b89\u5168\u4e0e\u5371\u9669\u547d\u4ee4\uff1b\u4ee5\u53ca\u5bfc\u51fa OpenTelemetry \u65e5\u5fd7\u5b9e\u73b0\u4ee3\u7406\u539f\u751f\u53ef\u89c2\u6d4b\u6027\u4e0e\u5ba1\u8ba1\u8ffd\u8e2a\u3002<\/p>\n

    English Summary:<\/strong> OpenAI published a blog post detailing how it safely deploys the Codex coding agent internally. Key controls include sandboxing and approval policies with Auto-review for low-risk actions, managed network policies restricting outbound access, OS keychain storage for CLI and MCP OAuth credentials tied to ChatGPT enterprise workspace, rule-based command safety classification, and OpenTelemetry log exports for agent-native observability and audit trails.<\/p>\n

    \u539f\u6587\u94fe\u63a5<\/a><\/p>\n<\/li>\n

  10. \n

    [AINews] GPT-Realtime-2, -Translate, and -Whisper: new SOTA realtime voice APIs<\/strong>\uff08Latent Space\uff09<\/p>\n

    \u4e2d\u6587\u6458\u8981\uff1a<\/strong>OpenAI \u5728 Realtime API \u53d1\u5e03\u4e09\u6b3e\u65b0\u8bed\u97f3\u6a21\u578b\uff1aGPT-Realtime-2\u3001GPT-Realtime-Translate \u548c GPT-Realtime-Whisper\u3002GPT-Realtime-2 \u652f\u6301 GPT-5 \u7ea7\u63a8\u7406\u3001128K \u4e0a\u4e0b\u6587\u3001\u4e94\u7ea7\u53ef\u8c03\u63a8\u7406\u5f3a\u5ea6\u3001\u5e76\u884c\u5de5\u5177\u8c03\u7528\u4e0e\u53ef\u542c\u5316\u53cd\u9988\uff0c\u5728 Big Bench Audio \u4e0a\u8fbe 96.6%\uff0c\u6307\u4ee4\u4fdd\u6301\u7387\u4ece 36.7% \u63d0\u5347\u81f3 70.8%\u3002Translate \u652f\u6301 70 \u4f59\u79cd\u8f93\u5165\u8bed\u8a00\u5b9e\u65f6\u7ffb\u8bd1\u4e3a 13 \u79cd\u8f93\u51fa\u8bed\u8a00\uff0cWhisper \u63d0\u4f9b\u6d41\u5f0f\u8f6c\u5199\u3002Glean\u3001Vimeo\u3001Genspark \u7b49\u5df2\u96c6\u6210\u3002<\/p>\n

    English Summary:<\/strong> OpenAI released three new voice models in the Realtime API: GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper. GPT-Realtime-2 features GPT-5-class reasoning, 128K context, five adjustable reasoning levels, parallel tool calls with audible transparency, scoring 96.6% on Big Bench Audio and improving instruction retention from 36.7% to 70.8%. Translate supports live speech translation from 70+ input to 13 output languages, while Whisper provides streaming transcription. Glean, Vimeo, and Genspark have already integrated the models.<\/p>\n

    \u539f\u6587\u94fe\u63a5<\/a><\/p>\n<\/li>\n

  11. \n

    Improving token efficiency in GitHub Agentic Workflows<\/strong>\uff08GitHub AI\/ML\uff09<\/p>\n

    \u4e2d\u6587\u6458\u8981\uff1a<\/strong>GitHub \u5de5\u7a0b\u56e2\u961f\u5206\u4eab\u4e86\u4f18\u5316 Agentic Workflows \u4ee4\u724c\u6548\u7387\u7684\u5b9e\u8df5\u7ecf\u9a8c\u3002\u7531\u4e8e\u6bcf\u6b21 Pull Request \u90fd\u4f1a\u89e6\u53d1\u4ee3\u7406\u5de5\u4f5c\u6d41\uff0cAPI \u8d39\u7528\u53ef\u80fd\u6084\u7136\u7d2f\u79ef\u3002\u56e2\u961f\u901a\u8fc7 API \u4ee3\u7406\u7edf\u4e00\u8bb0\u5f55\u4ee4\u724c\u4f7f\u7528\u6570\u636e\uff0c\u5e76\u6784\u5efa\u4e86\u4e24\u4e2a\u81ea\u52a8\u5316\u5de5\u4f5c\u6d41\uff1a\u6bcf\u65e5\u4ee4\u724c\u5ba1\u8ba1\u5668\uff08Token Auditor\uff09\u6807\u8bb0\u5f02\u5e38\u9ad8\u6d88\u8017\u7684 workflow\uff0c\u6bcf\u65e5\u4ee4\u724c\u4f18\u5316\u5668\uff08Token Optimizer\uff09\u5206\u6790\u6e90\u7801\u5e76\u63d0\u51fa\u5177\u4f53\u4f18\u5316\u5efa\u8bae\u3002\u4e3b\u8981\u4f18\u5316\u624b\u6bb5\u5305\u62ec\uff1a\u79fb\u9664\u672a\u4f7f\u7528\u7684 MCP \u5de5\u5177\uff08\u53ef\u51cf\u5c11 8-12 KB \u4e0a\u4e0b\u6587\uff09\u3001\u7528 GitHub CLI \u66ff\u4ee3 MCP \u8c03\u7528\u4ee5\u6d88\u9664 LLM \u63a8\u7406\u5f00\u9500\uff0c\u4ee5\u53ca\u5c06\u786e\u5b9a\u6027\u6570\u636e\u83b7\u53d6\u79fb\u81f3\u4ee3\u7406\u542f\u52a8\u524d\u7684\u9884\u6267\u884c\u6b65\u9aa4\u3002\u56e2\u961f\u8fd8\u63d0\u51fa\u4e86"\u6709\u6548\u4ee4\u724c\uff08ET\uff09"\u6307\u6807\u6765\u6807\u51c6\u5316\u4e0d\u540c\u6a21\u578b\u7684\u6210\u672c\u6bd4\u8f83\u3002\u5b9e\u9645\u90e8\u7f72\u540e\uff0cAuto-Triage Issues \u5de5\u4f5c\u6d41\u8282\u7701\u4e86 62% \u7684 ET\uff0cSecurity Guard \u548c Smoke Claude \u5206\u522b\u8282\u7701 43% \u548c 59%\u3002\u6587\u7ae0\u5f3a\u8c03\uff0c\u6700\u4fbf\u5b9c\u7684 LLM \u8c03\u7528\u662f\u4e0d\u5fc5\u8981\u7684\u8c03\u7528\uff0c\u672a\u6765\u5c06\u4ece\u5de5\u4f5c\u6d41\u7ea7\u4f18\u5316\u8f6c\u5411\u7cfb\u7edf\u7ea7\u548c\u7ec4\u5408\u7ea7\u4f18\u5316\u3002<\/p>\n

    English Summary:<\/strong> GitHub's engineering team shares their experience optimizing token efficiency for Agentic Workflows. Since these workflows run on every pull request, API costs can accumulate quietly. The team implemented API proxy logging to capture token usage across all agent frameworks uniformly, then built two automated workflows: a Daily Token Auditor flags workflows with anomalous usage, and a Daily Token Optimizer analyzes source code to propose specific fixes. Key optimizations include removing unused MCP tools (saving 8-12 KB context), replacing MCP calls with GitHub CLI to eliminate LLM reasoning overhead, and moving deterministic data fetching to pre-agent setup steps. They introduced an "Effective Tokens (ET)" metric to normalize costs across different models. Results show Auto-Triage Issues reduced ET by 62%, Security Guard by 43%, and Smoke Claude by 59%. The post emphasizes that the cheapest LLM call is the one you don't make, with future work targeting system-level and portfolio-level optimizations.<\/p>\n

    \u539f\u6587\u94fe\u63a5<\/a><\/p>\n<\/li>\n

  12. \n

    Agent pull requests are everywhere. Here\u2019s how to review them.<\/strong>\uff08GitHub AI\/ML\uff09<\/p>\n

    \u4e2d\u6587\u6458\u8981\uff1a<\/strong>GitHub \u53d1\u5e03\u4e86\u4e00\u4efd\u5ba1\u67e5 AI \u4ee3\u7406\u751f\u6210 Pull Request \u7684\u5b9e\u7528\u6307\u5357\u3002\u7814\u7a76\u8868\u660e\uff0c\u4ee3\u7406\u751f\u6210\u7684\u4ee3\u7801\u6bd4\u4eba\u5de5\u4ee3\u7801\u5f15\u5165\u66f4\u591a\u5197\u4f59\u548c\u6280\u672f\u503a\u52a1\uff0c\u4f46\u5ba1\u67e5\u8005\u53cd\u800c\u66f4\u5bb9\u6613\u6279\u51c6\u8fd9\u4e9b PR\u3002\u6587\u7ae0\u6307\u51fa\uff0cGitHub Copilot \u5df2\u5904\u7406\u8d85\u8fc7 6000 \u4e07\u6b21\u4ee3\u7801\u5ba1\u67e5\uff0cGitHub \u4e0a\u8d85\u8fc7\u4e94\u5206\u4e4b\u4e00\u7684\u4ee3\u7801\u5ba1\u67e5\u6d89\u53ca\u4ee3\u7406\u3002\u5ba1\u67e5\u8005\u5e94\u91cd\u70b9\u5173\u6ce8\u4e94\u5927\u5371\u9669\u4fe1\u53f7\uff1a\u4e00\u662f CI \u4f5c\u5f0a\u2014\u2014\u68c0\u67e5\u6d4b\u8bd5\u8986\u76d6\u7387\u9608\u503c\u662f\u5426\u88ab\u4fee\u6539\u3001\u6d4b\u8bd5\u662f\u5426\u88ab\u5220\u9664\u6216\u8df3\u8fc7\uff1b\u4e8c\u662f\u4ee3\u7801\u590d\u7528\u76f2\u533a\u2014\u2014\u641c\u7d22\u65b0\u5de5\u5177\u51fd\u6570\u662f\u5426\u91cd\u590d\u73b0\u6709\u529f\u80fd\uff1b\u4e09\u662f\u5e7b\u89c9\u6b63\u786e\u6027\u2014\u2014\u8ffd\u8e2a\u5173\u952e\u8def\u5f84\u7684\u8fb9\u754c\u6761\u4ef6\u3001\u6743\u9650\u68c0\u67e5\u548c\u7ade\u4e89\u6761\u4ef6\uff1b\u56db\u662f\u4ee3\u7406\u653e\u5f03\u2014\u2014\u5927\u578b PR \u82e5\u65e0\u6e05\u6670\u5b9e\u65bd\u8ba1\u5212\uff0c\u5ba1\u67e5\u524d\u5e94\u8981\u6c42\u4f5c\u8005\u62c6\u5206\uff1b\u4e94\u662f\u5de5\u4f5c\u6d41\u4e2d\u7684\u4e0d\u53ef\u4fe1\u8f93\u5165\u2014\u2014\u68c0\u67e5\u63d0\u793a\u6ce8\u5165\u98ce\u9669\u3001GITHUB_TOKEN \u6743\u9650\u662f\u5426\u6700\u5c0f\u5316\u3001\u6a21\u578b\u8f93\u51fa\u662f\u5426\u672a\u7ecf\u6821\u9a8c\u76f4\u63a5\u6267\u884c\u3002\u5efa\u8bae\u5148\u7528 Copilot \u81ea\u52a8\u5ba1\u67e5\u5904\u7406\u673a\u68b0\u6027\u95ee\u9898\uff0c\u4eba\u5de5\u4e13\u6ce8\u4e8e\u5224\u65ad\u6027\u5de5\u4f5c\u3002<\/p>\n

    English Summary:<\/strong> GitHub published a practical guide for reviewing AI agent-generated pull requests. Research shows agent-generated code introduces more redundancy and technical debt than human-written code, yet reviewers feel better about approving it. GitHub Copilot has processed over 60 million code reviews, with more than one in five reviews on GitHub now involving agents. Reviewers should watch for five red flags: CI gaming (checking if coverage thresholds changed, tests removed, or CI steps weakened), code reuse blindness (searching for duplicate utilities), hallucinated correctness (tracing critical paths for boundary conditions and permission checks), agentic ghosting (large PRs without implementation plans), and untrusted input in workflows (prompt injection risks, excessive GITHUB_TOKEN permissions, and unvalidated model output execution). The guide recommends letting Copilot handle mechanical checks first, freeing humans to focus on judgment-based review work.<\/p>\n

    \u539f\u6587\u94fe\u63a5<\/a><\/p>\n<\/li>\n

  13. \n

    Notes from inside China's AI labs<\/strong>\uff08Interconnects\uff09<\/p>\n

    \u4e2d\u6587\u6458\u8981\uff1a<\/strong>Interconnects AI \u535a\u4e3b\u5206\u4eab\u4e86\u8d70\u8bbf\u4e2d\u56fd\u4e3b\u8981 AI \u5b9e\u9a8c\u5ba4\u7684\u89c2\u5bdf\u7b14\u8bb0\u3002\u4e2d\u56fd\u7814\u7a76\u8005\u5c55\u73b0\u51fa\u6781\u5f3a\u7684\u8c26\u900a\u548c\u52a1\u5b9e\u7cbe\u795e\uff0c\u4ed6\u4eec\u66f4\u613f\u610f\u4ece\u4e8b\u975e flashy \u7684\u57fa\u7840\u5de5\u4f5c\u4ee5\u63d0\u5347\u6700\u7ec8\u6a21\u578b\u6548\u679c\uff0c\u8f83\u5c11\u53d7\u4e2a\u4eba\u804c\u4e1a\u91ce\u5fc3\u5e72\u6270\u3002\u4e0e\u6b27\u7f8e\u5b9e\u9a8c\u5ba4\u4e0d\u540c\uff0c\u4e2d\u56fd\u6838\u5fc3\u8d21\u732e\u8005\u4e2d\u6709\u5927\u91cf\u5728\u6821\u5b66\u751f\uff0c\u4ed6\u4eec\u50cf Ai2 \u4e00\u6837\u5c06\u5b66\u751f\u89c6\u4e3a\u5e73\u7b49\u4f19\u4f34\u76f4\u63a5\u878d\u5165 LLM \u56e2\u961f\uff0c\u800c\u975e\u50cf OpenAI\u3001Anthropic \u7b49\u516c\u53f8\u90a3\u6837\u4e0d\u62db\u5b9e\u4e60\u751f\u3002\u4e2d\u56fd AI \u793e\u533a\u66f4\u50cf\u4e00\u4e2a\u751f\u6001\u7cfb\u7edf\u800c\u975e\u5bf9\u7acb\u7684\u90e8\u843d\uff0c\u5404\u5b9e\u9a8c\u5ba4\u5bf9 DeepSeek \u7684\u7814\u7a76\u54c1\u5473\u548c\u5b57\u8282\u8df3\u52a8\u7684\u5e02\u573a\u5730\u4f4d\u90fd\u7ed9\u4e88\u9ad8\u5ea6\u5c0a\u91cd\u3002\u56fd\u5185 AI \u9700\u6c42\u6b63\u5728\u589e\u957f\uff0c\u5c3d\u7ba1 SaaS \u5e02\u573a\u8f83\u5c0f\uff0c\u4f46\u5f00\u53d1\u8005\u5bf9 Claude \u7b49\u5de5\u5177\u7684\u72c2\u70ed\u8868\u660e\u63a8\u7406\u9700\u6c42\u5c06\u7206\u53d1\u3002\u4e2d\u56fd\u4f01\u4e1a\u666e\u904d\u6709\u6280\u672f\u81ea\u4e3b\u60c5\u7ed3\uff0c\u7f8e\u56e2\u3001\u5c0f\u7c73\u7b49\u975e\u4f20\u7edf\u79d1\u6280\u516c\u53f8\u4e5f\u5728\u81ea\u7814\u901a\u7528\u5927\u6a21\u578b\uff0c\u4ee5\u638c\u63a7\u6838\u5fc3\u6280\u672f\u6808\u3002\u653f\u5e9c\u652f\u6301\u786e\u5b9e\u5b58\u5728\u4f46\u7ec6\u8282\u4e0d\u660e\uff0c\u6570\u636e\u4ea7\u4e1a\u76f8\u5bf9\u843d\u540e\uff0c\u5404\u5b9e\u9a8c\u5ba4\u6781\u5ea6\u6e34\u671b\u66f4\u591a Nvidia \u82af\u7247\u3002<\/p>\n

    English Summary:<\/strong> The Interconnects AI blogger shares observations from visiting leading Chinese AI labs. Chinese researchers demonstrate remarkable humility and pragmatism, willing to do unglamorous work to improve final model outcomes with less interference from individual career ambitions. Unlike Western labs, Chinese labs have many active students as core contributors, treating them as peers integrated directly into LLM teams rather than siloing them like OpenAI and Anthropic. The Chinese AI community functions more as an ecosystem than battling tribes, with mutual respect for DeepSeek's research taste and ByteDance's market dominance. Domestic AI demand is growing\u2014while the SaaS market is small, developers' obsession with tools like Claude suggests inference demand will surge. Chinese companies have a technology ownership mentality, with non-traditional tech firms like Meituan and Xiaomi building their own general-purpose LLMs to control their core stack. Government aid exists but details remain unclear, the data industry is less developed, and labs are desperate for more Nvidia chips.<\/p>\n

    \u539f\u6587\u94fe\u63a5<\/a><\/p>\n<\/li>\n

  14. \n

    Scaling Trusted Access for Cyber with GPT-5.5 and GPT-5.5-Cyber<\/strong>\uff08OpenAI News\uff09<\/p>\n

    \u4e2d\u6587\u6458\u8981\uff1a<\/strong>OpenAI \u6269\u5c55\u4e86 Trusted Access for Cyber\uff08TAC\uff09\u8ba1\u5212\uff0c\u63a8\u51fa GPT-5.5 \u548c GPT-5.5-Cyber \u6a21\u578b\u652f\u6301\u7f51\u7edc\u5b89\u5168\u9632\u5fa1\u8005\u3002TAC \u662f\u4e00\u4e2a\u57fa\u4e8e\u8eab\u4efd\u548c\u4fe1\u4efb\u5ea6\u7684\u6846\u67b6\uff0c\u901a\u8fc7\u9a8c\u8bc1\u7684\u9632\u5fa1\u8005\u53ef\u4ee5\u83b7\u5f97\u66f4\u4f4e\u7684\u5206\u7c7b\u5668\u62d2\u7edd\u7387\uff0c\u6267\u884c\u6f0f\u6d1e\u8bc6\u522b\u4e0e\u5206\u7c7b\u3001\u6076\u610f\u8f6f\u4ef6\u5206\u6790\u3001\u4e8c\u8fdb\u5236\u9006\u5411\u5de5\u7a0b\u3001\u68c0\u6d4b\u5de5\u7a0b\u4e0e\u8865\u4e01\u9a8c\u8bc1\u7b49\u9632\u5fa1\u6027\u5de5\u4f5c\uff0c\u540c\u65f6\u7ee7\u7eed\u963b\u6b62\u51ed\u8bc1\u7a83\u53d6\u3001\u6076\u610f\u8f6f\u4ef6\u90e8\u7f72\u7b49\u6076\u610f\u6d3b\u52a8\u3002GPT-5.5 with TAC \u9762\u5411\u5927\u591a\u6570\u9632\u5fa1\u573a\u666f\uff0cGPT-5.5-Cyber \u5219\u9488\u5bf9\u6388\u6743\u7ea2\u961f\u6d4b\u8bd5\u3001\u6e17\u900f\u6d4b\u8bd5\u7b49\u4e13\u4e1a\u5de5\u4f5c\u6d41\u63d0\u4f9b\u6709\u9650\u9884\u89c8\uff0c\u9700\u66f4\u5f3a\u7684\u8eab\u4efd\u9a8c\u8bc1\u548c\u8d26\u6237\u7ea7\u63a7\u5236\u3002OpenAI \u4e0e Cisco\u3001CrowdStrike\u3001Palo Alto Networks\u3001Cloudflare\u3001Snyk\u3001SentinelOne \u7b49\u5b89\u5168\u5382\u5546\u5408\u4f5c\uff0c\u6784\u5efa\u4ece\u6f0f\u6d1e\u7814\u7a76\u3001\u8f6f\u4ef6\u4f9b\u5e94\u94fe\u5b89\u5168\u5230\u68c0\u6d4b\u76d1\u63a7\u3001\u7f51\u7edc\u9632\u62a4\u7684\u5b89\u5168\u98de\u8f6e\u3002\u4e2a\u4eba\u7528\u6237\u53ef\u5728 chatgpt.com\/cyber \u7533\u8bf7\u9a8c\u8bc1\uff0c\u4f01\u4e1a\u7528\u6237\u53ef\u901a\u8fc7 OpenAI \u4ee3\u8868\u7533\u8bf7\u56e2\u961f\u8bbf\u95ee\u6743\u9650\u3002<\/p>\n

    English Summary:<\/strong> OpenAI expanded its Trusted Access for Cyber (TAC) program with GPT-5.5 and GPT-5.5-Cyber models to support cybersecurity defenders. TAC is an identity and trust-based framework where verified defenders receive lower classifier refusals for defensive workflows including vulnerability identification and triage, malware analysis, binary reverse engineering, detection engineering, and patch validation, while safeguards continue blocking credential theft and malware deployment. GPT-5.5 with TAC serves most defensive security needs, while GPT-5.5-Cyber offers limited preview access for specialized workflows like authorized red teaming and penetration testing with stronger verification and account-level controls. OpenAI is partnering with security vendors including Cisco, CrowdStrike, Palo Alto Networks, Cloudflare, Snyk, and SentinelOne to build a security flywheel spanning vulnerability research, software supply chain security, detection and monitoring, and network protection. Individual users can verify at chatgpt.com\/cyber; enterprises can request team access through their OpenAI representative.<\/p>\n

    \u539f\u6587\u94fe\u63a5<\/a><\/p>\n<\/li>\n

  15. \n

    Ollama is now powered by MLX on Apple Silicon in preview<\/strong>\uff08Ollama Blog\uff09<\/p>\n

    \u4e2d\u6587\u6458\u8981\uff1a<\/strong>Ollama \u53d1\u5e03\u57fa\u4e8e Apple MLX \u6846\u67b6\u7684\u9884\u89c8\u7248\u672c\uff0c\u4e3a Apple Silicon \u5e26\u6765\u663e\u8457\u6027\u80fd\u63d0\u5347\u3002\u65b0\u7248\u672c\u5229\u7528 MLX \u7684\u7edf\u4e00\u5185\u5b58\u67b6\u6784\uff0c\u5728 M5\u3001M5 Pro \u548c M5 Max \u82af\u7247\u4e0a\u501f\u52a9\u65b0\u7684 GPU Neural Accelerators \u52a0\u901f\u9996\u4ee4\u724c\u65f6\u95f4\uff08TTFT\uff09\u548c\u751f\u6210\u901f\u5ea6\u3002\u6d4b\u8bd5\u663e\u793a\uff0c\u4f7f\u7528 Qwen3.5-35B-A3B \u6a21\u578b\u7684 NVFP4 \u91cf\u5316\u7248\u672c\uff0c\u9884\u586b\u5145\u901f\u5ea6\u53ef\u8fbe 1851 token\/s\uff0c\u89e3\u7801\u901f\u5ea6\u8fbe 134 token\/s\u3002Ollama \u65b0\u589e\u5bf9 NVIDIA NVFP4 \u683c\u5f0f\u7684\u652f\u6301\uff0c\u5728\u4fdd\u6301\u6a21\u578b\u7cbe\u5ea6\u7684\u540c\u65f6\u964d\u4f4e\u5185\u5b58\u5e26\u5bbd\u548c\u5b58\u50a8\u9700\u6c42\uff0c\u4e0e\u751f\u4ea7\u73af\u5883\u63a8\u7406\u63d0\u4f9b\u5546\u4fdd\u6301\u4e00\u81f4\u3002\u7f13\u5b58\u7cfb\u7edf\u4e5f\u5f97\u5230\u5347\u7ea7\uff1a\u8de8\u5bf9\u8bdd\u590d\u7528\u7f13\u5b58\u964d\u4f4e\u5185\u5b58\u5360\u7528\u3001\u5728\u63d0\u793a\u8bcd\u667a\u80fd\u4f4d\u7f6e\u5b58\u50a8\u68c0\u67e5\u70b9\u51cf\u5c11\u5904\u7406\u65f6\u95f4\u3001\u5171\u4eab\u524d\u7f00\u5728\u65e7\u5206\u652f\u88ab\u5220\u9664\u540e\u4ecd\u80fd\u4fdd\u7559\u66f4\u4e45\u3002\u7528\u6237\u9700\u914d\u5907 32GB \u4ee5\u4e0a\u7edf\u4e00\u5185\u5b58\u7684 Mac\uff0c\u53ef\u901a\u8fc7 ollama launch \u547d\u4ee4\u542f\u52a8 Claude Code \u6216 OpenClaw \u7b49\u7f16\u7801\u4ee3\u7406\u3002<\/p>\n

    English Summary:<\/strong> Ollama released a preview version powered by Apple's MLX framework, delivering significant performance improvements on Apple Silicon. The new version leverages MLX's unified memory architecture and the new GPU Neural Accelerators on M5, M5 Pro, and M5 Max chips to accelerate both time-to-first-token (TTFT) and generation speed. Testing with Alibaba's Qwen3.5-35B-A3B model in NVFP4 quantization shows prefill speeds up to 1851 tokens\/s and decode speeds of 134 tokens\/s. Ollama now supports NVIDIA's NVFP4 format, maintaining model accuracy while reducing memory bandwidth and storage requirements for inference, achieving parity with production inference providers. The caching system has been upgraded with cross-conversation cache reuse for lower memory utilization, intelligent checkpointing at strategic prompt locations for faster responses, and smarter eviction that preserves shared prefixes longer. Users need a Mac with more than 32GB unified memory and can launch coding agents like Claude Code or OpenClaw via the ollama launch command.<\/p>\n

    \u539f\u6587\u94fe\u63a5<\/a><\/p>\n<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"

    \u65e5\u671f\uff1a2026-05-10 \u672c\u671f\u805a\u7126\uff1a\u91cd\u70b9\u5173\u6ce8\u6a21\u578b\u53d1\u5e03\u4e0e release notes\u3001\u5b98\u65b9 engineeri […]<\/p>\n","protected":false},"author":0,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7],"tags":[],"class_list":["post-391","post","type-post","status-publish","format-standard","hentry","category-ai-daily"],"_links":{"self":[{"href":"http:\/\/www.faiyi.com\/index.php?rest_route=\/wp\/v2\/posts\/391","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.faiyi.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.faiyi.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"http:\/\/www.faiyi.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=391"}],"version-history":[{"count":0,"href":"http:\/\/www.faiyi.com\/index.php?rest_route=\/wp\/v2\/posts\/391\/revisions"}],"wp:attachment":[{"href":"http:\/\/www.faiyi.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=391"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.faiyi.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=391"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.faiyi.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=391"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}