{"id":13221,"date":"2025-09-19T12:04:43","date_gmt":"2025-09-19T04:04:43","guid":{"rendered":"https:\/\/ihower.tw\/blog\/?p=13221"},"modified":"2025-12-20T19:19:33","modified_gmt":"2025-12-20T11:19:33","slug":"ai-evals-debate","status":"publish","type":"post","link":"https:\/\/ihower.tw\/blog\/13221-ai-evals-debate","title":{"rendered":"AI Evals \u5927\u8faf\u8ad6: \u5f9e Claude Code \u8a2a\u8ac7\u5f15\u767c\u7684\u53cd\u601d"},"content":{"rendered":"\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"800\" height=\"533\" data-attachment-id=\"13249\" data-permalink=\"https:\/\/ihower.tw\/blog\/13221-ai-evals-debate\/ai-evals-debate-2\" data-orig-file=\"https:\/\/ihower.tw\/blog\/wp-content\/uploads\/2025\/09\/ai-evals-debate.png\" data-orig-size=\"800,533\" data-comments-opened=\"1\" data-image-meta=\"{&quot;aperture&quot;:&quot;0&quot;,&quot;credit&quot;:&quot;&quot;,&quot;camera&quot;:&quot;&quot;,&quot;caption&quot;:&quot;&quot;,&quot;created_timestamp&quot;:&quot;0&quot;,&quot;copyright&quot;:&quot;&quot;,&quot;focal_length&quot;:&quot;0&quot;,&quot;iso&quot;:&quot;0&quot;,&quot;shutter_speed&quot;:&quot;0&quot;,&quot;title&quot;:&quot;&quot;,&quot;orientation&quot;:&quot;0&quot;}\" data-image-title=\"ai-evals-debate\" data-image-description=\"\" data-image-caption=\"\" data-medium-file=\"https:\/\/ihower.tw\/blog\/wp-content\/uploads\/2025\/09\/ai-evals-debate-300x200.png\" data-large-file=\"https:\/\/ihower.tw\/blog\/wp-content\/uploads\/2025\/09\/ai-evals-debate.png\" src=\"https:\/\/ihower.tw\/blog\/wp-content\/uploads\/2025\/09\/ai-evals-debate.png\" alt=\"\" class=\"wp-image-13249\" style=\"width:699px;height:auto\" srcset=\"https:\/\/ihower.tw\/blog\/wp-content\/uploads\/2025\/09\/ai-evals-debate.png 800w, https:\/\/ihower.tw\/blog\/wp-content\/uploads\/2025\/09\/ai-evals-debate-300x200.png 300w, https:\/\/ihower.tw\/blog\/wp-content\/uploads\/2025\/09\/ai-evals-debate-768x512.png 768w\" sizes=\"auto, (max-width: 800px) 100vw, 800px\" \/><\/figure>\n\n\n\n<p>\u6700\u8fd1\u770b\u5230\u4e00\u5834\u95dc\u65bc AI Evals \u7684\u7cbe\u5f69\u8ad6\u6230\uff0c\u722d\u8ad6\u7126\u9ede\u4e0d\u5728\u6a21\u578b\u8a13\u7df4\u5c64\u9762\u7684\u8a55\u4f30(\u9019\u500b\u5927\u5bb6\u90fd\u6709\u5171\u8b58\u8981\u505a)\uff0c\u800c\u662f AI \u61c9\u7528\u5c64\u9762\u5230\u5e95\u8981\u505a\u591a\u5c11\u8a55\u4f30\u3002\u9019\u8b93\u6211\u60f3\u8d77\u81ea\u5df1\u5728\u8edf\u9ad4\u958b\u767c\u7684\u7d93\u9a57: \u5982\u4f55\u5beb\u6e2c\u8a66\u4e5f\u662f\u6211\u66fe\u95dc\u6ce8\u7684\u554f\u984c\uff0c\u4f46\u8aaa\u5be6\u8a71\uff0c\u6211\u5f9e\u4f86\u4e0d\u8ffd\u6c42 100% test coverage\u3002\u5373\u4f7f Ruby \u793e\u7fa4\u4e5f\u5f37\u8abf\u6bcf\u4ef6\u4e8b\u90fd\u8981\u6709\u6e2c\u8a66\u6db5\u84cb\uff0c\u4f46\u6211\u9084\u662f\u6bd4\u8f03\u8003\u91cf\u6210\u672c\u6548\u76ca\uff0c\u81ea\u52d5\u5316\u6e2c\u8a66\u5c0d\u6211\u4f86\u8aaa\u662f\u503c\u5f97\u505a\u624d\u6703\u505a\u7684\u4e8b\u3002<\/p>\n\n\n\n<p>\u73fe\u5728 AI Evals \u4e5f\u8655\u65bc\u985e\u4f3c\u968e\u6bb5\u3002\u6211\u53bb\u5e74\u5c31\u958b\u59cb\u95dc\u6ce8\u4e26\u5206\u4eab\u5982\u4f55\u505a\u8a55\u4f30\uff0c\u4f46\u8981\u6c42\u6bcf\u500b\u9762\u5411\u90fd 100% \u6709\u8a55\u4f30\u5176\u5be6\u662f\u4e0d\u5be6\u969b\u7684\u3002AI \u662f\u6a5f\u7387\u6027\u8edf\u9ad4\uff0c\u8a55\u4f30\u96e3\u5ea6\u66f4\u9ad8\uff0cAI \u7684\u8f38\u51fa\u597d\u4e0d\u597d\u4e5f\u975e\u5e38\u6709\u4e3b\u89c0\u6210\u5206\uff0c\u76ee\u524d\u600e\u9ebc\u505a\u5f88\u4f9d\u8cf4\u5be6\u52d9\u7d93\u9a57\u4ea4\u6d41\u3002\u6700\u8fd1\u525b\u4e0a\u5b8c <a href=\"https:\/\/maven.com\/parlance-labs\/evals\">AI Evals For Engineers &amp; PMs<\/a> \u8ab2\u7a0b\uff0c\u6709\u4e86\u65b0\u7684\u9ad4\u6703\u3002\u9996\u5148\uff0c\u300c\u8a55\u4f30\u9a45\u52d5\u958b\u767c\u300d(\u6307\u5148\u5beb\u8a55\u4f30\u518d\u958b\u767c) \u7adf\u7136\u53ef\u80fd\u662f\u932f\u7684\u65b9\u5411 &#8211; \u5c0d\u65bc\u6c92\u6709\u6a19\u6e96\u8f38\u51fa\u7684 AI \u4efb\u52d9\uff0c\u4f60\u7121\u6cd5\u7121\u9650\u6295\u8cc7\u5728\u8a55\u4f30\u4e0a\u3002<\/p>\n\n\n\n<!--more-->\n\n\n\n<p>\u6211\u76ee\u524d\u8a8d\u70ba\u8a55\u4f30\u53ef\u5206\u5169\u7a2e: <\/p>\n\n\n\n<p><strong>Soft preferences (graded helpfulness)<\/strong>: \u5927\u591a\u6578\u7bc4\u4f8b\u7528 LLM as Judge \u7d66 1-5 \u5206\u8a55\u4f30\u300c\u6709\u591a\u5c11\u5e6b\u52a9\u300d\u3001\u300c\u6709\u591a\u6e05\u6670\u300d\u7b49\uff0c\u90fd\u5c6c\u65bc\u9019\u985e\u3002\u5766\u767d\u8aaa\uff0c\u9019\u7a2e\u8a55\u4f30\u5e6b\u52a9\u6709\u9650\uff0c\u9802\u591a\u505a\u76f8\u5c0d\u6bd4\u8f03\uff0c\u5be6\u969b\u6548\u7528\u4e0d\u5927\u3002<\/p>\n\n\n\n<p><strong>Hard rules (binary guardrails)<\/strong>: \u9019\u7a2e\u5176\u5be6\u66f4\u5be6\u7528\uff0c\u5c31\u662f\u6703\u6709\u6a19\u6e96\u7b54\u6848\u7684\u4e8c\u5143\u5224\u65b7\uff0c\u7528 Code-based \u6216 LLM as Judge \u5224\u65b7\u662f\u4e0d\u662f\u6709\u505a\u5230\u4f60\u8981\u6c42\u7684\u9650\u5236\u3002\u9019\u9580\u8ab2\u7a0b\u91cd\u9ede\u5c31\u662f\u6559\u5982\u4f55\u7528 LLM \u79d1\u5b78\u5730\u505a\u51fa\u5c0d\u9f4a\u5f8c\u7684 binary classifier\uff0c\u8001\u5e2b\u5c31\u662f\u9019\u7bc7\u77e5\u540d paper: <a href=\"https:\/\/arxiv.org\/abs\/2404.12272\">Who Validates the Validators? Aligning LLM-Assisted Evaluation of LLM Outputs with Human Preferences<\/a> \u7684\u7b2c\u4e00\u4f5c\u8005\u3002<\/p>\n\n\n\n<p>\u6c92\u6709\u6a19\u6e96\u7b54\u6848\u7684 LLM \u8981\u8f38\u51fa\u600e\u9ebc\u505a binary classification \u7684\u8a55\u4f30? \u6211\u5011\u4e0d\u505a\u901a\u7528\u7684 LLM-as-Judge \u8a55\u4f30\uff0c\u800c\u662f\u5148\u4eba\u5de5\u505a\u932f\u8aa4\u5206\u6790\uff0c\u627e\u51fa\u771f\u6b63\u91cd\u8981\u7684\u932f\u8aa4\u6a21\u5f0f\uff0c\u7136\u5f8c\u6bcf\u500b\u932f\u8aa4\u6a21\u5f0f\u5206\u5225\u505a LLM as Judge \u53ea\u5224\u65b7 Y\/N \u8f38\u51fa\u6709\u6c92\u6709\u72af\u9019\u500b\u7279\u5b9a\u932f\u8aa4\u3002\u9019\u500b Judge \u9700\u8981\u8a8d\u771f\u505a\u5c0d\u9f4a\uff0c\u5f97\u6709\u5f88\u9ad8\u7684\u6e96\u78ba\u7387\uff0c\u56e0\u70ba\u662f classification \u4efb\u52d9\u6240\u4ee5\u4e5f\u76f8\u5c0d\u597d\u505a prompt \u6700\u4f73\u5316\u3002\u5982\u6b64\u4e0b\u4f86\uff0c\u4f60\u6703\u5f97\u5230\u771f\u6b63\u6709\u7528\u7684\u6e96\u78ba\u6307\u6a19\uff0c\u80fd\u5be6\u969b\u7528\u4f86\u5e6b\u52a9\u6539\u9032 AI \u61c9\u7528\u3002<\/p>\n\n\n\n<p>\u5c31\u5982 Hamel Husain [0] \u8aaa\u7684\uff1a\u300cGenerally no. Eval-driven development (writing evaluators before implementing features) sounds appealing but creates more problems than it solves.\u300d\u91cd\u9ede\u662f\u627e\u5230\u771f\u6b63\u5f71\u97ff\u7528\u6236\u9ad4\u9a57\u7684\u95dc\u9375\u6307\u6a19\uff0c\u800c\u4e0d\u662f\u70ba\u4e86\u8a55\u4f30\u800c\u8a55\u4f30\u3002<\/p>\n\n\n\n<p>[0] <a href=\"https:\/\/x.com\/HamelHusain\/status\/1950247467444031828\">x.com\/HamelHusain\/status\/1950247467444031828<\/a><\/p>\n\n\n\n<p>\u90a3 Soft preferences \u600e\u9ebc\u8fa6\uff1f\u9019\u90e8\u5206\u9084\u662f\u5f97\u4f9d\u8cf4\u4eba\u7684\u5224\u65b7\uff0cAI Judge \u7121\u6cd5\u5b8c\u5168\u53d6\u4ee3\u3002\u7121\u8ad6\u662f\u5718\u968a\u81ea\u5df1 dogfooding\u3001\u76e3\u63a7\u7dda\u4e0a\u6307\u6a19\u3001\u6536\u96c6\u7528\u6236\u56de\u994b\uff0c\u6216\u662f\u8dd1 A\/B testing\uff0c\u9084\u662f\u5f97\u9760\u50b3\u7d71\u7522\u54c1\u958b\u767c\u7684\u65b9\u6cd5\u3002AI \u61c9\u7528\u7684\u8a55\u4f30\u4e0d\u662f\u975e\u9ed1\u5373\u767d\u7684\u9078\u64c7\u984c\uff0c\u5f97\u5728\u4e0d\u540c\u968e\u6bb5\u4e0d\u540c\u9700\u6c42\uff0c\u9748\u6d3b\u904b\u7528\u4e0d\u540c\u5de5\u5177\u7684\u667a\u6167\u3002<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>Hamel Husain \u9084\u6709\u4e00\u7bc7 <a href=\"https:\/\/decodingml.substack.com\/p\/the-mirage-of-generic-ai-metrics\">The Mirage of Generic AI Metrics<\/a> \u4e5f\u6307\u51fa\u901a\u7528\u6307\u6a19\u7684\u554f\u984c\uff0c\u4e26\u4ecb\u7d39\u932f\u8aa4\u5206\u6790\u3001\u5efa\u7acb\u5ba2\u88fd\u5316\u6307\u6a19\u3001\u5efa\u7acb\u53ef\u4fe1\u7684 LLM Judge\u3002<\/p>\n<\/blockquote>\n\n\n\n<p>(\u4ee5\u4e0b\u6574\u7406\u6574\u500b\u8ad6\u6230\u7684\u7cbe\u5f69\u5167\u5bb9\uff0c\u900f\u904e Claude \u6458\u8981\u5beb\u6210)<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">AI Evals \u8ad6\u6230\u6574\u7406<\/h2>\n\n\n\n<p>\u9019\u5834\u8faf\u8ad6\u59cb\u65bc Anthropic Claude Code \u5275\u59cb\u4eba Boris \u5728\u8a2a\u8ac7\u4e2d[1]\u7684\u5766\u767d\u3002\u7576\u88ab\u554f\u5230\u5982\u4f55\u8a55\u4f30\u65b0\u6a21\u578b\u6642\uff0c\u4ed6\u7684\u56de\u7b54\u4ee4\u4eba\u610f\u5916\u5730\u7c21\u55ae\uff1a\u300c\u6211\u5c31\u505a\u6211\u7576\u5929\u7684\u5de5\u4f5c\u300d\u3002\u6c92\u6709\u8907\u96dc\u7684\u6e2c\u8a66\u5957\u4ef6\uff0c\u6c92\u6709\u7cbe\u5bc6\u7684\u6307\u6a19\uff0c\u7d14\u7cb9\u5c31\u662f\u611f\u89ba\u6a21\u578b\u8b8a\u8070\u660e\u4e86\u55ce\u3002Boris \u627f\u8a8d\u4ed6\u5011\u8a66\u904e\u5efa\u7acb\u7522\u54c1\u8a55\u4f30\u7cfb\u7d71\uff0c\u4f46\u767c\u73fe\u5efa\u7acb\u8a55\u4f30\u771f\u7684\u5f88\u56f0\u96e3\uff0c\u6700\u7d42\u6700\u5927\u7684\u4fe1\u865f\u5c31\u662f\u300c\u611f\u89ba\u300d\u3002\u4ed6\u8a8d\u70ba\u5f88\u96e3\u627e\u5230\u80fd\u6355\u6349\u8edf\u9ad4\u5de5\u7a0b\u6240\u6709\u8907\u96dc\u6027\u7684\u5408\u6210\u8a55\u4f30\uff0cSWE-bench \u9019\u985e\u57fa\u6e96\u6e2c\u8a66\u96d6\u7136\u8868\u73fe\u51fa\u8272\uff0c\u4f46\u771f\u5be6\u7684\u8edf\u9ad4\u958b\u767c\u9060\u6bd4\u9019\u4e9b\u6e2c\u8a66\u8907\u96dc\u3002<\/p>\n\n\n\n<p>\u9019\u500b\u5766\u767d\u88ab swyx \u7206\u6599\u5f8c\u5f15\u767c\u8ed2\u7136\u5927\u6ce2\u3002\u4ed6\u5728\u63a8\u7279\u4e0a\u8af7\u523a\u5730\u5217\u51fa\u4e86\u7522\u696d\u73fe\u6cc1\uff1aClaude Code \u6c92\u6709 evals\uff0c\u67d0\u77e5\u540d code agent \u516c\u53f8\u6c92\u6709 evals\uff0c\u53e6\u4e00\u5bb6\u77e5\u540d\u516c\u53f8\u53ea\u6709\u6577\u884d\u7684 evals\uff0c\u9818\u5148\u7684 vibe coding \u516c\u53f8\u4e5f\u6c92\u6709 evals\u3002\u4f46\u8ce3 evals \u7684\u516c\u53f8 CEO \u6703\u8aaa\u300c\u6211\u6240\u6709\u7684\u9802\u7d1a\u5ba2\u6236\u90fd\u505a evals\uff0c\u4f60\u4e5f\u61c9\u8a72\u505a\u300d\uff0c\u611b\u4e0a evals \u516c\u53f8\u7684\u5275\u6295\u4e5f\u6703\u9644\u548c\u300c\u6211\u6240\u6709\u7684\u9802\u7d1a\u5275\u8fa6\u4eba\u90fd\u505a evals\uff0c\u5fc5\u9808\u505a evals\u300d\u3002swyx \u88dc\u5145\u8aaa\u4ed6\u78ba\u5be6\u8a8d\u70ba evals \u5f88\u91cd\u8981\uff0c\u4f46\u90a3\u4e9b\u72c2\u71b1\u652f\u6301 evals \u7684 AI \u5de5\u7a0b\u5e2b\u4e5f\u6ce8\u610f\u5230\uff0c\u505a\u4e0d\u505a evals \u4e26\u4e0d\u662f\u6210\u529f\u7684\u56b4\u683c\u8981\u6c42\uff0c\u81f3\u5c11\u5728\u5f9e 0 \u5230 1 \u7684\u968e\u6bb5\uff0c\u4e0d\u505a evals \u751a\u81f3\u53ef\u80fd\u8207\u6210\u529f\u5448\u6b63\u76f8\u95dc\u3002<\/p>\n\n\n\n<p>[1] <a href=\"https:\/\/www.youtube.com\/watch?v=iF9iV4xponk\">www.youtube.com\/watch?v=iF9iV4xponk<\/a><br>[2] <a href=\"https:\/\/x.com\/swyx\/status\/1963725773355057249\">x.com\/swyx\/status\/1963725773355057249<\/a><\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u652f\u6301 Evals \u7684\u8072\u97f3<\/h3>\n\n\n\n<p>\u5b78\u8853\u6d3e\u4ee3\u8868 Shreya Shankar [3] \u8207 Hamel Husain \u806f\u624b\u70ba evals \u8faf\u8b77\uff0c\u4ed6\u5011\u5c07 evals \u5b9a\u7fa9\u70ba\u300c\u5c0d\u61c9\u7528\u54c1\u8cea\u7684\u7cfb\u7d71\u6027\u6e2c\u91cf\u300d\uff0c\u5f37\u8abf\u9019\u4e0d\u610f\u5473\u8457\u7279\u5b9a\u7684\u6307\u6a19\u6216\u65b9\u6cd5\uff0c\u4e5f\u4e0d\u5fc5\u662f\u55ae\u4e00\u6578\u5b57\u6216\u5b8c\u5168\u6e96\u78ba\u3002\u5979\u6307\u51fa\u90a3\u4e9b\u8072\u7a31\u300c\u4e0d\u505a evals\u300d\u7684\u5718\u968a\u5176\u5be6\u5728\u81ea\u6b3a\u6b3a\u4eba\uff0c\u6bcf\u500b\u6210\u529f\u7684\u7522\u54c1\u90fd\u5728\u67d0\u8655\u505a\u8457\u8a55\u4f30\uff0c\u67e5\u770b\u8f38\u51fa\u3001\u6ce8\u610f\u554f\u984c\u3001dogfooding \u7522\u54c1\u4e26\u505a\u51fa\u6539\u8b8a\uff0c\u9019\u5c31\u662f\u8a55\u4f30\uff0c\u53ea\u662f\u6c92\u6709\u8cbc\u4e0a\u300cevals\u300d\u7684\u6a19\u7c64\u3002<\/p>\n\n\n\n<p>\u5979\u7279\u5225\u9ede\u51fa\u4e00\u500b\u95dc\u9375\u4e8b\u5be6\uff1a\u5373\u4f7f\u50cf coding agent \u9019\u6a23\u7684\u5718\u968a\u611f\u89ba\u53ef\u4ee5\u4e0d\u505a\u592a\u591a\u8a55\u4f30\uff0c\u90a3\u662f\u56e0\u70ba\u4ed6\u5011\u5df2\u7d93\u5f9e\u4e0a\u6e38\u7684\u56b4\u683c\u8a55\u4f30\u4e2d\u53d7\u76ca\u3002\u7de8\u78bc evals \u5728\u5f8c\u8a13\u7df4\u4e2d\u4f54\u64da\u5982\u6b64\u91cd\u8981\u7684\u5730\u4f4d\uff0c\u5be6\u969b\u4e0a\u662f\u5176\u4ed6\u4eba\u5df2\u7d93\u70ba\u4f60\u5b8c\u6210\u4e86\u5927\u90e8\u5206\u7684 evals \u5de5\u4f5c\u3002\u57fa\u790e\u6a21\u578b\u4f9b\u61c9\u5546\u70ba\u6bcf\u500b\u65b0\u80fd\u529b\u9818\u57df\u6295\u5165\u5927\u91cf\u8cc7\u91d1\u9032\u884c\u8a55\u4f30\uff0c\u56e0\u70ba\u4ed6\u5011\u77e5\u9053\u6c92\u6709\u7cfb\u7d71\u6027\u6e2c\u91cf\u5c31\u7121\u6cd5\u6539\u9032\u6548\u80fd\u3002<\/p>\n\n\n\n<p>\u5979\u7279\u5225\u64d4\u5fc3\u53cd eval \u60c5\u7dd2\u5c0d\u793e\u7fa4\u7684\u50b7\u5bb3\uff0c\u56e0\u70ba\u8a31\u591a\u65b0\u624b\u6b63\u5728\u5c0b\u6c42\u5efa\u7acb AI \u7522\u54c1\uff0c\u4ed6\u5011\u7684\u4efb\u52d9\u53ef\u80fd\u4e0d\u5728\u5f8c\u8a13\u7df4\u4e2d\u88ab\u5145\u5206\u4ee3\u8868\uff0c\u5718\u968a\u4e5f\u9084\u6c92\u6709\u8db3\u5920\u7d93\u9a57\u4f9d\u8cf4\u76f4\u89ba\uff0c\u5c0d\u9019\u4e9b\u5718\u968a\u4f86\u8aaa\uff0c\u5426\u5b9a evals \u7b49\u65bc\u79fb\u9664\u4e86\u5e6b\u52a9\u4ed6\u5011\u7406\u89e3\u4ec0\u9ebc\u6709\u6548\u3001\u4ec0\u9ebc\u7121\u6548\u7684\u5de5\u5177\u3002<\/p>\n\n\n\n<p>[3] <a href=\"https:\/\/www.sh-reya.com\/blog\/in-defense-ai-evals\/\">www.sh-reya.com\/blog\/in-defense-ai-evals\/<\/a><\/p>\n\n\n\n<p>Braintrust CEO Ankur Goyal [4] \u767c\u8868\u4e86\u4e00\u7bc7\u5f15\u6230\u6587\u7ae0\uff0c\u5ba3\u7a31 A\/B \u6e2c\u8a66\u5df2\u7d93\u8ddf\u4e0d\u4e0a AI \u6642\u4ee3\u7684\u6b65\u4f10\u3002\u4ed6\u8a8d\u70ba\u50b3\u7d71 A\/B \u6e2c\u8a66\u5efa\u7acb\u5728\u300c\u5275\u5efa\u8b8a\u9ad4\u5f88\u6602\u8cb4\u300d\u7684\u5047\u8a2d\u4e0a\uff0c\u4f60\u9700\u8981\u70ba\u6bcf\u500b\u9ad4\u9a57\u6295\u5165\u5927\u91cf\u8a2d\u8a08\u548c\u5de5\u7a0b\u5de5\u4f5c\uff0c\u6240\u4ee5\u53ea\u80fd\u6e2c\u8a66\u5c11\u6578\u9078\u9805\u3002\u4f46 AI \u6539\u8b8a\u4e86\u9019\u500b\u904a\u6232\u898f\u5247\uff0c\u7576 AI \u80fd\u81ea\u52d5\u70ba\u6bcf\u500b\u7528\u6236\u751f\u6210\u500b\u6027\u5316\u4ecb\u9762\u6642\uff0c\u70ba\u4ec0\u9ebc\u9084\u8981\u70ba\u300c\u6700\u4f73\u5e73\u5747\u9ad4\u9a57\u300d\u505a\u512a\u5316\uff1f\u4ed6\u6307\u51fa OpenAI \u6536\u8cfc Statsig\u3001Datadog \u6536\u8cfc Eppo \u6697\u793a\u8457\u8f49\u6298\u9ede\u5df2\u7d93\u5230\u4f86\uff0c\u672a\u4f86\u5c6c\u65bc evals \u800c\u975e A\/B \u6e2c\u8a66\u3002\u5728\u4ed6\u7684\u9858\u666f\u4e2d\uff0c\u7cfb\u7d71\u80fd\u5373\u6642\u5b78\u7fd2\u548c\u6539\u9032\uff0c\u5718\u968a\u5f9e\u624b\u5de5\u8abf\u6574\u6bcf\u500b\u7d30\u7bc0\u7684\u5de5\u5320\uff0c\u8b8a\u6210\u81ea\u52d5\u5316\u6539\u9032\u7cfb\u7d71\u7684\u67b6\u69cb\u5e2b\u3002<\/p>\n\n\n\n<p>[4] <a href=\"https:\/\/www.braintrust.dev\/blog\/ab-testing-evals\"> www.braintrust.dev\/blog\/ab-testing-evals<\/a><\/p>\n\n\n\n<p>\u5be6\u52d9\u7d93\u9a57\u8c50\u5bcc\u7684 Eugene Yan [5][6][7]\u628a evals \u985e\u6bd4\u70ba\u6e2c\u8a66\u9a45\u52d5\u958b\u767c\uff08TDD\uff09\uff0c\u5728\u958b\u767c\u529f\u80fd\u524d\u5148\u5b9a\u7fa9\u6210\u529f\u6a19\u6e96\u4e26\u5f9e\u7b2c\u4e00\u5929\u5c31\u958b\u59cb\u6e2c\u91cf\u3002\u4ed6\u5206\u4eab\u4e86\u81ea\u5df1\u7684\u7d93\u9a57\uff1a\u4e00\u65e6\u8a2d\u7f6e\u597d eval \u548c\u5be6\u9a57\u6846\u67b6\uff0c\u8b93\u8abf\u6574\u914d\u7f6e\u548c prompt \u8b8a\u6210\u4e00\u9375\u904b\u884c\u52a0\u8a55\u4f30\uff0c\u5718\u968a\u6703\u4eab\u53d7\u904b\u884c\u5be6\u9a57\u548c\u63d0\u5347\u6307\u6a19\u7684\u904e\u7a0b\uff0c\u9032\u5c55\u6703\u5f88\u5feb\u3002\u4f46\u4ed6\u4e5f\u627f\u8a8d\u8a2d\u7f6e\u9019\u500b\u6846\u67b6\u5c0d\u6bcf\u500b\u65b0\u5c08\u6848\u90fd\u662f\u6311\u6230\uff0c\u9700\u8981\u8655\u7406\u6a21\u7cca\u7684\u5de5\u4f5c\uff0c\u5373\u4f7f\u8a2d\u7f6e\u597d\u4e86\uff0c\u4e5f\u5f88\u5c11\u4eba\u60f3\u770b\u751f\u6210\u7684\u56de\u61c9\u4f86\u505a\u5b9a\u6027\u8a55\u4f30\u3002\u4ed6\u5f37\u8abf\u901a\u7528\u7684 evals \u50cf\u300cfaithfulness\u300d\u6216\u300chelpfulness\u300d\u662f\u6c92\u7528\u7684\uff0c\u4f60\u7684 evals \u5fc5\u9808\u8207\u7528\u6236\u554f\u984c\u5c0d\u9f4a\u3002<\/p>\n\n\n\n<p>[5] <a href=\"https:\/\/x.com\/eugeneyan\/status\/1960148508495020234\">x.com\/eugeneyan\/status\/1960148508495020234<\/a><br>[6] <a href=\"https:\/\/x.com\/eugeneyan\/status\/1964103249230774682\">x.com\/eugeneyan\/status\/1964103249230774682<\/a><br>[7] <a href=\"https:\/\/x.com\/eugeneyan\/status\/1964334356006391882\">x.com\/eugeneyan\/status\/1964334356006391882<\/a><\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u53cd\u5c0d Evals \u7684\u6279\u5224<\/h3>\n\n\n\n<p>AgentOps CEO Alex Reibman \u767c\u8868\u4e86\u6700\u6fc0\u9032\u7684\u6279\u8a55\u6587\u7ae0\u300aEvals are a scam\u300b [8]\uff0c\u76f4\u6307\u6574\u500b\u7522\u696d\u90fd\u5728\u88ab\u300ceval \u6d17\u8166\u300d\u3002\u4ed6\u8a8d\u70ba\u5927\u591a\u6578 AI \u5718\u968a\u4e0d\u9700\u8981 evals\uff0c\u9700\u8981\u7684\u662f logging\u3001QA \u548c\u54c1\u5473\u3002\u4ed6\u6307\u51fa\u57fa\u790e\u6a21\u578b\u7684 evals\uff08\u6e2c\u8a66 LLM \u5728\u5404\u7a2e\u4efb\u52d9\u7684\u4e00\u822c\u6548\u80fd\uff09\u548c\u7522\u54c1 evals\uff08\u6e2c\u8a66\u61c9\u7528\u5728\u771f\u5be6\u4f7f\u7528\u6848\u4f8b\u7684\u6548\u679c\uff09\u662f\u5b8c\u5168\u4e0d\u540c\u7684\u6771\u897f\uff0c\u9664\u975e\u4f60\u5728 OpenAI\u3001Anthropic \u6216 Meta \u5de5\u4f5c\uff0c\u5426\u5247\u4f60\u4e0d\u9700\u8981\u6a21\u578b evals\u3002\u800c\u7522\u54c1 evals \u662f\u4e3b\u89c0\u3001\u6a21\u7cca\u4e14\u6df7\u4e82\u7684\uff0ceval \u4f9b\u61c9\u5546\u8a66\u5716\u8ce3\u7d66\u4f60\u4ed6\u5011\u6c38\u9060\u7121\u6cd5\u63d0\u4f9b\u7684\u6771\u897f\uff1a\u5c0d\u4f60\u81ea\u5df1\u7522\u54c1\u7684\u5c08\u696d\u77e5\u8b58\u3002\u4ed6\u8a8d\u70ba\u9019\u4e9b\u516c\u53f8\u5176\u5be6\u662f\u300c\u8907\u96dc\u6027\u5546\u4eba\u300d\uff0c\u900f\u904e\u8ab2\u7a0b\u3001\u96fb\u5b50\u66f8\u3001\u90e8\u843d\u683c\u3001\u7db2\u8def\u7814\u8a0e\u6703\u63a8\u92b7\u4ed6\u5011\u7684\u6846\u67b6\uff0c\u6700\u7d42\u771f\u6b63\u8ce3\u7684\u662f\u6602\u8cb4\u7684 logging \u670d\u52d9\u3002<\/p>\n\n\n\n<p>[8] <a href=\"https:\/\/hacktrace.substack.com\/p\/evals-are-a-scam-and-were-being-gaslit\">hacktrace.substack.com\/p\/evals-are-a-scam-and-were-being-gaslit<\/a><\/p>\n\n\n\n<p>\u524d GitHub Copilot \u5718\u968a\u3001\u73fe Quotient \u5275\u8fa6\u4eba Julia Neagu [9] \u5f9e\u5de5\u7a0b\u5e2b\u8996\u89d2\u63d0\u4f9b\u4e86\u6df1\u5165\u5206\u6790\u3002\u5979\u6307\u51fa LLM \u4f5c\u70ba API \u51fa\u73fe\uff0c\u9019\u5c0d\u5de5\u7a0b\u5e2b\u4f86\u8aaa\u662f\u719f\u6089\u7684\u9818\u57df\uff0c\u6240\u6709\u6210\u529f\u7684 AI \u5de5\u5177\u90fd\u6709\u4e00\u500b\u5171\u540c\u9ede\uff1a\u5c0d\u5de5\u7a0b\u5e2b\u6709\u826f\u597d\u7684\u4eba\u9ad4\u5de5\u5b78\u8a2d\u8a08\u3002\u4f46 eval \u5de5\u5177\u6c92\u6709\u6620\u5c04\u5230\u5df2\u77e5\u7684\u958b\u767c\u8005\u5de5\u4f5c\u6d41\u7a0b\uff0c\u9019\u5c31\u662f\u70ba\u4ec0\u9ebc\u5169\u5e74\u4f86\u300cevals \u5f88\u91cd\u8981\u300d\u7684\u8ad6\u8ff0\u4e0b\uff0ceval \u5de5\u5177\u4ecd\u7136\u6c92\u6709\u7a81\u7834\u7684\u539f\u56e0\u3002\u5979\u5728 GitHub Copilot \u7684\u7d93\u9a57\u8b49\u5be6\u4e86\u9019\u9ede\uff1a\u96d6\u7136\u6709\u8907\u96dc\u7684 eval \u6846\u67b6\u7528\u65bc\u57fa\u6e96\u6e2c\u8a66\uff0c\u4f46\u771f\u6b63\u6c7a\u5b9a\u662f\u5426\u4e0a\u7dda\u7684\u9084\u662f A\/B \u6e2c\u8a66\uff0c\u800c\u4e14\u63a8\u51fa\u7684\u58d3\u529b\u5982\u6b64\u4e4b\u9ad8\uff0c\u5982\u679c\u901a\u904e A\/B \u6e2c\u8a66\u5c31\u6703\u76f4\u63a5\u4e0a\u7dda\u3002\u5979\u7e3d\u7d50\u8aaa\uff0c\u5118\u7ba1\u7db2\u4e0a\u8ad6\u8ff0\u8072\u7a31\u76f8\u53cd\uff0c\u5927\u90e8\u5206 AI \u7522\u54c1\u4ecd\u7136\u662f\u57fa\u65bc vibes \u5728\u51fa\u8ca8\u3002<\/p>\n\n\n\n<p>[9] <a href=\"https:\/\/x.com\/JuliaANeagu\/status\/1964704824299253888\">x.com\/JuliaANeagu\/status\/1964704824299253888 <\/a><\/p>\n\n\n\n<p>Raindrop CTO Ben Hylak [10] \u5f9e\u76e3\u63a7\u89d2\u5ea6\u6279\u8a55 evals\uff0c\u8a8d\u70ba evals \u53ea\u662f\u5df2\u77e5\u554f\u984c\u7684\u96c6\u5408\uff0c\u662f\u4f60\u900f\u904e\u5c0d\u6297\u6027\u9078\u64c7\u627e\u5230\u7684\u5931\u6557\u6848\u4f8b\uff0c\u7121\u6cd5\u767c\u73fe\u672a\u77e5\u554f\u984c\u3002\u4ed6\u5f15\u7528 Replit \u5de5\u7a0b\u5e2b\u7684\u8a71\uff1a\u7531\u65bc\u8f38\u5165\u7a7a\u9593\u4e0d\u518d\u6709\u9650\uff0c\u4f60\u7121\u6cd5\u7c21\u55ae\u5730\u5beb\u5e7e\u500b\u9078\u5b9a\u7684\u6e2c\u8a66\u7136\u5f8c\u9075\u5faa\u6e2c\u8a66\u9a45\u52d5\u958b\u767c\uff0c\u4f60\u6703\u5728\u4e0d\u77e5\u60c5\u7684\u60c5\u6cc1\u4e0b\u7834\u58de\u95dc\u9375\u529f\u80fd\u3002\u9019\u5c31\u662f\u70ba\u4ec0\u9ebc\u5728\u751f\u7522\u74b0\u5883\u4e2d\u6e2c\u8a66\uff0c\u4f7f\u7528\u50b3\u7d71 A\/B \u6e2c\u8a66\u4e5f\u5f88\u95dc\u9375\uff0c\u80fd\u76e1\u53ef\u80fd\u63a5\u8fd1\u4f60\u670d\u52d9\u7684\u4e00\u822c\u4eba\u7fa4\uff0c\u6709\u66f4\u9ad8\u6a5f\u6703\u6e2c\u8a66\u9577\u5c3e\u7d50\u679c\u3002\u7576 AI agents \u8b8a\u5f97\u8d8a\u4f86\u8d8a\u5f37\u5927\uff0c\u80fd\u57f7\u884c\u7684\u4efb\u52d9\u8d8a\u4f86\u8d8a\u958b\u653e\uff0c\u904b\u884c\u6642\u9593\u8d8a\u4f86\u8d8a\u9577\uff0c\u5982\u679c\u69cb\u5efa\u6b63\u78ba\uff0c\u5b83\u5011\u80fd\u4ee5\u4f60\u7121\u6cd5\u9810\u6e2c\u6216\u60f3\u50cf\u7684\u65b9\u5f0f\u57f7\u884c\u4efb\u52d9\uff0c\u6e2c\u8a66\u6240\u6709\u53ef\u80fd\u6027\u8b8a\u5f97\u4e0d\u53ef\u80fd\uff0c\u4f46\u76e3\u63a7\u4ecd\u7136\u53ef\u884c\u3002<\/p>\n\n\n\n<p>[10] <a href=\"https:\/\/www.raindrop.ai\/blog\/thoughts-on-evals\">www.raindrop.ai\/blog\/thoughts-on-evals<\/a><\/p>\n\n\n\n<p>\u5be6\u8e10\u8005 Matt Shumer [11] [12] \u5206\u4eab\u4e86 Otherside \u7684\u7d93\u9a57\uff0c\u4ed6\u5011\u4f7f\u7528\u57fa\u65bc\u771f\u5be6\u6d41\u91cf\u7684 A\/B \u6e2c\u8a66\uff0c\u6e2c\u91cf\u8a02\u95b1\u548c\u7559\u5b58\u7387\u3002\u4ed6\u76f4\u8a00 evals \u6709\u5e6b\u52a9\u4f46\u8207\u5be6\u969b\u6548\u7528\u7684\u76f8\u95dc\u6027\u4e0d\u9ad8\uff0c\u8a66\u904e\u6240\u6709\u65b9\u6cd5\u5f8c\uff0cA\/B \u6e2c\u8a66\u624d\u662f\u6b63\u9053\u3002\u4ed6\u7684\u89c0\u9ede\u7c21\u55ae\u660e\u77ad\uff1a\u5982\u679c\u4f60\u662f\u5be6\u9a57\u5ba4\uff0c\u5efa\u7acb\u901a\u7528\u524d\u6cbf\u6a21\u578b\uff0cevals \u5f88\u91cd\u8981\uff1b\u5982\u679c\u4f60\u662f\u7522\u54c1\u5efa\u69cb\u8005\uff0cevals \u5927\u591a\u4e0d\u91cd\u8981\u3002<\/p>\n\n\n\n<p>[11] <a href=\"https:\/\/x.com\/mattshumer_\/status\/1964178712452354551\">x.com\/mattshumer_\/status\/1964178712452354551<\/a><br>[12] <a href=\"https:\/\/x.com\/mattshumer_\/status\/1963789428029042822\">x.com\/mattshumer_\/status\/1963789428029042822<\/a><\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\u5c0b\u6c42\u5e73\u8861\u7684\u8072\u97f3<\/h3>\n\n\n\n<p>Hamel Husain [13] \u8a66\u5716\u8abf\u89e3\u9019\u5834\u7531\u5169\u5bb6\u4f9b\u61c9\u5546\u5f15\u767c\u7684\u6230\u722d\uff0c\u4ed6\u6307\u51fa\u4f60\u9700\u8981\u540c\u6642\u9032\u884c\u7dda\u4e0a\u548c\u96e2\u7dda\u6e2c\u8a66\uff0c\u5b83\u5011\u5404\u6709\u4e0d\u540c\u7684\u6b0a\u8861\u3002\u96e2\u7dda\u6307\u6a19\u662f\u884c\u70ba\u548c\u7d50\u679c\u7684\u66ff\u4ee3\u6307\u6a19\uff0c\u80fd\u5920\u66f4\u5feb\u901f\u5730\u8fed\u4ee3\uff1b\u7dda\u4e0a\u6307\u6a19\u548c A\/B \u6e2c\u8a66\u5247\u662f\u5fc5\u8981\u7684\uff0c\u7528\u4f86\u9a57\u8b49\u4f60\u7684\u96e2\u7dda\u6307\u6a19\u78ba\u5be6\u662f\u826f\u597d\u7684\u66ff\u4ee3\u6307\u6a19\uff0c\u800c\u4e14\u67d0\u4e9b\u6771\u897f\u53ea\u80fd\u900f\u904e\u7dda\u4e0a\u6e2c\u8a66\u4f86\u6e2c\u91cf\u3002\u4ed6\u611f\u5606\u9700\u8981\u5c07\u8cc7\u6599\u79d1\u5b78\u5e36\u56de AI \u5de5\u7a0b\u9818\u57df\uff0c\u56e0\u70ba\u9019\u7a2e\u7f3a\u5931\u5728\u7576\u524d\u7684\u8a0e\u8ad6\u4e2d\u986f\u800c\u6613\u898b\u3002\u4ed6\u88dc\u5145\u8aaa\uff0cOpenAI \u7684\u6280\u8853\u6210\u529f\u5718\u968a\u82b1\u8cbb\u5927\u91cf\u6642\u9593\u8207\u5ba2\u6236\u8a0e\u8ad6\u8a55\u4f30\uff08evals\uff09\uff0c\u4ed6\u5011\u8a8d\u70ba\u9019\u662f\u9ad8\u69d3\u687f\u7684\u6d3b\u52d5\uff0c\u8a55\u4f30\u6c7a\u5b9a\u4e86\u5c08\u6848\u80fd\u5426\u5f9e\u6982\u5ff5\u9a57\u8b49(PoC)\u968e\u6bb5\u9806\u5229\u63a8\u9032\u5230\u6210\u529f\u90e8\u7f72\u3002<\/p>\n\n\n\n<p>[13] <a href=\"https:\/\/x.com\/HamelHusain\/status\/1964110406596907170\">x.com\/HamelHusain\/status\/1964110406596907170<\/a><\/p>\n\n\n\n<p>Bryan Bischof [14] \u8a8d\u70ba swyx \u6bd4\u5176\u4ed6\u8a0e\u8ad6\u8005\u66f4\u6df1\u5165\u7406\u89e3\u9019\u500b\u8b70\u984c\u7684\u7d30\u5fae\u5dee\u7570\u3002\u4ed6\u6307\u51fa\u5e7e\u500b\u95dc\u9375\u9ede\uff1aevals \u4f5c\u70ba\u55ae\u5143\u6e2c\u8a66\u4e00\u76f4\u5927\u591a\u7121\u95dc\u7dca\u8981\uff0c\u9664\u4e86\u65e9\u671f\u8fed\u4ee3\u6216\u56de\u6b78\u6e2c\u8a66\uff1b\u8d85\u7d1a\u5ee3\u6cdb\u7684\u9ec3\u91d1 eval \u96c6\u5f80\u5f80\u8001\u5316\u5f97\u5f88\u5dee\uff1b\u90a3\u4e9b\u8aaa\u300c\u5c31 A\/B \u6e2c\u8a66\u554a\uff0c\u5144\u5f1f\u300d\u7684\u4eba\uff0c\u901a\u5e38\u9023 A\/B \u6e2c\u8a66\u90fd\u4e0d\u61c2\uff0c\u66f4\u5225\u8aaa\u5728 LLM \u7bc4\u5f0f\u4e2d\u505a\u6709\u6548\u7684 A\/B \u6e2c\u8a66\u3002\u4ed6\u89e3\u91cb\u70ba\u4ec0\u9ebc code agent \u516c\u53f8\u8f03\u5c11\u8ff7\u4fe1 evals\uff1a\u56e0\u70ba\u7528\u6236\u89d2\u8272\u548c\u7522\u54c1\u958b\u767c\u8005\u4e4b\u9593\u7684\u5dee\u8ddd\u5f88\u5c0f\uff0c\u5927\u5bb6\u90fd\u540c\u610f dogfooding \u975e\u5e38\u6709\u50f9\u503c\uff0c\u4f46\u98db\u8f2a\u662f\uff1adogfooding \u2192 \u932f\u8aa4\u5206\u6790 \u2192 \u7de8\u78bc\u5316\u7684 eval\u3002\u4ed6\u76f4\u8a00\u4e0d\u5728\u61c9\u7528\u5c64\u5de5\u4f5c\u7684\u4eba\u61c9\u8a72\u5728\u9019\u500b\u8a71\u984c\u4e0a\u9589\u5634\u3002<\/p>\n\n\n\n<p>[14] <a href=\"https:\/\/x.com\/BEBischof\/status\/1963739648792117484\">x.com\/BEBischof\/status\/1963739648792117484<\/a><\/p>\n\n\n\n<p>Brooke Hopkins [15] \u6307\u51fa\u9019\u5834\u8faf\u8ad6\u7684\u6838\u5fc3\u554f\u984c\uff1a\u6bcf\u500b\u4eba\u7528\u300cevals\u300d\u6307\u4ee3\u81f3\u5c11 6 \u7a2e\u4e0d\u540c\u7684\u6771\u897f\uff0c\u7136\u5f8c\u5728\u5c0d\u8a71\u6beb\u7121\u4ea4\u96c6\u6642\u611f\u5230\u9707\u9a5a\u3002\u5979\u8a8d\u70ba\u96d9\u65b9\u90fd\u932f\u904e\u4e86\u91cd\u9ede\uff0c\u771f\u6b63\u7684\u554f\u984c\u4e0d\u662f evals \u6709\u6548\u6216\u7121\u6548\uff0c\u800c\u662f\u5b9a\u7fa9\u6df7\u4e82\u3002\u5979\u4e3b\u5f35\u4e0d\u80fd\u628a\u751f\u7522\u524d\u548c\u751f\u7522\u5f8c\u7684\u8a55\u4f30\u5206\u958b\u770b\u5f85\uff0c\u9700\u8981\u6574\u5408\u7684\u7cfb\u7d71\uff0c\u8b93\u751f\u7522\u524d\u6a21\u64ec\u80fd\u5be6\u969b\u53cd\u6620\u771f\u5be6\u7528\u6236\u6a21\u5f0f\uff0c\u751f\u7522\u76e3\u63a7\u80fd\u56de\u994b\u6210\u66f4\u597d\u7684\u6e2c\u8a66\u6848\u4f8b\uff0c\u4eba\u5de5\u5be9\u67e5\u54e1\u80fd\u6709\u6548\u767c\u73fe\u6f14\u7b97\u6cd5\u6c38\u9060\u4e0d\u6703\u6355\u6349\u5230\u7684\u908a\u7de3\u6848\u4f8b\u3002\u5979\u7279\u5225\u5f37\u8abf\u9700\u8981\u66f4\u591a\u8cc7\u6599\u79d1\u5b78\u5bb6\u53c3\u8207 evals\uff0c\u800c\u4e0d\u53ea\u662f ML \u5de5\u7a0b\u5e2b\uff0c\u56e0\u70ba\u5927\u591a\u6578 eval \u6846\u67b6\u662f\u7531\u78ba\u5b9a\u6027\u601d\u7dad\u7684\u4eba\u69cb\u5efa\u7684\uff0c\u4f46 AI \u7cfb\u7d71\u662f\u6a5f\u7387\u6027\u7684\uff0c\u6211\u5011\u4f7f\u7528\u4e86\u932f\u8aa4\u7684\u5fc3\u667a\u6a21\u578b\u3002<br>[15] <a href=\"https:\/\/x.com\/bnicholehopkins\/status\/1965130607790264452\">x.com\/bnicholehopkins\/status\/1965130607790264452<\/a><\/p>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6700\u8fd1\u770b\u5230\u4e00\u5834\u95dc\u65bc AI Evals \u7684\u7cbe\u5f69\u8ad6\u6230\uff0c\u722d\u8ad6\u7126\u9ede\u4e0d\u5728\u6a21\u578b\u8a13\u7df4\u5c64\u9762\u7684\u8a55\u4f30(\u9019\u500b\u5927\u5bb6\u90fd\u6709\u5171\u8b58\u8981\u505a)\uff0c\u800c\u662f  &hellip; <\/p>\n<p class=\"link-more\"><a href=\"https:\/\/ihower.tw\/blog\/13221-ai-evals-debate\" class=\"more-link\">\u95b1\u8b80\u5168\u6587<span class=\"screen-reader-text\">\u3008AI Evals \u5927\u8faf\u8ad6: \u5f9e Claude Code \u8a2a\u8ac7\u5f15\u767c\u7684\u53cd\u601d\u3009<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"jetpack_post_was_ever_published":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_publicize_message":"","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":true,"jetpack_social_options":{"image_generator_settings":{"template":"highway","default_image_id":0,"font":"","enabled":false},"version":2}},"categories":[80],"tags":[],"class_list":["post-13221","post","type-post","status-publish","format-standard","hentry","category-llm","entry"],"jetpack_publicize_connections":[],"jetpack_featured_media_url":"","jetpack_shortlink":"https:\/\/wp.me\/p1q6tG-3rf","jetpack_sharing_enabled":true,"jetpack_likes_enabled":true,"_links":{"self":[{"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/posts\/13221","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/comments?post=13221"}],"version-history":[{"count":35,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/posts\/13221\/revisions"}],"predecessor-version":[{"id":13576,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/posts\/13221\/revisions\/13576"}],"wp:attachment":[{"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/media?parent=13221"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/categories?post=13221"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/tags?post=13221"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}